Release v2.5.5

- Add knowledge export/import row to features table with CLI link
- Document `praisonai knowledge export <file.json>` command
- Document `praisonai knowledge import <file.json>` command

.gitignore

@@ -80,4 +80,6 @@ CopilotKit*
 issue_details.json
 
 # MCP Registry tokens
-.mcpregistry_*
+.mcpregistry_*
+
+*.log

docker/Dockerfile.chat

@@ -16,7 +16,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
     praisonai_tools \
-    "praisonai>=2.4.0" \
+    "praisonai>=2.5.5" \
     "praisonai[chat]" \
     "embedchain[github,youtube]"
 

docker/Dockerfile.dev

@@ -20,7 +20,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
     praisonai_tools \
-    "praisonai>=2.4.0" \
+    "praisonai>=2.5.5" \
     "praisonai[ui]" \
     "praisonai[chat]" \
     "praisonai[realtime]" \

docker/Dockerfile.ui

@@ -16,7 +16,7 @@ RUN mkdir -p /root/.praison
 # Install Python packages (using latest versions)
 RUN pip install --no-cache-dir \
     praisonai_tools \
-    "praisonai>=2.4.0" \
+    "praisonai>=2.5.5" \
     "praisonai[ui]" \
     "praisonai[crewai]"
 

examples/async_runs/README.py

@@ -0,0 +1,35 @@
+"""
+Async Runs Examples
+===================
+
+This directory contains examples for the PraisonAI Async Jobs API.
+
+Examples:
+---------
+1. sdk_submit_poll_result.py - Python SDK example for submit/poll/result
+2. cli_submit_poll_result.sh - Bash/curl example for submit/poll/result
+3. idempotency.sh - Demonstrates idempotency key usage
+4. cancel.sh - Demonstrates job cancellation
+5. sse_stream.sh - Demonstrates SSE streaming
+
+Prerequisites:
+--------------
+1. Set OPENAI_API_KEY environment variable
+2. Start the jobs server:
+   
+   cd /path/to/praisonai-package/src/praisonai
+   python -m uvicorn praisonai.jobs.server:create_app --port 8005 --factory
+
+3. Run examples:
+   
+   python sdk_submit_poll_result.py
+   bash cli_submit_poll_result.sh
+   bash idempotency.sh
+   bash cancel.sh
+   bash sse_stream.sh
+
+Environment Variables:
+----------------------
+- OPENAI_API_KEY: Required for agent execution
+- PRAISONAI_BASE_URL: API server URL (default: http://127.0.0.1:8005)
+"""

examples/async_runs/cancel.sh

@@ -0,0 +1,54 @@
+#!/bin/bash
+# CLI Example: Cancel a Running Job
+#
+# This example demonstrates how to cancel a running job.
+#
+# Requirements:
+# - Jobs server running on http://127.0.0.1:8005
+#
+# Usage:
+#     bash cancel.sh
+
+set -e
+
+API_URL="${PRAISONAI_BASE_URL:-http://127.0.0.1:8005}"
+
+echo "============================================================"
+echo "PraisonAI Job Cancellation Example"
+echo "============================================================"
+
+# Submit a long-running job
+echo ""
+echo "1. Submitting a job..."
+RESPONSE=$(curl -s -X POST "$API_URL/api/v1/runs" \
+    -H "Content-Type: application/json" \
+    -d '{"prompt": "Write a detailed essay about the history of computing."}')
+
+JOB_ID=$(echo "$RESPONSE" | python3 -c "import sys,json; print(json.load(sys.stdin)['job_id'])")
+echo "   Job ID: $JOB_ID"
+
+# Wait briefly
+echo ""
+echo "2. Waiting 1 second..."
+sleep 1
+
+# Cancel the job
+echo ""
+echo "3. Cancelling job..."
+CANCEL_RESPONSE=$(curl -s -X POST "$API_URL/api/v1/runs/$JOB_ID/cancel")
+STATUS=$(echo "$CANCEL_RESPONSE" | python3 -c "import sys,json; print(json.load(sys.stdin)['status'])")
+echo "   Status after cancel: $STATUS"
+
+# Verify cancellation
+echo ""
+echo "4. Verification:"
+if [ "$STATUS" = "cancelled" ]; then
+    echo "   ✓ SUCCESS: Job was cancelled"
+else
+    echo "   ℹ Job status: $STATUS (may have completed before cancel)"
+fi
+
+echo ""
+echo "============================================================"
+echo "Cancellation example completed!"
+echo "============================================================"

examples/async_runs/cli_submit_poll_result.sh

@@ -0,0 +1,75 @@
+#!/bin/bash
+# CLI Example: Submit, Poll, and Get Result for Async Jobs
+#
+# This example demonstrates how to use the PraisonAI CLI to:
+# 1. Submit an async job
+# 2. Check job status
+# 3. Get the result
+#
+# Requirements:
+# - OPENAI_API_KEY environment variable set
+# - Jobs server running on http://127.0.0.1:8005
+#
+# Usage:
+#     bash cli_submit_poll_result.sh
+
+set -e
+
+API_URL="${PRAISONAI_BASE_URL:-http://127.0.0.1:8005}"
+
+echo "============================================================"
+echo "PraisonAI Async Jobs CLI Example"
+echo "============================================================"
+
+# Check if server is running
+echo ""
+echo "Checking server health..."
+if ! curl -s "$API_URL/health" > /dev/null 2>&1; then
+    echo "✗ Server not available at $API_URL"
+    echo "  Start the server with: python -m uvicorn praisonai.jobs.server:create_app --port 8005 --factory"
+    exit 1
+fi
+echo "✓ Server is healthy"
+
+# 1. Submit a job
+echo ""
+echo "1. Submitting job..."
+RESPONSE=$(curl -s -X POST "$API_URL/api/v1/runs" \
+    -H "Content-Type: application/json" \
+    -d '{"prompt": "What is 2+2? Answer with just the number."}')
+
+JOB_ID=$(echo "$RESPONSE" | python3 -c "import sys,json; print(json.load(sys.stdin)['job_id'])")
+echo "   Job ID: $JOB_ID"
+
+# 2. Poll for status
+echo ""
+echo "2. Polling for status..."
+while true; do
+    STATUS_RESPONSE=$(curl -s "$API_URL/api/v1/runs/$JOB_ID")
+    STATUS=$(echo "$STATUS_RESPONSE" | python3 -c "import sys,json; print(json.load(sys.stdin)['status'])")
+    PROGRESS=$(echo "$STATUS_RESPONSE" | python3 -c "import sys,json; print(json.load(sys.stdin)['progress']['percentage'])")
+    
+    echo "   Status: $STATUS | Progress: ${PROGRESS}%"
+    
+    if [ "$STATUS" = "succeeded" ] || [ "$STATUS" = "failed" ] || [ "$STATUS" = "cancelled" ]; then
+        break
+    fi
+    
+    sleep 2
+done
+
+# 3. Get result
+echo ""
+echo "3. Getting result..."
+if [ "$STATUS" = "succeeded" ]; then
+    RESULT=$(curl -s "$API_URL/api/v1/runs/$JOB_ID/result")
+    echo "   Result: $(echo "$RESULT" | python3 -c "import sys,json; print(json.load(sys.stdin)['result'])")"
+    echo "   Duration: $(echo "$RESULT" | python3 -c "import sys,json; print(f\"{json.load(sys.stdin)['duration_seconds']:.2f}s\")")"
+else
+    echo "   Job did not succeed: $STATUS"
+fi
+
+echo ""
+echo "============================================================"
+echo "Example completed!"
+echo "============================================================"

examples/async_runs/idempotency.sh

@@ -0,0 +1,57 @@
+#!/bin/bash
+# CLI Example: Idempotency Key Usage
+#
+# This example demonstrates how idempotency keys prevent duplicate job submissions.
+# Submitting the same request with the same idempotency key returns the existing job.
+#
+# Requirements:
+# - Jobs server running on http://127.0.0.1:8005
+#
+# Usage:
+#     bash idempotency.sh
+
+set -e
+
+API_URL="${PRAISONAI_BASE_URL:-http://127.0.0.1:8005}"
+IDEM_KEY="example-idem-key-$(date +%s)"
+
+echo "============================================================"
+echo "PraisonAI Idempotency Key Example"
+echo "============================================================"
+
+# First submission
+echo ""
+echo "1. First submission with Idempotency-Key: $IDEM_KEY"
+RESPONSE1=$(curl -s -X POST "$API_URL/api/v1/runs" \
+    -H "Content-Type: application/json" \
+    -H "Idempotency-Key: $IDEM_KEY" \
+    -d '{"prompt": "What is 1+1?"}')
+
+JOB_ID1=$(echo "$RESPONSE1" | python3 -c "import sys,json; print(json.load(sys.stdin)['job_id'])")
+echo "   Job ID: $JOB_ID1"
+
+# Second submission with same key
+echo ""
+echo "2. Second submission with SAME Idempotency-Key: $IDEM_KEY"
+RESPONSE2=$(curl -s -X POST "$API_URL/api/v1/runs" \
+    -H "Content-Type: application/json" \
+    -H "Idempotency-Key: $IDEM_KEY" \
+    -d '{"prompt": "Different prompt but same key"}')
+
+JOB_ID2=$(echo "$RESPONSE2" | python3 -c "import sys,json; print(json.load(sys.stdin)['job_id'])")
+echo "   Job ID: $JOB_ID2"
+
+# Verify same job ID
+echo ""
+echo "3. Verification:"
+if [ "$JOB_ID1" = "$JOB_ID2" ]; then
+    echo "   ✓ SUCCESS: Same job ID returned (no duplicate created)"
+else
+    echo "   ✗ FAILURE: Different job IDs returned"
+    exit 1
+fi
+
+echo ""
+echo "============================================================"
+echo "Idempotency example completed!"
+echo "============================================================"

examples/async_runs/sdk_submit_poll_result.py

@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+"""
+SDK Example: Submit, Poll, and Get Result for Async Jobs
+
+This example demonstrates how to use the PraisonAI Jobs SDK to:
+1. Submit an async job
+2. Poll for status
+3. Get the result
+
+Requirements:
+- OPENAI_API_KEY environment variable set
+- Jobs server running on http://127.0.0.1:8005
+
+Usage:
+    python sdk_submit_poll_result.py
+"""
+
+import os
+import time
+import httpx
+
+# Configuration
+API_URL = os.getenv("PRAISONAI_BASE_URL", "http://127.0.0.1:8005")
+POLL_INTERVAL = 2  # seconds
+
+
+def submit_job(prompt: str, idempotency_key: str = None) -> dict:
+    """Submit a job and return the response."""
+    headers = {}
+    if idempotency_key:
+        headers["Idempotency-Key"] = idempotency_key
+    
+    with httpx.Client(timeout=30.0) as client:
+        response = client.post(
+            f"{API_URL}/api/v1/runs",
+            json={"prompt": prompt},
+            headers=headers if headers else None
+        )
+        response.raise_for_status()
+        return response.json()
+
+
+def get_status(job_id: str) -> dict:
+    """Get job status."""
+    with httpx.Client(timeout=30.0) as client:
+        response = client.get(f"{API_URL}/api/v1/runs/{job_id}")
+        response.raise_for_status()
+        return response.json()
+
+
+def get_result(job_id: str) -> dict:
+    """Get job result."""
+    with httpx.Client(timeout=30.0) as client:
+        response = client.get(f"{API_URL}/api/v1/runs/{job_id}/result")
+        response.raise_for_status()
+        return response.json()
+
+
+def wait_for_completion(job_id: str, max_wait: int = 120) -> dict:
+    """Poll until job completes."""
+    start = time.time()
+    while time.time() - start < max_wait:
+        status = get_status(job_id)
+        print(f"  Status: {status['status']} | Progress: {status['progress']['percentage']:.0f}%")
+        
+        if status["status"] in ("succeeded", "failed", "cancelled"):
+            return status
+        
+        # Honor retry_after if present
+        wait_time = status.get("retry_after") or POLL_INTERVAL
+        time.sleep(wait_time)
+    
+    raise TimeoutError(f"Job {job_id} did not complete within {max_wait}s")
+
+
+def main():
+    print("=" * 60)
+    print("PraisonAI Async Jobs SDK Example")
+    print("=" * 60)
+    
+    # Check if server is running
+    try:
+        with httpx.Client(timeout=5.0) as client:
+            health = client.get(f"{API_URL}/health")
+            health.raise_for_status()
+            print(f"✓ Server healthy: {health.json()['status']}")
+    except Exception as e:
+        print(f"✗ Server not available at {API_URL}: {e}")
+        print("  Start the server with: python -m uvicorn praisonai.jobs.server:create_app --port 8005 --factory")
+        return
+    
+    # 1. Submit a job
+    print("\n1. Submitting job...")
+    prompt = "What is the capital of France? Answer in one word."
+    result = submit_job(prompt)
+    job_id = result["job_id"]
+    print(f"   Job ID: {job_id}")
+    print(f"   Status: {result['status']}")
+    print(f"   Poll URL: {result['poll_url']}")
+    
+    # 2. Wait for completion
+    print("\n2. Waiting for completion...")
+    final_status = wait_for_completion(job_id)
+    print(f"   Final status: {final_status['status']}")
+    
+    # 3. Get result
+    if final_status["status"] == "succeeded":
+        print("\n3. Getting result...")
+        result = get_result(job_id)
+        print(f"   Result: {result['result']}")
+        print(f"   Duration: {result['duration_seconds']:.2f}s")
+    else:
+        print(f"\n3. Job did not succeed: {final_status.get('error', 'Unknown error')}")
+    
+    print("\n" + "=" * 60)
+    print("Example completed successfully!")
+    print("=" * 60)
+
+
+if __name__ == "__main__":
+    main()

examples/async_runs/sse_stream.sh

@@ -0,0 +1,51 @@
+#!/bin/bash
+# CLI Example: SSE Streaming
+#
+# This example demonstrates how to stream job progress via Server-Sent Events.
+#
+# Requirements:
+# - Jobs server running on http://127.0.0.1:8005
+#
+# Usage:
+#     bash sse_stream.sh
+
+set -e
+
+API_URL="${PRAISONAI_BASE_URL:-http://127.0.0.1:8005}"
+
+echo "============================================================"
+echo "PraisonAI SSE Streaming Example"
+echo "============================================================"
+
+# Submit a job
+echo ""
+echo "1. Submitting job..."
+RESPONSE=$(curl -s -X POST "$API_URL/api/v1/runs" \
+    -H "Content-Type: application/json" \
+    -d '{"prompt": "Count from 1 to 5."}')
+
+JOB_ID=$(echo "$RESPONSE" | python3 -c "import sys,json; print(json.load(sys.stdin)['job_id'])")
+echo "   Job ID: $JOB_ID"
+
+# Stream progress
+echo ""
+echo "2. Streaming progress (Ctrl+C to stop)..."
+echo "   Connecting to: $API_URL/api/v1/runs/$JOB_ID/stream"
+echo ""
+
+# Use curl to stream SSE events (timeout after 60s)
+curl -s -N --max-time 60 "$API_URL/api/v1/runs/$JOB_ID/stream" | while read -r line; do
+    if [[ "$line" == data:* ]]; then
+        data="${line#data: }"
+        if [ "$data" = "[DONE]" ]; then
+            echo "   Stream completed"
+            break
+        fi
+        echo "   Event: $data"
+    fi
+done
+
+echo ""
+echo "============================================================"
+echo "Streaming example completed!"
+echo "============================================================"

examples/cli/persistence_commands.md

@@ -0,0 +1,93 @@
+# Persistence CLI Commands
+
+## Quick Reference
+
+```bash
+# Check database connectivity
+praisonai persistence doctor \
+  --conversation-url postgresql://localhost:5432/praisonai \
+  --state-url redis://localhost:6379
+
+# Run agent with persistence
+praisonai persistence run \
+  --conversation-url postgresql://localhost:5432/praisonai \
+  --session-id my-session
+
+# Resume a session
+praisonai persistence resume \
+  --conversation-url postgresql://localhost:5432/praisonai \
+  --session-id my-session
+
+# Export session to file
+praisonai persistence export \
+  --conversation-url postgresql://localhost:5432/praisonai \
+  --session-id my-session \
+  --output session_backup.jsonl
+
+# Import session from file
+praisonai persistence import \
+  --conversation-url postgresql://localhost:5432/praisonai \
+  --file session_backup.jsonl
+
+# Check schema status
+praisonai persistence status \
+  --conversation-url postgresql://localhost:5432/praisonai \
+  --state-url redis://localhost:6379
+
+# Apply migrations
+praisonai persistence migrate \
+  --conversation-url postgresql://localhost:5432/praisonai \
+  --state-url redis://localhost:6379
+```
+
+## Environment Variables
+
+Instead of CLI flags, you can use environment variables:
+
+```bash
+export PRAISON_CONVERSATION_URL=postgresql://localhost:5432/praisonai
+export PRAISON_STATE_URL=redis://localhost:6379
+export PRAISON_KNOWLEDGE_URL=http://localhost:6333
+
+# Then run without URL flags
+praisonai persistence doctor
+praisonai persistence status
+```
+
+## Docker Setup
+
+```bash
+# Start all services
+docker run -d --name praison-postgres \
+  -e POSTGRES_PASSWORD=praison123 \
+  -e POSTGRES_DB=praisonai \
+  -p 5432:5432 postgres:16
+
+docker run -d --name praison-redis \
+  -p 6379:6379 redis:7
+
+docker run -d --name praison-qdrant \
+  -p 6333:6333 -p 6334:6334 qdrant/qdrant
+
+# Verify
+docker ps
+```
+
+## Example Workflow
+
+```bash
+# 1. Start services
+docker-compose up -d
+
+# 2. Check connectivity
+praisonai persistence doctor
+
+# 3. Run agent
+praisonai persistence run --session-id demo-session
+
+# 4. Export for backup
+praisonai persistence export --session-id demo-session --output backup.jsonl
+
+# 5. Import to another environment
+praisonai persistence import --file backup.jsonl
+```

examples/cli/persistence_end_to_end.md

@@ -0,0 +1,124 @@
+# PraisonAI Persistence CLI - End-to-End Guide
+
+This guide demonstrates the CLI commands for database persistence.
+
+## Prerequisites
+
+```bash
+pip install "praisonai[tools]"
+
+# Start Docker containers
+docker run -d --name praison-postgres -p 5432:5432 \
+    -e POSTGRES_PASSWORD=praison123 \
+    -e POSTGRES_DB=praisonai \
+    postgres:16
+
+docker run -d --name praison-qdrant -p 6333:6333 qdrant/qdrant
+docker run -d --name praison-redis -p 6379:6379 redis:7
+```
+
+## 1. Doctor Command - Validate Connectivity
+
+```bash
+# Check all stores
+praisonai persistence doctor \
+    --conversation-url "postgresql://postgres:praison123@localhost:5432/praisonai" \
+    --knowledge-url "http://localhost:6333" \
+    --state-url "redis://localhost:6379"
+```
+
+**Expected Output:**
+```
+==================================================
+PraisonAI Persistence Doctor
+==================================================
+
+[Conversation Store] Testing: postgresql://postgres:****@localhost:5432/praisonai
+  ✅ Connected (postgres)
+
+[Knowledge Store] Testing: http://localhost:6333
+  ✅ Connected (qdrant)
+
+[State Store] Testing: redis://localhost:6379
+  ✅ Connected (redis)
+
+==================================================
+Results: 3/3 stores connected successfully
+==================================================
+```
+
+## 2. Run Command - Execute Agent with Persistence
+
+```bash
+# Run agent with persistence (dry-run first)
+praisonai persistence run \
+    --session-id "cli-demo-001" \
+    --conversation-url "postgresql://postgres:praison123@localhost:5432/praisonai" \
+    --dry-run \
+    "Hello, my name is Bob"
+
+# Actually run it
+praisonai persistence run \
+    --session-id "cli-demo-001" \
+    --conversation-url "postgresql://postgres:praison123@localhost:5432/praisonai" \
+    "Hello, my name is Bob"
+```
+
+**Expected Output:**
+```
+[Session: cli-demo-001]
+Agent: Hello Bob! Nice to meet you. How can I help you today?
+```
+
+## 3. Resume Command - Continue a Session
+
+```bash
+# Show history
+praisonai persistence resume \
+    --session-id "cli-demo-001" \
+    --conversation-url "postgresql://postgres:praison123@localhost:5432/praisonai" \
+    --show-history
+
+# Continue conversation
+praisonai persistence resume \
+    --session-id "cli-demo-001" \
+    --conversation-url "postgresql://postgres:praison123@localhost:5432/praisonai" \
+    --continue "What's my name?"
+```
+
+**Expected Output:**
+```
+[Session: cli-demo-001]
+Messages in history: 2
+
+--- Conversation History ---
+[USER] Hello, my name is Bob
+[ASSISTANT] Hello Bob! Nice to meet you...
+--- End History ---
+
+Agent: Your name is Bob!
+```
+
+## Environment Variables
+
+Instead of passing URLs on command line, use environment variables:
+
+```bash
+export PRAISON_CONVERSATION_URL="postgresql://postgres:praison123@localhost:5432/praisonai"
+export PRAISON_KNOWLEDGE_URL="http://localhost:6333"
+export PRAISON_STATE_URL="redis://localhost:6379"
+export OPENAI_API_KEY="your-key"
+
+# Now commands are simpler
+praisonai persistence doctor --all
+praisonai persistence run --session-id "my-session" "Hello!"
+```
+
+## Full CLI Reference
+
+```
+praisonai persistence --help
+praisonai persistence doctor --help
+praisonai persistence run --help
+praisonai persistence resume --help
+```

examples/middleware/basic_middleware.py

@@ -0,0 +1,55 @@
+"""
+Basic Middleware Example - PraisonAI Agents
+
+Demonstrates before/after hooks and wrap decorators for model and tool calls.
+"""
+
+from praisonaiagents import Agent, tool
+from praisonaiagents.hooks import (
+    before_model, after_model, wrap_model_call,
+    before_tool, after_tool, wrap_tool_call,
+    InvocationContext, ModelRequest, ModelResponse
+)
+
+# Simple tool for demonstration
+@tool
+def get_weather(city: str) -> str:
+    """Get weather for a city."""
+    return f"Sunny, 22°C in {city}"
+
+# Before model hook - adds context
+@before_model
+def add_context(request: ModelRequest) -> ModelRequest:
+    print(f"[before_model] Adding context to request")
+    return request
+
+# After model hook - logs response
+@after_model  
+def log_response(response: ModelResponse) -> ModelResponse:
+    print(f"[after_model] Response received")
+    return response
+
+# Wrap tool call - retry on error
+@wrap_tool_call
+def retry_on_error(tool_call, call_next):
+    print(f"[wrap_tool_call] Executing tool")
+    try:
+        return call_next(tool_call)
+    except Exception as e:
+        print(f"[wrap_tool_call] Retrying after error: {e}")
+        return call_next(tool_call)
+
+# Create agent with hooks
+agent = Agent(
+    name="WeatherBot",
+    instructions="You help with weather queries.",
+    tools=[get_weather],
+    hooks=[add_context, log_response, retry_on_error]
+)
+
+if __name__ == "__main__":
+    # Test the tool directly
+    result = get_weather("London")
+    print(f"Weather: {result}")
+    
+    print("\n✓ Middleware example complete")

examples/middleware/configurable_model.py

@@ -0,0 +1,31 @@
+"""
+Configurable Model Example - PraisonAI Agents
+
+Demonstrates runtime model switching without recreating the agent.
+"""
+
+from praisonaiagents import Agent
+
+# Create agent with configurable model enabled
+agent = Agent(
+    name="FlexBot",
+    instructions="You are a helpful assistant.",
+    llm="gpt-4o-mini",
+    llm_config={"configurable": True}
+)
+
+if __name__ == "__main__":
+    # Default model call
+    print("Using default model (gpt-4o-mini)...")
+    # response = agent.chat("Say hello in 5 words")
+    
+    # Override model per-call
+    print("\nOverriding to different model...")
+    # response = agent.chat("Say hello in 5 words", config={"model": "gpt-4o"})
+    
+    # Override temperature
+    print("\nOverriding temperature...")
+    # response = agent.chat("Say hello creatively", config={"temperature": 0.9})
+    
+    print("\n✓ Configurable model example complete")
+    print("Note: Uncomment agent.chat() calls to run with API key")

examples/middleware/injected_state.py

@@ -0,0 +1,50 @@
+"""
+Injected State Example - PraisonAI Agents
+
+Demonstrates injecting agent state into tools without exposing it in the schema.
+"""
+
+from praisonaiagents import Agent, tool
+from praisonaiagents.tools import Injected
+from praisonaiagents.tools.injected import AgentState, with_injection_context
+
+# Tool with injected state - state param is NOT in the public schema
+@tool
+def show_context(query: str, state: Injected[dict]) -> str:
+    """Show the current agent context."""
+    session_id = state.get('session_id', 'unknown')
+    agent_id = state.get('agent_id', 'unknown')
+    return f"Query: {query}, Session: {session_id}, Agent: {agent_id}"
+
+# Create agent with the tool
+agent = Agent(
+    name="ContextBot",
+    instructions="You help show context information.",
+    tools=[show_context],
+    session_id="my-session-123"
+)
+
+if __name__ == "__main__":
+    # Verify injected param is not in schema
+    schema = show_context.get_schema()
+    params = schema['function']['parameters']['properties']
+    print(f"Schema params: {list(params.keys())}")
+    assert 'state' not in params, "state should NOT be in schema"
+    print("✓ 'state' correctly excluded from schema")
+    
+    # Test with manual injection context
+    mock_state = AgentState(
+        agent_id="test-agent",
+        run_id="run-1",
+        session_id="session-abc"
+    )
+    
+    with with_injection_context(mock_state):
+        result = show_context(query="hello")
+        print(f"Result: {result}")
+    
+    # Test via agent.execute_tool
+    result = agent.execute_tool("show_context", {"query": "test"})
+    print(f"Agent result: {result}")
+    
+    print("\n✓ Injected state example complete")

examples/middleware/rate_limiter.py

@@ -0,0 +1,29 @@
+"""
+Rate Limiter Example - PraisonAI Agents
+
+Demonstrates token bucket rate limiting for LLM API calls.
+"""
+
+from praisonaiagents import Agent
+from praisonaiagents.llm import RateLimiter
+
+# Create rate limiter: 60 requests per minute
+limiter = RateLimiter(requests_per_minute=60, burst=5)
+
+# Create agent with rate limiter
+agent = Agent(
+    name="RateLimitedBot",
+    instructions="You are a helpful assistant.",
+    rate_limiter=limiter
+)
+
+if __name__ == "__main__":
+    print(f"Rate limiter: {limiter}")
+    print(f"Available tokens: {limiter.available_tokens}")
+    
+    # Demonstrate rate limiting
+    for i in range(3):
+        limiter.acquire()
+        print(f"Request {i+1} acquired, tokens left: {limiter.available_tokens:.1f}")
+    
+    print("\n✓ Rate limiter example complete")

examples/multi_agent/multi_agent_persistence.py

@@ -0,0 +1,122 @@
+"""
+Multi-Agent Persistence Example
+
+Demonstrates how multiple agents can share or isolate conversation
+history using the persistence layer.
+"""
+
+from praisonai.persistence.factory import create_conversation_store, create_state_store
+from praisonai.persistence.conversation.base import ConversationSession, ConversationMessage
+import uuid
+
+
+def main():
+    # Create stores
+    conv_store = create_conversation_store("sqlite", path=":memory:")
+    state_store = create_state_store("memory")
+    
+    print("=== Multi-Agent Persistence Demo ===\n")
+    
+    # Scenario 1: Shared session (agents collaborate)
+    print("--- Scenario 1: Shared Session ---")
+    shared_session_id = f"shared-{uuid.uuid4().hex[:8]}"
+    
+    session = ConversationSession(
+        session_id=shared_session_id,
+        user_id="user-123",
+        name="Shared Conversation"
+    )
+    conv_store.create_session(session)
+    
+    # Agent 1 (Researcher) adds context
+    conv_store.add_message(shared_session_id, ConversationMessage(
+        role="assistant",
+        content="[Researcher] I found relevant information about the topic.",
+        metadata={"agent_id": "researcher", "agent_role": "research"}
+    ))
+    
+    # Agent 2 (Writer) uses the context
+    conv_store.add_message(shared_session_id, ConversationMessage(
+        role="assistant", 
+        content="[Writer] Based on the research, here's a summary...",
+        metadata={"agent_id": "writer", "agent_role": "writing"}
+    ))
+    
+    # Agent 3 (Editor) reviews
+    conv_store.add_message(shared_session_id, ConversationMessage(
+        role="assistant",
+        content="[Editor] I've reviewed and polished the content.",
+        metadata={"agent_id": "editor", "agent_role": "editing"}
+    ))
+    
+    messages = conv_store.get_messages(shared_session_id)
+    print(f"Shared session has {len(messages)} messages from different agents")
+    for msg in messages:
+        agent = msg.metadata.get("agent_id", "unknown") if msg.metadata else "unknown"
+        print(f"  - {agent}: {msg.content[:50]}...")
+    
+    # Scenario 2: Isolated sessions (agents work independently)
+    print("\n--- Scenario 2: Isolated Sessions ---")
+    
+    agent_sessions = {}
+    for agent_name in ["planner", "executor", "validator"]:
+        session_id = f"{agent_name}-{uuid.uuid4().hex[:8]}"
+        agent_sessions[agent_name] = session_id
+        
+        conv_store.create_session(ConversationSession(
+            session_id=session_id,
+            agent_id=agent_name,
+            name=f"{agent_name.title()} Session"
+        ))
+        
+        # Each agent has its own conversation
+        conv_store.add_message(session_id, ConversationMessage(
+            role="user",
+            content=f"Task for {agent_name}"
+        ))
+        conv_store.add_message(session_id, ConversationMessage(
+            role="assistant",
+            content=f"[{agent_name.title()}] Processing task..."
+        ))
+    
+    print("Created isolated sessions for each agent:")
+    for agent, sid in agent_sessions.items():
+        msgs = conv_store.get_messages(sid)
+        print(f"  - {agent}: {len(msgs)} messages")
+    
+    # Scenario 3: Shared state across agents
+    print("\n--- Scenario 3: Shared State ---")
+    
+    # Agents share state for coordination
+    state_store.set("workflow:status", {"phase": "research", "progress": 0})
+    state_store.set("workflow:context", {"topic": "AI Agents", "deadline": "2024-12-31"})
+    
+    # Agent 1 updates progress
+    state = state_store.get("workflow:status")
+    state["progress"] = 33
+    state["phase"] = "writing"
+    state_store.set("workflow:status", state)
+    print(f"Agent 1 updated state: {state}")
+    
+    # Agent 2 reads and updates
+    state = state_store.get("workflow:status")
+    state["progress"] = 66
+    state["phase"] = "editing"
+    state_store.set("workflow:status", state)
+    print(f"Agent 2 updated state: {state}")
+    
+    # Agent 3 completes
+    state = state_store.get("workflow:status")
+    state["progress"] = 100
+    state["phase"] = "complete"
+    state_store.set("workflow:status", state)
+    print(f"Agent 3 completed: {state}")
+    
+    # Cleanup
+    conv_store.close()
+    state_store.close()
+    print("\n✅ Multi-agent persistence demo complete!")
+
+
+if __name__ == "__main__":
+    main()

examples/multi_agent/shared_db_session.py

@@ -0,0 +1,40 @@
+"""Multi-Agent Shared Session - DB Persistence Test"""
+import sys
+sys.path.insert(0, 'src/praisonai')
+from praisonai.persistence import create_conversation_store
+from praisonai.persistence.conversation.base import ConversationSession, ConversationMessage
+
+# Create shared session for multiple agents
+store = create_conversation_store("sqlite", path="/tmp/multi_agent.db")
+session_id = "shared-session-123"
+
+# Create session
+session = ConversationSession(session_id=session_id, agent_id="multi")
+try:
+    store.create_session(session)
+except Exception:
+    pass
+
+# Agent 1 writes
+msg1 = ConversationMessage(session_id=session_id, role="user", content="Agent1: Hello")
+store.add_message(session_id, msg1)
+msg2 = ConversationMessage(session_id=session_id, role="assistant", content="Agent1: Hi there!")
+store.add_message(session_id, msg2)
+
+# Agent 2 writes to same session
+msg3 = ConversationMessage(session_id=session_id, role="user", content="Agent2: What did Agent1 say?")
+store.add_message(session_id, msg3)
+msg4 = ConversationMessage(session_id=session_id, role="assistant", content="Agent2: Agent1 said Hello")
+store.add_message(session_id, msg4)
+
+# Verify both agents' messages are in session
+messages = store.get_messages(session_id)
+print(f"Total messages in shared session: {len(messages)}")
+assert len(messages) >= 4, "Should have messages from both agents"
+
+# Verify content
+contents = [m.content for m in messages]
+assert any("Agent1" in c for c in contents), "Should have Agent1 messages"
+assert any("Agent2" in c for c in contents), "Should have Agent2 messages"
+
+print("PASSED: Multi-agent shared session persistence")

examples/multi_agent/shared_session_wow.py

@@ -0,0 +1,16 @@
+"""Multi-Agent Shared Session Demo"""
+from praisonaiagents import Agent
+from praisonai.persistence import create_conversation_store
+
+store = create_conversation_store("sqlite", path="/tmp/multi_agent.db")
+
+agent1 = Agent(name="Researcher", llm="gpt-4o-mini", db=store, session_id="shared-session")
+agent2 = Agent(name="Writer", llm="gpt-4o-mini", db=store, session_id="shared-session")
+
+r1 = agent1.chat("Research topic: AI in healthcare. Key finding: reduces diagnosis time by 50%")
+print(f"Researcher: {r1[:50]}...")
+
+r2 = agent2.chat("Based on the research, what was the key finding?")
+print(f"Writer: {r2[:50]}...")
+assert "50" in r2 or "diagnosis" in r2.lower(), "Writer should see Researcher's context"
+print("PASSED: Multi-agent shared session works")

examples/observability/basic_usage.py

@@ -0,0 +1,48 @@
+"""
+Basic Observability Usage Example
+
+This example demonstrates how to use the observability module
+with PraisonAI agents.
+
+Usage:
+    # Set your provider's API key first
+    export LANGFUSE_PUBLIC_KEY=pk-lf-xxx
+    export LANGFUSE_SECRET_KEY=sk-lf-xxx
+    
+    python basic_usage.py
+"""
+
+from praisonai_tools.observability import obs
+from praisonaiagents import Agent
+
+# Initialize observability (auto-detects provider from env vars)
+obs.init()
+
+# Or specify a provider explicitly
+# obs.init(provider="langfuse", project_name="my-project")
+
+# Create an agent
+agent = Agent(
+    instructions="You are a helpful assistant.",
+    model="gpt-4o-mini",
+)
+
+# Use trace context manager for complete workflow tracing
+with obs.trace("chat-session", session_id="user-123"):
+    # Use span context manager for individual operations
+    with obs.span("user-query", kind=obs.SpanKind.AGENT):
+        response = agent.chat("What is the capital of France?")
+        print(response)
+
+# Log LLM calls manually if needed
+obs.log_llm_call(
+    model="gpt-4o-mini",
+    input_messages="What is 2+2?",
+    output="4",
+    input_tokens=10,
+    output_tokens=1,
+)
+
+# Check diagnostics
+print("\nDiagnostics:")
+print(obs.doctor())

examples/observability/langfuse_example.py

@@ -0,0 +1,48 @@
+"""
+Langfuse Integration Example
+
+This example shows how to use Langfuse for LLM observability.
+
+Setup:
+    1. Sign up at https://langfuse.com/
+    2. Get your API keys from the project settings
+    3. Set environment variables:
+       export LANGFUSE_PUBLIC_KEY=pk-lf-xxx
+       export LANGFUSE_SECRET_KEY=sk-lf-xxx
+    4. Install dependencies:
+       pip install opentelemetry-sdk opentelemetry-exporter-otlp
+
+Usage:
+    python langfuse_example.py
+"""
+
+import os
+from praisonai_tools.observability import obs
+from praisonaiagents import Agent
+
+# Initialize Langfuse
+success = obs.init(
+    provider="langfuse",
+    project_name="praisonai-demo",
+)
+
+if not success:
+    print("Failed to initialize Langfuse. Check your API keys.")
+    print("Required: LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY")
+    exit(1)
+
+print("Langfuse initialized successfully!")
+print(f"View traces at: https://cloud.langfuse.com/")
+
+# Create agent
+agent = Agent(
+    instructions="You are a helpful coding assistant.",
+    model="gpt-4o-mini",
+)
+
+# Run with tracing
+with obs.trace("coding-session", user_id="developer-1"):
+    response = agent.chat("Write a Python function to calculate fibonacci numbers")
+    print(response)
+
+print("\nCheck Langfuse dashboard for traces!")

examples/observability/mlflow_local_wow.py

@@ -0,0 +1,22 @@
+"""MLflow Local Observability Demo (no API key needed)"""
+import mlflow
+import time
+
+mlflow.set_tracking_uri("file:///tmp/mlflow_demo")
+mlflow.set_experiment("praisonai-demo")
+
+with mlflow.start_run(run_name="agent-trace"):
+    mlflow.log_param("agent_name", "Assistant")
+    mlflow.log_param("model", "gpt-4o-mini")
+    
+    # Simulate agent work
+    start = time.time()
+    time.sleep(0.1)
+    latency = time.time() - start
+    
+    mlflow.log_metric("latency_ms", latency * 1000)
+    mlflow.log_metric("tokens_used", 150)
+    
+    print(f"Logged run to MLflow: latency={latency*1000:.1f}ms")
+
+print("PASSED: MLflow local observability works")

examples/observability/multi_agent_tracing.py

@@ -0,0 +1,53 @@
+"""
+Multi-Agent Tracing Example
+
+This example shows how to trace multi-agent workflows with
+parent/child span relationships.
+
+Usage:
+    # Set your provider's API key
+    export LANGFUSE_PUBLIC_KEY=pk-lf-xxx
+    export LANGFUSE_SECRET_KEY=sk-lf-xxx
+    
+    python multi_agent_tracing.py
+"""
+
+from praisonai_tools.observability import obs
+from praisonai_tools.observability.base import SpanKind
+from praisonaiagents import Agent, PraisonAIAgents
+
+# Initialize observability
+obs.init()
+
+# Create agents
+researcher = Agent(
+    name="Researcher",
+    instructions="You research topics and provide detailed information.",
+    model="gpt-4o-mini",
+)
+
+writer = Agent(
+    name="Writer", 
+    instructions="You write clear, engaging content based on research.",
+    model="gpt-4o-mini",
+)
+
+# Trace the multi-agent workflow
+with obs.trace("content-creation", metadata={"workflow": "research-write"}):
+    
+    # Research phase
+    with obs.span("research-phase", kind=SpanKind.AGENT) as research_span:
+        research_span.attributes["agent"] = "Researcher"
+        research_result = researcher.chat("Research the benefits of AI in healthcare")
+    
+    # Writing phase  
+    with obs.span("writing-phase", kind=SpanKind.AGENT) as write_span:
+        write_span.attributes["agent"] = "Writer"
+        write_span.attributes["input_length"] = len(str(research_result))
+        final_content = writer.chat(f"Write a blog post based on: {research_result}")
+
+print("Final Content:")
+print(final_content)
+
+# Check trace info
+print("\nDiagnostics:", obs.doctor())

examples/persistence/knowledge_qdrant.py

@@ -0,0 +1,71 @@
+"""
+Qdrant Knowledge Store - Agent-First Example
+
+Demonstrates using Qdrant for knowledge-based RAG with an Agent.
+
+Requirements:
+    pip install "praisonai[tools]"
+
+Docker Setup:
+    docker run -d --name praison-qdrant -p 6333:6333 qdrant/qdrant
+
+Run:
+    python knowledge_qdrant.py
+
+Expected Output:
+    Agent responds using knowledge from documents
+"""
+
+from praisonaiagents import Agent
+import tempfile
+import os
+import shutil
+
+# Create sample knowledge document
+sample_doc = """
+# AI and Programming Guide
+
+## Python Programming
+Python is a versatile programming language used for web development, data science, and AI.
+It has a simple syntax and extensive libraries.
+
+## Machine Learning
+Machine learning enables computers to learn patterns from data without explicit programming.
+Common frameworks include TensorFlow, PyTorch, and scikit-learn.
+
+## Databases
+PostgreSQL is a powerful open-source relational database system.
+It supports both SQL and JSON data types.
+"""
+
+# Save to temp file
+temp_dir = tempfile.mkdtemp()
+doc_path = os.path.join(temp_dir, "guide.txt")
+with open(doc_path, "w") as f:
+    f.write(sample_doc)
+
+print("=== Qdrant Knowledge Store Demo (Agent-First) ===")
+
+# Agent-first approach: use knowledge parameter with Qdrant
+agent = Agent(
+    name="KnowledgeAssistant",
+    instructions="You are a helpful assistant with access to technical documentation.",
+    knowledge=[doc_path],
+    knowledge_config={
+        "vector_store": "qdrant",
+        "url": "http://localhost:6333"
+    }
+)
+
+# Chat - agent uses knowledge for RAG
+response = agent.chat("What programming language is good for AI?")
+print(f"Response: {response}")
+
+# Cleanup
+shutil.rmtree(temp_dir)
+
+print("\n=== Demo Complete ===")
+
+# --- Advanced: Direct Store Usage ---
+# from praisonai.persistence.factory import create_knowledge_store
+# store = create_knowledge_store("qdrant", url="http://localhost:6333")

examples/persistence/minimal_agent_db.py

@@ -0,0 +1,43 @@
+"""
+Minimal Agent with Database Persistence
+
+This is the simplest way to add persistence to your agent.
+Messages, runs, and traces are automatically saved.
+
+Run:
+    python minimal_agent_db.py
+
+Expected output:
+    - Agent responds to your question
+    - Data is persisted to PostgreSQL
+    - Session can be resumed later with same session_id
+"""
+
+from praisonaiagents import Agent, db
+
+# Create database connection (PostgreSQL + Redis)
+my_db = db(
+    database_url="postgresql://postgres:praison123@localhost:5432/praisonai",
+    state_url="redis://localhost:6379"  # Optional: for state/tracing
+)
+
+# Create agent with persistence
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant. Be concise.",
+    db=my_db,
+    # session_id is optional - defaults to per-hour ID (YYYYMMDDHH-hash)
+    # Set explicitly for session continuity across runs:
+    session_id="my-persistent-session"
+)
+
+# Chat - messages are automatically persisted
+response = agent.chat("What is the capital of France?")
+print(f"Response: {response}")
+
+# Verify persistence
+data = my_db.export_session("my-persistent-session")
+print(f"\nMessages stored: {len(data.get('messages', []))}")
+
+# Clean up
+my_db.close()

examples/persistence/postgres_runs_traces.py

@@ -0,0 +1,61 @@
+"""
+PostgreSQL with Runs and Traces
+
+Demonstrates full persistence with run tracking and tracing.
+
+Run:
+    python postgres_runs_traces.py
+
+Expected output:
+    - Agent responds with tool usage
+    - Runs and traces are persisted
+    - Export shows full session data
+"""
+
+from praisonaiagents import Agent, db
+
+def get_weather(city: str) -> str:
+    """Get weather for a city."""
+    return f"Weather in {city}: Sunny, 72°F"
+
+# Create database with PostgreSQL + Redis
+my_db = db(
+    database_url="postgresql://postgres:praison123@localhost:5432/praisonai",
+    state_url="redis://localhost:6379"
+)
+
+# Create agent with tools
+agent = Agent(
+    name="WeatherBot",
+    instructions="You help with weather. Use the get_weather tool when asked.",
+    tools=[get_weather],
+    db=my_db,
+    session_id="weather-session-001"
+)
+
+# First chat - triggers run tracking
+print("=== First Question ===")
+response1 = agent.chat("What's the weather in Tokyo?")
+print(f"Response: {response1}\n")
+
+# Second chat - same session, new run
+print("=== Second Question ===")
+response2 = agent.chat("How about New York?")
+print(f"Response: {response2}\n")
+
+# Export session data
+print("=== Session Data ===")
+data = my_db.export_session("weather-session-001")
+print(f"Session ID: {data.get('session_id')}")
+print(f"Messages: {len(data.get('messages', []))}")
+print(f"Agent: {data.get('agent_name')}")
+
+# Show messages
+print("\n=== Messages ===")
+for msg in data.get('messages', []):
+    role = msg.get('role', 'unknown')
+    content = msg.get('content', '')[:60]
+    print(f"  [{role}] {content}...")
+
+my_db.close()
+print("\n✅ Done")

examples/persistence/redis_state.py

@@ -0,0 +1,46 @@
+"""
+Redis State Store - Agent-First Example
+
+Use Redis for state management with an Agent.
+
+Docker Setup:
+    docker run -d --name redis -p 6379:6379 redis:7
+
+Run:
+    python redis_state.py
+
+Expected output:
+    Agent responds with state persisted to Redis
+"""
+
+from praisonaiagents import Agent, db
+
+print("=== Redis State Store (Agent-First) ===")
+
+# Agent-first approach: use db parameter with Redis state store
+my_db = db(
+    database_url="sqlite:///conversations.db",  # Conversations
+    state_url="redis://localhost:6379"          # State/tracing
+)
+
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant.",
+    db=my_db,
+    session_id="redis-state-session"
+)
+
+# Chat - state is automatically managed via Redis
+response = agent.chat("Hello! Remember my name is Alice.")
+print(f"Response: {response}")
+
+# Second message to test state
+response2 = agent.chat("What is my name?")
+print(f"Response: {response2}")
+
+my_db.close()
+print("\n✅ Done")
+
+# --- Advanced: Direct Store Usage ---
+# from praisonaiagents import db
+# redis_db = db.RedisDB(host="localhost", port=6379)

examples/persistence/session_resume.py

@@ -0,0 +1,66 @@
+"""
+Session Resume Example
+
+Demonstrates how to resume a conversation from a previous session.
+
+Run:
+    python session_resume.py
+
+Expected output:
+    - First run: Creates new session
+    - Second run: Resumes with history
+"""
+
+from praisonaiagents import Agent, db
+
+# Create database
+my_db = db(
+    database_url="postgresql://postgres:praison123@localhost:5432/praisonai",
+    state_url="redis://localhost:6379"
+)
+
+SESSION_ID = "resume-demo-session"
+
+# Create agent with explicit session_id
+agent = Agent(
+    name="MemoryBot",
+    instructions="You remember our conversation. Be helpful and concise.",
+    db=my_db,
+    session_id=SESSION_ID
+)
+
+# Check if this is a new or resumed session
+data = my_db.export_session(SESSION_ID)
+existing_messages = len(data.get('messages', []))
+
+if existing_messages == 0:
+    print("=== NEW SESSION ===")
+    print("This is a new conversation.\n")
+    
+    response = agent.chat("My name is Alice and I love Python programming.")
+    print(f"Response: {response}\n")
+    
+    response = agent.chat("What's my favorite programming language?")
+    print(f"Response: {response}\n")
+else:
+    print(f"=== RESUMED SESSION ({existing_messages} previous messages) ===")
+    print("Continuing previous conversation.\n")
+    
+    # Show previous messages
+    print("Previous messages:")
+    for msg in data.get('messages', [])[:4]:
+        role = msg.get('role', 'unknown')
+        content = msg.get('content', '')[:60]
+        print(f"  [{role}] {content}...")
+    print()
+    
+    # Continue conversation
+    response = agent.chat("What's my name and what do I like?")
+    print(f"Response: {response}\n")
+
+# Show final state
+final_data = my_db.export_session(SESSION_ID)
+print(f"Total messages now: {len(final_data.get('messages', []))}")
+
+my_db.close()
+print("✅ Done")

examples/persistence/simple_db_agent.py

@@ -0,0 +1,58 @@
+"""
+Simple Database Persistence Example for PraisonAI Agents.
+
+This example shows how to enable automatic conversation persistence
+with just 2 lines of code.
+
+Requirements:
+    pip install "praisonai[tools]"
+    
+Docker Setup:
+    docker run -d --name praison-postgres -p 5432:5432 \
+        -e POSTGRES_PASSWORD=praison123 \
+        -e POSTGRES_DB=praisonai \
+        postgres:16
+
+Environment:
+    export OPENAI_API_KEY=your-key
+
+Run:
+    python simple_agent_postgres.py
+
+Expected Output:
+    === First Run ===
+    Agent: Nice to meet you, Alice! Python is a great choice...
+    Session ID: demo-session-001
+    Messages persisted: 2
+"""
+
+from praisonaiagents import Agent, db
+
+# Create database adapter (simplified import)
+db_instance = db(
+    database_url="postgresql://postgres:praison123@localhost:5432/praisonai"
+)
+
+# Create agent with persistence enabled
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant. Keep responses brief.",
+    db=db_instance,
+    session_id="demo-session-001",  # Same session_id = resume conversation
+    verbose=False
+)
+
+# Chat with persistence
+print("=== First Run ===")
+response = agent.chat("My name is Alice and I love Python programming.")
+print(f"Agent: {response}\n")
+
+# Show session info
+print(f"Session ID: {agent.session_id}")
+print(f"Messages in history: {len(agent.chat_history)}")
+
+# Close DB connection
+db_instance.close()
+
+print("\n=== Session persisted! ===")
+print("Run again with same session_id to resume.")

examples/persistence/sqlite_local.py

@@ -0,0 +1,41 @@
+"""
+SQLite Local Database
+
+Use SQLite for simple local persistence without external services.
+
+Run:
+    python sqlite_local.py
+
+Expected output:
+    - Agent responds
+    - Data persisted to local SQLite file
+"""
+
+from praisonaiagents import Agent, db
+import tempfile
+import os
+
+# Create SQLite database (local file)
+db_path = os.path.join(tempfile.gettempdir(), "my_agent.db")
+my_db = db.SQLiteDB(path=db_path)
+
+print(f"Database: {db_path}")
+
+# Create agent
+agent = Agent(
+    name="LocalBot",
+    instructions="You are a helpful assistant.",
+    db=my_db,
+    session_id="local-session"
+)
+
+# Chat
+response = agent.chat("Hello! What can you help me with?")
+print(f"Response: {response}")
+
+# Verify
+data = my_db.export_session("local-session")
+print(f"\nMessages stored: {len(data.get('messages', []))}")
+
+my_db.close()
+print("✅ Done")

examples/persistence/state_redis.py

@@ -0,0 +1,46 @@
+"""
+Redis State Store - Agent-First Example
+
+Use Redis for state management with an Agent.
+
+Docker Setup:
+    docker run -d --name praison-redis -p 6379:6379 redis:7
+
+Run:
+    python state_redis.py
+
+Expected Output:
+    Agent responds with state persisted to Redis
+"""
+
+from praisonaiagents import Agent, db
+
+print("=== Redis State Store (Agent-First) ===")
+
+# Agent-first approach: use db parameter with Redis state store
+my_db = db(
+    database_url="sqlite:///conversations.db",  # Conversations
+    state_url="redis://localhost:6379"          # State/tracing
+)
+
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant.",
+    db=my_db,
+    session_id="redis-state-example"
+)
+
+# Chat - state is automatically managed via Redis
+response = agent.chat("Hello! Remember my favorite color is blue.")
+print(f"Response: {response}")
+
+# Second message to test state
+response2 = agent.chat("What is my favorite color?")
+print(f"Response: {response2}")
+
+my_db.close()
+print("\n=== Demo Complete ===")
+
+# --- Advanced: Direct Store Usage ---
+# from praisonai.persistence.factory import create_state_store
+# store = create_state_store("redis", url="redis://localhost:6379")

examples/python/knowledge/advanced-knowledge-management.py

@@ -15,6 +15,7 @@ Features demonstrated:
 from praisonaiagents import Agent
 import tempfile
 import os
+import shutil
 
 # Create a sample knowledge document for demonstration
 sample_document = """
@@ -91,38 +92,15 @@ doc_path = os.path.join(temp_dir, "ai_healthcare_guide.txt")
 with open(doc_path, "w") as f:
     f.write(sample_document)
 
-# Create an agent with advanced knowledge management configuration
+# Create an agent with knowledge management
 knowledge_agent = Agent(
     name="AdvancedKnowledgeAgent",
     role="Healthcare AI Knowledge Expert", 
     goal="Provide comprehensive answers about AI in healthcare using advanced knowledge retrieval",
     backstory="You are an expert in healthcare AI with access to comprehensive knowledge bases and advanced retrieval capabilities.",
     
-    # Advanced knowledge configuration
+    # Knowledge configuration - simple and working
     knowledge=[doc_path],
-    knowledge_config={
-        "vector_store": {
-            "provider": "chroma",
-            "collection_name": "healthcare_ai_knowledge"
-        },
-        "chunking": {
-            "strategy": "semantic",  # Advanced semantic chunking
-            "chunk_size": 500,       # Optimal chunk size for retrieval
-            "chunk_overlap": 50,     # Overlap for context preservation
-            "separators": ["\n\n", "\n", ".", "!", "?"]  # Smart separators
-        },
-        "retrieval": {
-            "search_type": "similarity",
-            "k": 5,  # Retrieve top 5 most relevant chunks
-            "score_threshold": 0.7,  # Minimum relevance score
-            "rerank": True  # Enable reranking for better results
-        },
-        "embedding": {
-            "provider": "openai",
-            "model": "text-embedding-3-small",  # Efficient embedding model
-            "dimensions": 1536
-        }
-    },
     
     instructions="""You are an expert in healthcare AI. Use the knowledge base to provide 
     comprehensive, accurate answers. Always cite specific information from the documents 
@@ -180,7 +158,6 @@ relationship_result = knowledge_agent.start(
 print(f"Relationship analysis result:\n{relationship_result}\n")
 
 # Clean up temporary files
-import shutil
 shutil.rmtree(temp_dir)
 
 print("="*80)

examples/run_all_examples.py

@@ -0,0 +1,69 @@
+#!/usr/bin/env python3
+"""Run all examples and collect PASS/SKIP/FAIL results."""
+import subprocess
+import sys
+import os
+from pathlib import Path
+
+EXAMPLES_DIR = Path(__file__).parent
+RESULTS = []
+
+def run_example(path: Path) -> tuple:
+    """Run an example and return (name, status, reason)."""
+    name = f"{path.parent.name}/{path.name}"
+    try:
+        result = subprocess.run(
+            [sys.executable, str(path)],
+            capture_output=True,
+            text=True,
+            timeout=60,
+            env={**os.environ, "PYTHONPATH": str(EXAMPLES_DIR.parent / "src")}
+        )
+        output = result.stdout + result.stderr
+        if "PASSED" in output:
+            return (name, "PASS", "")
+        elif "SKIPPED" in output:
+            return (name, "SKIP", output.split("SKIPPED:")[-1].strip()[:50])
+        elif result.returncode != 0:
+            error = output.strip().split("\n")[-1][:80]
+            return (name, "FAIL", error)
+        else:
+            return (name, "PASS", "")
+    except subprocess.TimeoutExpired:
+        return (name, "FAIL", "Timeout")
+    except Exception as e:
+        return (name, "FAIL", str(e)[:50])
+
+def main():
+    print("=" * 70)
+    print("EXAMPLE RUNNER - PASS/SKIP/FAIL SUMMARY")
+    print("=" * 70)
+    
+    # Find all example files
+    examples = []
+    for subdir in ["db", "vector", "observability", "multi_agent", "workflows"]:
+        subpath = EXAMPLES_DIR / subdir
+        if subpath.exists():
+            examples.extend(sorted(subpath.glob("*_wow.py")))
+    
+    print(f"\nFound {len(examples)} examples to run\n")
+    
+    for ex in examples:
+        name, status, reason = run_example(ex)
+        RESULTS.append((name, status, reason))
+        icon = {"PASS": "✅", "SKIP": "⏭️", "FAIL": "❌"}.get(status, "?")
+        reason_str = f" ({reason})" if reason else ""
+        print(f"{icon} {status:4} | {name:45} {reason_str}")
+    
+    # Summary
+    print("\n" + "=" * 70)
+    passed = sum(1 for _, s, _ in RESULTS if s == "PASS")
+    skipped = sum(1 for _, s, _ in RESULTS if s == "SKIP")
+    failed = sum(1 for _, s, _ in RESULTS if s == "FAIL")
+    print(f"SUMMARY: {passed} PASS | {skipped} SKIP | {failed} FAIL | {len(RESULTS)} TOTAL")
+    print("=" * 70)
+    
+    return 0 if failed == 0 else 1
+
+if __name__ == "__main__":
+    sys.exit(main())

examples/tools/external/arxiv/tool.py

@@ -0,0 +1,44 @@
+"""ArXiv Tool Example.
+
+This example demonstrates how to use the ArXiv tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+
+Usage:
+    python tool.py
+
+Note: No API key required!
+"""
+
+from praisonai_tools import ArxivTool
+
+
+def main():
+    # Initialize ArXiv tool (no API key needed)
+    arxiv = ArxivTool()
+    
+    # Example 1: Search ArXiv
+    print("=" * 60)
+    print("Example 1: Search ArXiv")
+    print("=" * 60)
+    
+    results = arxiv.search("transformer neural networks", max_results=3)
+    
+    if results and "error" in results[0]:
+        print(f"Error: {results[0]['error']}")
+    else:
+        print(f"Found {len(results)} papers:")
+        for paper in results:
+            print(f"\n  Title: {paper.get('title', 'N/A')[:60]}...")
+            authors = paper.get('authors', [])
+            if authors:
+                print(f"  Authors: {', '.join(authors[:2])}...")
+            print(f"  Published: {paper.get('published', 'N/A')}")
+            print(f"  PDF: {paper.get('pdf_url', 'N/A')}")
+    
+    print("\n✅ ArXiv tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/calculator/tool.py

@@ -0,0 +1,53 @@
+"""Calculator Tool Example.
+
+This example demonstrates how to use the Calculator tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+
+Usage:
+    python tool.py
+
+Note: No API key required!
+"""
+
+from praisonai_tools import CalculatorTool
+
+
+def main():
+    # Initialize Calculator tool
+    calc = CalculatorTool()
+    
+    # Example 1: Basic arithmetic
+    print("=" * 60)
+    print("Example 1: Basic Arithmetic")
+    print("=" * 60)
+    
+    expressions = [
+        "2 + 2",
+        "10 * 5",
+        "100 / 4",
+        "2 ** 10",
+        "(10 + 5) * 2 - 3"
+    ]
+    
+    for expr in expressions:
+        result = calc.evaluate(expr)
+        print(f"  {expr} = {result.get('result', result)}")
+    
+    # Example 2: Percentage calculation
+    print("\n" + "=" * 60)
+    print("Example 2: Percentage Calculation")
+    print("=" * 60)
+    
+    result = calc.evaluate("250 * 0.15")
+    print(f"  15% of 250 = {result.get('result', result)}")
+    
+    result = calc.evaluate("1000 * (1 + 0.05) ** 3")
+    print(f"  $1000 at 5% for 3 years = ${result.get('result', 0):.2f}")
+    
+    print("\n✅ Calculator tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/duckduckgo/tool.py

@@ -0,0 +1,75 @@
+"""DuckDuckGo Search Tool Example.
+
+This example demonstrates how to use the DuckDuckGo search tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+
+Usage:
+    python tool.py
+
+Note: No API key required!
+"""
+
+from praisonai_tools import DuckDuckGoTool
+
+
+def main():
+    # Initialize DuckDuckGo tool (no API key needed)
+    ddg = DuckDuckGoTool()
+    
+    # Example 1: Web search
+    print("=" * 60)
+    print("Example 1: Web Search")
+    print("=" * 60)
+    
+    results = ddg.search("Python programming best practices", max_results=3)
+    
+    if results and "error" in results[0]:
+        print(f"Error: {results[0]['error']}")
+    else:
+        print(f"Found {len(results)} results:")
+        for r in results:
+            print(f"  - {r['title']}")
+            print(f"    URL: {r['url']}")
+            print(f"    Snippet: {r['snippet'][:100]}...")
+            print()
+    
+    # Example 2: News search
+    print("=" * 60)
+    print("Example 2: News Search")
+    print("=" * 60)
+    
+    news = ddg.news("artificial intelligence", max_results=3)
+    
+    if news and "error" in news[0]:
+        print(f"Error: {news[0]['error']}")
+    else:
+        print(f"Found {len(news)} news articles:")
+        for n in news:
+            print(f"  - {n['title']}")
+            print(f"    Source: {n.get('source', 'N/A')}")
+            print(f"    Date: {n.get('date', 'N/A')}")
+            print()
+    
+    # Example 3: Image search
+    print("=" * 60)
+    print("Example 3: Image Search")
+    print("=" * 60)
+    
+    images = ddg.images("sunset landscape", max_results=3)
+    
+    if images and "error" in images[0]:
+        print(f"Error: {images[0]['error']}")
+    else:
+        print(f"Found {len(images)} images:")
+        for img in images:
+            print(f"  - {img['title']}")
+            print(f"    Image URL: {img['image_url'][:60]}...")
+            print()
+    
+    print("✅ DuckDuckGo tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/exa/tool.py

@@ -0,0 +1,48 @@
+"""Exa Search Tool Example.
+
+This example demonstrates how to use the Exa search tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+    export EXA_API_KEY=your_api_key
+
+Usage:
+    python tool.py
+"""
+
+import os
+from praisonai_tools import ExaTool
+
+
+def main():
+    # Check for API key
+    if not os.getenv("EXA_API_KEY"):
+        print("Error: EXA_API_KEY environment variable not set")
+        print("Get your API key from https://exa.ai/")
+        return
+    
+    # Initialize Exa tool
+    exa = ExaTool()
+    
+    # Example 1: Neural search
+    print("=" * 60)
+    print("Example 1: Neural Search")
+    print("=" * 60)
+    
+    results = exa.search("best practices for Python development", num_results=3)
+    
+    if results and "error" in results[0]:
+        print(f"Error: {results[0]['error']}")
+    else:
+        print(f"Found {len(results)} results:")
+        for r in results:
+            print(f"  - {r.get('title', 'N/A')[:50]}...")
+            print(f"    URL: {r.get('url', 'N/A')[:60]}...")
+            print(f"    Score: {r.get('score', 'N/A')}")
+            print()
+    
+    print("✅ Exa tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/firecrawl/tool.py

@@ -0,0 +1,45 @@
+"""Firecrawl Tool Example.
+
+This example demonstrates how to use the Firecrawl tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+    export FIRECRAWL_API_KEY=your_api_key
+
+Usage:
+    python tool.py
+"""
+
+import os
+from praisonai_tools import FirecrawlTool
+
+
+def main():
+    # Check for API key
+    if not os.getenv("FIRECRAWL_API_KEY"):
+        print("Error: FIRECRAWL_API_KEY environment variable not set")
+        print("Get your API key from https://firecrawl.dev/")
+        return
+    
+    # Initialize Firecrawl tool
+    firecrawl = FirecrawlTool()
+    
+    # Example 1: Scrape a page
+    print("=" * 60)
+    print("Example 1: Scrape a Page")
+    print("=" * 60)
+    
+    result = firecrawl.scrape("https://example.com")
+    
+    if "error" in result:
+        print(f"Error: {result['error']}")
+    else:
+        print(f"URL: {result.get('url', 'N/A')}")
+        markdown = result.get('markdown', '')
+        print(f"Content preview: {markdown[:300]}...")
+    
+    print("\n✅ Firecrawl tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/hackernews/tool.py

@@ -0,0 +1,42 @@
+"""Hacker News Tool Example.
+
+This example demonstrates how to use the Hacker News tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+
+Usage:
+    python tool.py
+
+Note: No API key required!
+"""
+
+from praisonai_tools import HackerNewsTool
+
+
+def main():
+    # Initialize Hacker News tool
+    hn = HackerNewsTool()
+    
+    # Example 1: Get top stories
+    print("=" * 60)
+    print("Example 1: Top Stories")
+    print("=" * 60)
+    
+    stories = hn.get_top_stories(limit=5)
+    
+    if stories and "error" in stories[0]:
+        print(f"Error: {stories[0]['error']}")
+    else:
+        print(f"Found {len(stories)} top stories:")
+        for story in stories:
+            print(f"  - {story.get('title', 'N/A')[:50]}...")
+            print(f"    Score: {story.get('score', 'N/A')} | Comments: {story.get('descendants', 'N/A')}")
+            print(f"    URL: {story.get('url', 'N/A')[:60]}...")
+            print()
+    
+    print("✅ Hacker News tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/jina/tool.py

@@ -0,0 +1,40 @@
+"""Jina Reader Tool Example.
+
+This example demonstrates how to use the Jina Reader tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+
+Usage:
+    python tool.py
+
+Note: Works without API key (with rate limits)
+"""
+
+from praisonai_tools import JinaTool
+
+
+def main():
+    # Initialize Jina tool (works without API key)
+    jina = JinaTool()
+    
+    # Example 1: Read a URL
+    print("=" * 60)
+    print("Example 1: Read URL Content")
+    print("=" * 60)
+    
+    result = jina.read("https://example.com")
+    
+    if "error" in result:
+        print(f"Error: {result['error']}")
+    else:
+        print(f"URL: {result.get('url', 'N/A')}")
+        print(f"Title: {result.get('title', 'N/A')}")
+        content = result.get('content', '')
+        print(f"Content preview: {content[:300]}...")
+    
+    print("\n✅ Jina tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/postgres/tool.py

@@ -0,0 +1,56 @@
+"""PostgreSQL Tool Example.
+
+This example demonstrates how to use the PostgreSQL tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+    Docker: docker run -d --name postgres -e POSTGRES_PASSWORD=praison123 -e POSTGRES_DB=praisonai -p 5432:5432 postgres:16
+
+Usage:
+    python tool.py
+"""
+
+from praisonai_tools import PostgresTool
+
+
+def main():
+    # Initialize PostgreSQL tool
+    pg = PostgresTool(
+        host="localhost",
+        port=5432,
+        database="praisonai",
+        user="postgres",
+        password="praison123"
+    )
+    
+    # Example 1: List tables
+    print("=" * 60)
+    print("Example 1: List Tables")
+    print("=" * 60)
+    
+    tables = pg.list_tables()
+    
+    if isinstance(tables, dict) and "error" in tables:
+        print(f"Error: {tables['error']}")
+    else:
+        print(f"Found {len(tables)} tables:")
+        for t in tables[:10]:
+            print(f"  - {t.get('table_name', t)}")
+    
+    # Example 2: Query
+    print("\n" + "=" * 60)
+    print("Example 2: Execute Query")
+    print("=" * 60)
+    
+    result = pg.query("SELECT version()")
+    
+    if isinstance(result, dict) and "error" in result:
+        print(f"Error: {result['error']}")
+    else:
+        print(f"PostgreSQL version: {result}")
+    
+    print("\n✅ PostgreSQL tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/serper/tool.py

@@ -0,0 +1,47 @@
+"""Serper Tool Example.
+
+This example demonstrates how to use the Serper (Google Search) tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+    export SERPER_API_KEY=your_api_key
+
+Usage:
+    python tool.py
+"""
+
+import os
+from praisonai_tools import SerperTool
+
+
+def main():
+    # Check for API key
+    if not os.getenv("SERPER_API_KEY"):
+        print("Error: SERPER_API_KEY environment variable not set")
+        print("Get your API key from https://serper.dev/")
+        return
+    
+    # Initialize Serper tool
+    serper = SerperTool()
+    
+    # Example 1: Web search
+    print("=" * 60)
+    print("Example 1: Google Web Search")
+    print("=" * 60)
+    
+    results = serper.search("Python programming tutorials", max_results=3)
+    
+    if results and "error" in results[0]:
+        print(f"Error: {results[0]['error']}")
+    else:
+        print(f"Found {len(results)} results:")
+        for r in results:
+            print(f"  - {r.get('title', 'N/A')[:50]}...")
+            print(f"    URL: {r.get('link', 'N/A')[:60]}...")
+            print()
+    
+    print("✅ Serper tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/tavily/tool.py

@@ -0,0 +1,57 @@
+"""Tavily Search Tool Example.
+
+This example demonstrates how to use the Tavily search tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+    export TAVILY_API_KEY=your_api_key
+
+Usage:
+    python tool.py
+"""
+
+import os
+from praisonai_tools import TavilyTool
+
+def main():
+    # Check for API key
+    if not os.getenv("TAVILY_API_KEY"):
+        print("Error: TAVILY_API_KEY environment variable not set")
+        print("Get your API key from https://tavily.com/")
+        return
+    
+    # Initialize Tavily tool
+    tavily = TavilyTool()
+    
+    # Example 1: Basic search
+    print("=" * 60)
+    print("Example 1: Basic Search")
+    print("=" * 60)
+    
+    results = tavily.search("What is quantum computing?", max_results=3)
+    
+    if "error" in results:
+        print(f"Error: {results['error']}")
+    else:
+        print(f"Query: {results.get('query')}")
+        if "answer" in results:
+            print(f"\nAI Answer: {results['answer'][:200]}...")
+        print(f"\nResults ({len(results.get('results', []))}):")
+        for r in results.get("results", []):
+            print(f"  - {r['title']}")
+            print(f"    URL: {r['url']}")
+            print(f"    Score: {r.get('score', 'N/A')}")
+    
+    # Example 2: Search context for RAG
+    print("\n" + "=" * 60)
+    print("Example 2: Search Context (for RAG)")
+    print("=" * 60)
+    
+    context = tavily.search_context("Python best practices 2024")
+    print(f"Context (first 300 chars): {context[:300]}...")
+    
+    print("\n✅ Tavily tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/tools/external/wikipedia/tool.py

@@ -0,0 +1,69 @@
+"""Wikipedia Tool Example.
+
+This example demonstrates how to use the Wikipedia tool with PraisonAI agents.
+
+Requirements:
+    pip install "praisonai[tools]"
+
+Usage:
+    python tool.py
+
+Note: No API key required!
+"""
+
+from praisonai_tools import WikipediaTool
+
+
+def main():
+    # Initialize Wikipedia tool (no API key needed)
+    wiki = WikipediaTool()
+    
+    # Example 1: Search Wikipedia
+    print("=" * 60)
+    print("Example 1: Search Wikipedia")
+    print("=" * 60)
+    
+    results = wiki.search("Python programming", max_results=5)
+    
+    if results and "error" in results[0]:
+        print(f"Error: {results[0]['error']}")
+    else:
+        print(f"Found {len(results)} results:")
+        for r in results:
+            print(f"  - {r['title']}")
+    
+    # Example 2: Get page summary
+    print("\n" + "=" * 60)
+    print("Example 2: Get Page Summary")
+    print("=" * 60)
+    
+    summary = wiki.summary("Artificial intelligence", sentences=3)
+    
+    if "error" in summary:
+        print(f"Error: {summary['error']}")
+    else:
+        print(f"Title: {summary['title']}")
+        print(f"Summary: {summary['summary'][:300]}...")
+    
+    # Example 3: Get full page
+    print("\n" + "=" * 60)
+    print("Example 3: Get Full Page")
+    print("=" * 60)
+    
+    page = wiki.get_page("Machine learning")
+    
+    if "error" in page:
+        print(f"Error: {page['error']}")
+        if "options" in page:
+            print(f"Options: {page['options'][:5]}")
+    else:
+        print(f"Title: {page['title']}")
+        print(f"URL: {page['url']}")
+        print(f"Categories: {page['categories'][:3]}")
+        print(f"Content preview: {page['content'][:200]}...")
+    
+    print("\n✅ Wikipedia tool working correctly!")
+
+
+if __name__ == "__main__":
+    main()

examples/vector/chroma_wow.py

@@ -0,0 +1,23 @@
+"""ChromaDB Vector Store - Agent-First Example"""
+from praisonaiagents import Agent
+
+# Agent-first approach: use knowledge parameter with ChromaDB
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant with access to documents.",
+    knowledge=["./docs/guide.pdf"],  # Add your documents here
+    knowledge_config={
+        "vector_store": "chroma",
+        "path": "/tmp/chroma_demo"
+    }
+)
+
+# Chat - agent uses knowledge for RAG
+response = agent.chat("What information do you have?")
+print(f"Response: {response}")
+
+print("PASSED: ChromaDB with Agent")
+
+# --- Advanced: Direct Store Usage ---
+# from praisonai.persistence import create_knowledge_store
+# store = create_knowledge_store("chroma", path="/tmp/chroma_demo")

examples/vector/lancedb_real_wow.py

@@ -0,0 +1,25 @@
+"""LanceDB Vector Store - Agent-First Example"""
+from praisonaiagents import Agent
+
+# Agent-first approach: use knowledge parameter with LanceDB
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant with access to documents.",
+    knowledge=["./docs/guide.pdf"],  # Add your documents here
+    knowledge_config={
+        "vector_store": "lancedb",
+        "path": "/tmp/lancedb_test"
+    }
+)
+
+# Chat - agent uses knowledge for RAG
+response = agent.chat("What information do you have?")
+print(f"Response: {response}")
+
+print("PASSED: LanceDB with Agent")
+
+# --- Advanced: Direct LanceDB Usage ---
+# import lancedb
+# db = lancedb.connect("/tmp/lancedb_test")
+# data = [{"id": "1", "text": "ML is AI", "vector": [0.1] * 128}]
+# table = db.create_table("demo", data)

examples/vector/lancedb_wow.py

@@ -0,0 +1,23 @@
+"""LanceDB Vector Store - Agent-First Example"""
+from praisonaiagents import Agent
+
+# Agent-first approach: use knowledge parameter with LanceDB
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant with access to documents.",
+    knowledge=["./docs/guide.pdf"],  # Add your documents here
+    knowledge_config={
+        "vector_store": "lancedb",
+        "path": "/tmp/lancedb_demo"
+    }
+)
+
+# Chat - agent uses knowledge for RAG
+response = agent.chat("What information do you have?")
+print(f"Response: {response}")
+
+print("PASSED: LanceDB with Agent")
+
+# --- Advanced: Direct Store Usage ---
+# from praisonai.persistence import create_knowledge_store
+# store = create_knowledge_store("lancedb", path="/tmp/lancedb_demo")

examples/vector/pgvector_real_wow.py

@@ -0,0 +1,31 @@
+"""PGVector Vector Store - Agent-First Example
+
+Docker: docker run -d --name pgvector -e POSTGRES_PASSWORD=postgres -p 5433:5432 pgvector/pgvector:pg16
+"""
+import os
+from praisonaiagents import Agent
+
+url = os.getenv("PGVECTOR_URL", "postgresql://postgres:postgres@localhost:5433/postgres")
+
+# Agent-first approach: use knowledge parameter with PGVector
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant with access to documents.",
+    knowledge=["./docs/guide.pdf"],  # Add your documents here
+    knowledge_config={
+        "vector_store": "pgvector",
+        "url": url
+    }
+)
+
+# Chat - agent uses knowledge for RAG
+response = agent.chat("What information do you have?")
+print(f"Response: {response}")
+
+print("PASSED: PGVector with Agent")
+
+# --- Advanced: Direct PGVector Usage ---
+# import psycopg2
+# conn = psycopg2.connect(url)
+# cur = conn.cursor()
+# cur.execute("CREATE EXTENSION IF NOT EXISTS vector")

examples/vector/pgvector_wow.py

@@ -0,0 +1,29 @@
+"""PGVector Knowledge Store - Agent-First Example (requires Docker)
+
+Docker: docker run -d --name pgvector -e POSTGRES_PASSWORD=postgres -p 5433:5432 pgvector/pgvector:pg16
+"""
+import os
+from praisonaiagents import Agent
+
+url = os.getenv("PGVECTOR_URL", "postgresql://postgres:postgres@localhost:5433/postgres")
+
+# Agent-first approach: use knowledge parameter with PGVector
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant with access to documents.",
+    knowledge=["./docs/guide.pdf"],  # Add your documents here
+    knowledge_config={
+        "vector_store": "pgvector",
+        "url": url
+    }
+)
+
+# Chat - agent uses knowledge for RAG
+response = agent.chat("What information do you have?")
+print(f"Response: {response}")
+
+print("PASSED: PGVector with Agent")
+
+# --- Advanced: Direct Store Usage ---
+# from praisonai.persistence import create_knowledge_store
+# store = create_knowledge_store("pgvector", url=url)

examples/vector/pinecone_wow.py

@@ -0,0 +1,34 @@
+"""Pinecone Vector Store - Agent-First Example
+
+Requires: export PINECONE_API_KEY=...
+"""
+import os
+import sys
+from praisonaiagents import Agent
+
+api_key = os.getenv("PINECONE_API_KEY")
+if not api_key:
+    print("SKIPPED: Pinecone - PINECONE_API_KEY not set")
+    sys.exit(0)
+
+# Agent-first approach: use knowledge parameter with Pinecone
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant with access to documents.",
+    knowledge=["./docs/guide.pdf"],  # Add your documents here
+    knowledge_config={
+        "vector_store": "pinecone",
+        "api_key": api_key,
+        "environment": "us-east-1"
+    }
+)
+
+# Chat - agent uses knowledge for RAG
+response = agent.chat("What information do you have?")
+print(f"Response: {response}")
+
+print("PASSED: Pinecone with Agent")
+
+# --- Advanced: Direct Store Usage ---
+# from pinecone import Pinecone
+# pc = Pinecone(api_key=api_key)

examples/vector/qdrant_wow.py

@@ -0,0 +1,29 @@
+"""Qdrant Vector Store - Agent-First Example (requires Docker)
+
+Docker: docker run -d --name qdrant -p 6333:6333 qdrant/qdrant
+"""
+import os
+from praisonaiagents import Agent
+
+# Agent-first approach: use knowledge parameter with Qdrant
+url = os.getenv("QDRANT_URL", "http://localhost:6333")
+
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant with access to documents.",
+    knowledge=["./docs/guide.pdf"],  # Add your documents here
+    knowledge_config={
+        "vector_store": "qdrant",
+        "url": url
+    }
+)
+
+# Chat - agent uses knowledge for RAG
+response = agent.chat("What information do you have?")
+print(f"Response: {response}")
+
+print("PASSED: Qdrant with Agent")
+
+# --- Advanced: Direct Store Usage ---
+# from praisonai.persistence import create_knowledge_store
+# store = create_knowledge_store("qdrant", url=url)

examples/vector/weaviate_wow.py

@@ -0,0 +1,36 @@
+"""Weaviate Cloud Vector Store - Agent-First Example
+
+Requires: export WEAVIATE_URL=... and WEAVIATE_API_KEY=...
+"""
+import os
+import sys
+from praisonaiagents import Agent
+
+weaviate_url = os.getenv("WEAVIATE_URL", "https://your-cluster.weaviate.cloud")
+weaviate_key = os.getenv("WEAVIATE_API_KEY")
+if not weaviate_key:
+    print("SKIPPED: Weaviate - WEAVIATE_API_KEY not set")
+    sys.exit(0)
+
+# Agent-first approach: use knowledge parameter with Weaviate
+agent = Agent(
+    name="Assistant",
+    instructions="You are a helpful assistant with access to documents.",
+    knowledge=["./docs/guide.pdf"],  # Add your documents here
+    knowledge_config={
+        "vector_store": "weaviate",
+        "url": weaviate_url,
+        "api_key": weaviate_key
+    }
+)
+
+# Chat - agent uses knowledge for RAG
+response = agent.chat("What information do you have?")
+print(f"Response: {response}")
+
+print("PASSED: Weaviate with Agent")
+
+# --- Advanced: Direct Store Usage ---
+# import weaviate
+# from weaviate.classes.init import Auth
+# client = weaviate.connect_to_weaviate_cloud(cluster_url=weaviate_url, auth_credentials=Auth.api_key(weaviate_key))

src/praisonai-agents/.gitignore

@@ -62,6 +62,8 @@ docs/generated/
 *.log
 *.sqlite
 *.db
+# But keep the db module directory
+!praisonaiagents/db/
 
 # OS generated files
 .DS_Store

src/praisonai-agents/praisonaiagents/__init__.py

@@ -43,8 +43,7 @@ from .tools.base import BaseTool, ToolResult, ToolValidationError, validate_tool
 from .tools.decorator import tool, FunctionTool
 from .tools.registry import get_registry, register_tool, get_tool, ToolRegistry
 from .agents.autoagents import AutoAgents
-from .knowledge.knowledge import Knowledge
-from .knowledge.chunking import Chunking
+# Knowledge is lazy-loaded to avoid importing heavy deps (chromadb, mem0) at startup
 # MCP support (optional)
 try:
     from .mcp.mcp import MCP
@@ -53,7 +52,9 @@ except ImportError:
     _mcp_available = False
     MCP = None
 from .session import Session
-from .memory.memory import Memory
+# Memory is lazy-loaded to avoid importing chromadb at startup
+from .db import db
+from .obs import obs
 # Workflows - import from dedicated workflows module
 from .workflows import (
     Workflow, WorkflowStep, WorkflowContext, StepResult,
@@ -65,8 +66,16 @@ from .guardrails import GuardrailResult, LLMGuardrail
 
 # Fast Context support (lazy loaded to avoid performance impact)
 def __getattr__(name):
-    """Lazy load FastContext to avoid impacting package load time."""
-    if name == "FastContext":
+    """Lazy load heavy modules to avoid impacting package load time."""
+    # Knowledge module (imports chromadb, mem0)
+    if name == "Knowledge":
+        from praisonaiagents.knowledge.knowledge import Knowledge
+        return Knowledge
+    elif name == "Chunking":
+        from praisonaiagents.knowledge.chunking import Chunking
+        return Chunking
+    # FastContext support
+    elif name == "FastContext":
         from praisonaiagents.context.fast import FastContext
         return FastContext
     elif name == "FastContextResult":
@@ -91,6 +100,10 @@ def __getattr__(name):
     elif name == "SkillLoader":
         from praisonaiagents.skills import SkillLoader
         return SkillLoader
+    # Memory module (imports chromadb)
+    elif name == "Memory":
+        from praisonaiagents.memory.memory import Memory
+        return Memory
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
 
 # Planning mode support
@@ -257,6 +270,8 @@ __all__ = [
     'AutoAgents',
     'Session',
     'Memory',
+    'db',
+    'obs',
     'display_interaction',
     'display_self_reflection',
     'display_instruction',

src/praisonai-agents/praisonaiagents/agent/agent.py

@@ -252,7 +252,12 @@ class Agent:
         history_in_context: Optional[int] = None,
         auto_save: Optional[str] = None,
         skills: Optional[List[str]] = None,
-        skills_dirs: Optional[List[str]] = None
+        skills_dirs: Optional[List[str]] = None,
+        db: Optional[Any] = None,
+        session_id: Optional[str] = None,
+        hooks: Optional[List[Any]] = None,
+        llm_config: Optional[Dict[str, Any]] = None,
+        rate_limiter: Optional[Any] = None
     ):
         """Initialize an Agent instance.
 
@@ -393,6 +398,17 @@ class Agent:
         # Flag to track if final result has been displayed to prevent duplicates
         self._final_display_shown = False
         
+        # Store hooks for middleware system (zero overhead when empty)
+        self._hooks = hooks or []
+        self._middleware_manager = None  # Lazy init
+        
+        # Store llm_config for configurable model switching
+        self._llm_config = llm_config or {}
+        self._llm_configurable = self._llm_config.get('configurable', False)
+        
+        # Store rate limiter (optional, zero overhead when None)
+        self._rate_limiter = rate_limiter
+        
         # Store OpenAI client parameters for lazy initialization
         self._openai_api_key = api_key
         self._openai_base_url = base_url
@@ -615,6 +631,11 @@ Your Goal: {self.goal}
         self._skill_manager = None  # Lazy loaded
         self._skills_initialized = False
 
+        # Database persistence (lazy - no imports until used)
+        self._db = db
+        self._session_id = session_id
+        self._db_initialized = False
+
     @property
     def console(self):
         """Lazily initialize Rich Console only when needed."""
@@ -1615,8 +1636,32 @@ Your Goal: {self.goal}"""
     def execute_tool(self, function_name, arguments):
         """
         Execute a tool dynamically based on the function name and arguments.
+        Injects agent state for tools with Injected[T] parameters.
         """
         logging.debug(f"{self.name} executing tool {function_name} with arguments: {arguments}")
+        
+        # Set up injection context for tools with Injected parameters
+        from ..tools.injected import AgentState, with_injection_context
+        state = AgentState(
+            agent_id=self.name,
+            run_id=getattr(self, '_current_run_id', 'unknown'),
+            session_id=getattr(self, '_session_id', None) or 'default',
+            last_user_message=self.chat_history[-1].get('content') if self.chat_history else None,
+            metadata={'agent_name': self.name}
+        )
+        
+        # Execute within injection context
+        return self._execute_tool_with_context(function_name, arguments, state)
+    
+    def _execute_tool_with_context(self, function_name, arguments, state):
+        """Execute tool within injection context."""
+        from ..tools.injected import with_injection_context
+        
+        with with_injection_context(state):
+            return self._execute_tool_impl(function_name, arguments)
+    
+    def _execute_tool_impl(self, function_name, arguments):
+        """Internal tool execution implementation."""
 
         # Check if approval is required for this tool
         from ..approval import is_approval_required, console_approval_callback, get_risk_level, mark_approved, ApprovalDecision, get_approval_callback
@@ -1957,7 +2002,112 @@ Your Goal: {self.goal}"""
         #         expand=False
         #     )
 
-    def chat(self, prompt, temperature=1.0, tools=None, output_json=None, output_pydantic=None, reasoning_steps=False, stream=None, task_name=None, task_description=None, task_id=None):
+    def _init_db_session(self):
+        """Initialize DB session if db adapter is provided (lazy, first chat only)."""
+        if self._db is None or self._db_initialized:
+            return
+        
+        # Generate session_id if not provided: default to per-hour ID (YYYYMMDDHH-agentname)
+        if self._session_id is None:
+            import hashlib
+            from datetime import datetime, timezone
+            # Per-hour session ID: YYYYMMDDHH (UTC) + agent name hash for uniqueness
+            hour_str = datetime.now(timezone.utc).strftime("%Y%m%d%H")
+            agent_hash = hashlib.md5(self.name.encode()).hexdigest()[:6]
+            self._session_id = f"{hour_str}-{agent_hash}"
+        
+        # Call db adapter's on_agent_start to get previous messages
+        try:
+            history = self._db.on_agent_start(
+                agent_name=self.name,
+                session_id=self._session_id,
+                user_id=self.user_id,
+                metadata={"role": self.role, "goal": self.goal}
+            )
+            
+            # Restore chat history from previous session
+            if history:
+                for msg in history:
+                    self.chat_history.append({
+                        "role": msg.role,
+                        "content": msg.content
+                    })
+                logging.info(f"Resumed session {self._session_id} with {len(history)} messages")
+        except Exception as e:
+            logging.warning(f"Failed to initialize DB session: {e}")
+        
+        self._db_initialized = True
+        self._current_run_id = None  # Track current run
+
+    def _start_run(self, input_content: str):
+        """Start a new run (turn) for persistence tracking."""
+        if self._db is None:
+            return
+        
+        import uuid
+        self._current_run_id = f"run-{uuid.uuid4().hex[:12]}"
+        
+        try:
+            if hasattr(self._db, 'on_run_start'):
+                self._db.on_run_start(
+                    session_id=self._session_id,
+                    run_id=self._current_run_id,
+                    input_content=input_content,
+                    metadata={"agent_name": self.name}
+                )
+        except Exception as e:
+            logging.warning(f"Failed to start run: {e}")
+
+    def _end_run(self, output_content: str, status: str = "completed", metrics: dict = None):
+        """End the current run (turn)."""
+        if self._db is None or self._current_run_id is None:
+            return
+        
+        try:
+            if hasattr(self._db, 'on_run_end'):
+                self._db.on_run_end(
+                    session_id=self._session_id,
+                    run_id=self._current_run_id,
+                    output_content=output_content,
+                    status=status,
+                    metrics=metrics or {},
+                    metadata={"agent_name": self.name}
+                )
+        except Exception as e:
+            logging.warning(f"Failed to end run: {e}")
+        
+        self._current_run_id = None
+
+    def _persist_message(self, role: str, content: str):
+        """Persist a message to the DB if adapter is provided."""
+        if self._db is None:
+            return
+        
+        try:
+            if role == "user":
+                self._db.on_user_message(self._session_id, content)
+            elif role == "assistant":
+                self._db.on_agent_message(self._session_id, content)
+        except Exception as e:
+            logging.warning(f"Failed to persist message: {e}")
+
+    @property
+    def session_id(self) -> Optional[str]:
+        """Get the current session ID."""
+        return self._session_id
+
+    def chat(self, prompt, temperature=1.0, tools=None, output_json=None, output_pydantic=None, reasoning_steps=False, stream=None, task_name=None, task_description=None, task_id=None, config=None):
+        # Apply rate limiter if configured (before any LLM call)
+        if self._rate_limiter is not None:
+            self._rate_limiter.acquire()
+        
+        # Initialize DB session on first chat (lazy)
+        self._init_db_session()
+        
+        # Start a new run for this chat turn
+        prompt_str = prompt if isinstance(prompt, str) else str(prompt)
+        self._start_run(prompt_str)
+        
         # Reset the final display flag for each new conversation
         self._final_display_shown = False
         
@@ -2054,6 +2204,8 @@ Your Goal: {self.goal}"""
                         self.chat_history[-1].get("content") == normalized_content):
                     # Add user message to chat history BEFORE LLM call so handoffs can access it
                     self.chat_history.append({"role": "user", "content": normalized_content})
+                    # Persist user message to DB
+                    self._persist_message("user", normalized_content)
                 
                 try:
                     # Pass everything to LLM class
@@ -2083,6 +2235,8 @@ Your Goal: {self.goal}"""
                     )
 
                     self.chat_history.append({"role": "assistant", "content": response_text})
+                    # Persist assistant message to DB
+                    self._persist_message("assistant", response_text)
 
                     # Log completion time if in debug mode
                     if logging.getLogger().getEffectiveLevel() == logging.DEBUG:
@@ -2127,6 +2281,8 @@ Your Goal: {self.goal}"""
                     self.chat_history[-1].get("content") == normalized_content):
                 # Add user message to chat history BEFORE LLM call so handoffs can access it
                 self.chat_history.append({"role": "user", "content": normalized_content})
+                # Persist user message to DB (OpenAI path)
+                self._persist_message("user", normalized_content)
 
             reflection_count = 0
             start_time = time.time()
@@ -2166,6 +2322,8 @@ Your Goal: {self.goal}"""
                             # Add to chat history and return raw response
                             # User message already added before LLM call via _build_messages
                             self.chat_history.append({"role": "assistant", "content": response_text})
+                            # Persist assistant message to DB
+                            self._persist_message("assistant", response_text)
                             # Apply guardrail validation even for JSON output
                             try:
                                 validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
@@ -2181,6 +2339,8 @@ Your Goal: {self.goal}"""
                         if not self.self_reflect:
                             # User message already added before LLM call via _build_messages
                             self.chat_history.append({"role": "assistant", "content": response_text})
+                            # Persist assistant message to DB (non-reflect path)
+                            self._persist_message("assistant", response_text)
                             if self.verbose:
                                 logging.debug(f"Agent {self.name} final response: {response_text}")
                             # Return only reasoning content if reasoning_steps is True
@@ -2269,11 +2429,13 @@ Output MUST be JSON with 'reflection' and 'satisfactory'.
                                     validated_response = self._apply_guardrail_with_retry(response_text, original_prompt, temperature, tools, task_name, task_description, task_id)
                                     # Execute callback after validation
                                     self._execute_callback_and_display(original_prompt, validated_response, time.time() - start_time, task_name, task_description, task_id)
+                                    self._end_run(validated_response, "completed", {"duration_ms": (time.time() - start_time) * 1000})
                                     return validated_response
                                 except Exception as e:
                                     logging.error(f"Agent {self.name}: Guardrail validation failed after reflection: {e}")
                                     # Rollback chat history on guardrail failure
                                     self.chat_history = self.chat_history[:chat_history_length]
+                                    self._end_run(None, "error", {"error": str(e)})
                                     return None
 
                             # Check if we've hit max reflections

src/praisonai-agents/praisonaiagents/db/__init__.py

@@ -0,0 +1,85 @@
+"""
+Database adapter interface for PraisonAI Agents.
+
+This module provides the protocol and types for database persistence.
+Implementations are provided by the wrapper layer (praisonai.db).
+
+Usage (simplest - recommended):
+    from praisonaiagents import Agent, db
+    
+    agent = Agent(
+        name="Assistant",
+        db=db(database_url="postgresql://localhost/mydb"),  # db(...) shortcut
+        session_id="my-session"  # optional: defaults to per-hour ID
+    )
+    agent.chat("Hello!")  # auto-persists messages, runs, traces, tool calls
+
+Alternative (explicit backend):
+    db=db.PraisonDB(database_url="...")  # Auto-detect backend
+    db=db.PostgresDB(host="localhost")   # PostgreSQL
+    db=db.SQLiteDB(path="data.db")       # SQLite
+    db=db.RedisDB(host="localhost")      # Redis (state only)
+"""
+
+from .protocol import (
+    DbAdapter,
+    AsyncDbAdapter,
+    DbMessage,
+    DbToolCall,
+    DbRun,
+    DbSpan,
+    DbTrace,
+)
+
+
+class _LazyDbModule:
+    """
+    Lazy proxy for database backends.
+    
+    Allows `from praisonaiagents import db` without importing heavy deps.
+    Actual backend classes are loaded only when accessed.
+    """
+    
+    _BACKENDS = {
+        "PraisonDB": "praisonai.db.adapter",
+        "PostgresDB": "praisonai.db.adapter",
+        "SQLiteDB": "praisonai.db.adapter",
+        "RedisDB": "praisonai.db.adapter",
+    }
+    
+    def __getattr__(self, name: str):
+        if name in self._BACKENDS:
+            try:
+                import importlib
+                module = importlib.import_module(self._BACKENDS[name])
+                return getattr(module, name)
+            except ImportError as e:
+                raise ImportError(
+                    f"Database backend '{name}' requires the praisonai package. "
+                    f"Install with: pip install praisonai\n"
+                    f"Original error: {e}"
+                ) from e
+        raise AttributeError(f"module 'db' has no attribute {name!r}")
+    
+    def __repr__(self):
+        return "<module 'praisonaiagents.db' (lazy database backends)>"
+    
+    def __call__(self, **kwargs):
+        """Shortcut: db(...) is equivalent to db.PraisonDB(...)"""
+        return self.PraisonDB(**kwargs)
+
+
+# Singleton lazy module instance
+db = _LazyDbModule()
+
+
+__all__ = [
+    "DbAdapter",
+    "AsyncDbAdapter",
+    "DbMessage",
+    "DbToolCall",
+    "DbRun",
+    "DbSpan",
+    "DbTrace",
+    "db",
+]

src/praisonai-agents/praisonaiagents/db/protocol.py

@@ -0,0 +1,464 @@
+"""
+Database adapter protocol for PraisonAI Agents.
+
+This module uses only stdlib (typing, dataclasses) to define the interface.
+No external dependencies are imported here.
+
+Implementations are provided by the wrapper layer (praisonai.db).
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Protocol, runtime_checkable
+import time
+
+
+@dataclass
+class DbToolCall:
+    """A tool call to be persisted."""
+    tool_name: str
+    args: Dict[str, Any]
+    result: Any = None
+    call_id: Optional[str] = None
+    timestamp: float = field(default_factory=time.time)
+    duration_ms: Optional[float] = None
+    error: Optional[str] = None
+
+
+@dataclass
+class DbMessage:
+    """A message to be persisted."""
+    role: str  # "user", "assistant", "system", "tool"
+    content: str
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    timestamp: float = field(default_factory=time.time)
+    id: Optional[str] = None
+    tool_calls: Optional[List["DbToolCall"]] = None  # For assistant messages with tool calls
+    run_id: Optional[str] = None  # Group messages by run
+
+
+@dataclass
+class DbRun:
+    """A single agent run (turn) to be persisted."""
+    run_id: str
+    session_id: str
+    agent_name: str
+    started_at: float = field(default_factory=time.time)
+    ended_at: Optional[float] = None
+    status: str = "running"  # running, completed, error, cancelled
+    input_content: Optional[str] = None
+    output_content: Optional[str] = None
+    messages: List[DbMessage] = field(default_factory=list)
+    tool_calls: List[DbToolCall] = field(default_factory=list)
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    metrics: Dict[str, Any] = field(default_factory=dict)  # tokens, latency, etc.
+
+
+@dataclass
+class DbSpan:
+    """A span for tracing/observability."""
+    span_id: str
+    trace_id: str
+    name: str
+    started_at: float = field(default_factory=time.time)
+    ended_at: Optional[float] = None
+    parent_span_id: Optional[str] = None
+    status: str = "running"  # running, ok, error
+    attributes: Dict[str, Any] = field(default_factory=dict)
+    events: List[Dict[str, Any]] = field(default_factory=list)
+
+
+@dataclass
+class DbTrace:
+    """A trace for observability (groups spans)."""
+    trace_id: str
+    session_id: Optional[str] = None
+    run_id: Optional[str] = None
+    agent_name: Optional[str] = None
+    user_id: Optional[str] = None
+    started_at: float = field(default_factory=time.time)
+    ended_at: Optional[float] = None
+    status: str = "running"  # running, ok, error
+    spans: List[DbSpan] = field(default_factory=list)
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+
+@runtime_checkable
+class DbAdapter(Protocol):
+    """
+    Protocol for database adapters that provide persistence for agents.
+    
+    Implementations must provide these methods. The agent will call them
+    at appropriate lifecycle points if a db adapter is provided.
+    
+    All methods are synchronous. Async support can be added via wrapper.
+    """
+    
+    def on_agent_start(
+        self,
+        agent_name: str,
+        session_id: str,
+        user_id: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> List[DbMessage]:
+        """
+        Called when agent starts a session.
+        
+        Args:
+            agent_name: Name of the agent
+            session_id: Session identifier
+            user_id: Optional user identifier
+            metadata: Optional session metadata
+            
+        Returns:
+            List of previous messages for this session (for resume)
+        """
+        ...
+    
+    def on_user_message(
+        self,
+        session_id: str,
+        content: str,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when user sends a message.
+        
+        Args:
+            session_id: Session identifier
+            content: Message content
+            metadata: Optional message metadata
+        """
+        ...
+    
+    def on_agent_message(
+        self,
+        session_id: str,
+        content: str,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when agent produces a response.
+        
+        Args:
+            session_id: Session identifier
+            content: Response content
+            metadata: Optional message metadata
+        """
+        ...
+    
+    def on_tool_call(
+        self,
+        session_id: str,
+        tool_name: str,
+        args: Dict[str, Any],
+        result: Any,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when a tool is executed.
+        
+        Args:
+            session_id: Session identifier
+            tool_name: Name of the tool
+            args: Tool arguments
+            result: Tool result
+            metadata: Optional call metadata
+        """
+        ...
+    
+    def on_agent_end(
+        self,
+        session_id: str,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when agent session ends.
+        
+        Args:
+            session_id: Session identifier
+            metadata: Optional end metadata
+        """
+        ...
+    
+    def on_run_start(
+        self,
+        session_id: str,
+        run_id: str,
+        input_content: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when a new run (turn) starts.
+        
+        Args:
+            session_id: Session identifier
+            run_id: Run identifier
+            input_content: User input for this run
+            metadata: Optional run metadata
+        """
+        ...
+    
+    def on_run_end(
+        self,
+        session_id: str,
+        run_id: str,
+        output_content: Optional[str] = None,
+        status: str = "completed",
+        metrics: Optional[Dict[str, Any]] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when a run (turn) ends.
+        
+        Args:
+            session_id: Session identifier
+            run_id: Run identifier
+            output_content: Agent output for this run
+            status: Run status (completed, error, cancelled)
+            metrics: Optional metrics (tokens, latency)
+            metadata: Optional run metadata
+        """
+        ...
+    
+    def get_runs(
+        self,
+        session_id: str,
+        limit: Optional[int] = None,
+    ) -> List["DbRun"]:
+        """
+        Get runs for a session.
+        
+        Args:
+            session_id: Session identifier
+            limit: Maximum number of runs to return
+            
+        Returns:
+            List of runs for this session
+        """
+        ...
+    
+    def export_session(
+        self,
+        session_id: str,
+    ) -> Dict[str, Any]:
+        """
+        Export a session as a dictionary (for JSONL export).
+        
+        Args:
+            session_id: Session identifier
+            
+        Returns:
+            Dictionary with session data, messages, runs
+        """
+        ...
+    
+    def import_session(
+        self,
+        data: Dict[str, Any],
+    ) -> str:
+        """
+        Import a session from a dictionary.
+        
+        Args:
+            data: Session data dictionary
+            
+        Returns:
+            Session ID of imported session
+        """
+        ...
+    
+    # --- Tracing/Observability ---
+    
+    def on_trace_start(
+        self,
+        trace_id: str,
+        session_id: Optional[str] = None,
+        run_id: Optional[str] = None,
+        agent_name: Optional[str] = None,
+        user_id: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when a new trace starts.
+        
+        Args:
+            trace_id: Unique trace identifier
+            session_id: Optional session identifier
+            run_id: Optional run identifier
+            agent_name: Optional agent name
+            user_id: Optional user identifier
+            metadata: Optional trace metadata
+        """
+        ...
+    
+    def on_trace_end(
+        self,
+        trace_id: str,
+        status: str = "ok",
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when a trace ends.
+        
+        Args:
+            trace_id: Trace identifier
+            status: Trace status (ok, error)
+            metadata: Optional end metadata
+        """
+        ...
+    
+    def on_span_start(
+        self,
+        span_id: str,
+        trace_id: str,
+        name: str,
+        parent_span_id: Optional[str] = None,
+        attributes: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when a new span starts.
+        
+        Args:
+            span_id: Unique span identifier
+            trace_id: Parent trace identifier
+            name: Span name (e.g., "llm_call", "tool_execution")
+            parent_span_id: Optional parent span for nesting
+            attributes: Optional span attributes
+        """
+        ...
+    
+    def on_span_end(
+        self,
+        span_id: str,
+        status: str = "ok",
+        attributes: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """
+        Called when a span ends.
+        
+        Args:
+            span_id: Span identifier
+            status: Span status (ok, error)
+            attributes: Optional end attributes
+        """
+        ...
+    
+    def get_traces(
+        self,
+        session_id: Optional[str] = None,
+        user_id: Optional[str] = None,
+        limit: Optional[int] = None,
+    ) -> List["DbTrace"]:
+        """
+        Get traces with optional filters.
+        
+        Args:
+            session_id: Filter by session
+            user_id: Filter by user
+            limit: Maximum number of traces
+            
+        Returns:
+            List of traces
+        """
+        ...
+    
+    def close(self) -> None:
+        """Close the database connection."""
+        ...
+
+
+@runtime_checkable
+class AsyncDbAdapter(Protocol):
+    """
+    Async protocol for database adapters.
+    
+    Provides async variants of all DbAdapter methods for use with
+    async/await patterns and async backends like asyncpg.
+    """
+    
+    async def on_agent_start(
+        self,
+        agent_name: str,
+        session_id: str,
+        user_id: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> List["DbMessage"]:
+        """Async version of on_agent_start."""
+        ...
+    
+    async def on_user_message(
+        self,
+        session_id: str,
+        content: str,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Async version of on_user_message."""
+        ...
+    
+    async def on_agent_message(
+        self,
+        session_id: str,
+        content: str,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Async version of on_agent_message."""
+        ...
+    
+    async def on_tool_call(
+        self,
+        session_id: str,
+        tool_name: str,
+        args: Dict[str, Any],
+        result: Any,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Async version of on_tool_call."""
+        ...
+    
+    async def on_agent_end(
+        self,
+        session_id: str,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Async version of on_agent_end."""
+        ...
+    
+    async def on_run_start(
+        self,
+        session_id: str,
+        run_id: str,
+        input_content: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Async version of on_run_start."""
+        ...
+    
+    async def on_run_end(
+        self,
+        session_id: str,
+        run_id: str,
+        output_content: Optional[str] = None,
+        status: str = "completed",
+        metrics: Optional[Dict[str, Any]] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Async version of on_run_end."""
+        ...
+    
+    async def export_session(
+        self,
+        session_id: str,
+    ) -> Dict[str, Any]:
+        """Async version of export_session."""
+        ...
+    
+    async def import_session(
+        self,
+        data: Dict[str, Any],
+    ) -> str:
+        """Async version of import_session."""
+        ...
+    
+    async def close(self) -> None:
+        """Async close the database connection."""
+        ...

src/praisonai-agents/praisonaiagents/db/tests/__init__.py

@@ -0,0 +1 @@
+"""TDD tests for db integration."""

src/praisonai-agents/praisonaiagents/db/tests/test_db_integration.py

@@ -0,0 +1,246 @@
+"""
+TDD tests for PraisonAI Agents DB integration.
+
+Tests:
+1. Lazy import verification - no heavy deps on import
+2. Agent(db=MockDb) calls hook lifecycle in correct order
+3. Session resume injects history
+4. db=None behavior unchanged
+"""
+
+import sys
+import time
+
+
+def test_lazy_import_no_heavy_deps():
+    """Test that importing praisonaiagents.db does not import DB libs."""
+    print("=== Test: Lazy Import Verification ===")
+    
+    # Record modules before import
+    before_modules = set(sys.modules.keys())
+    
+    # Import only the db module (not full praisonaiagents which loads Knowledge)
+    from praisonaiagents.db import DbAdapter, DbMessage
+    
+    # Check that heavy DB libs are NOT loaded by db module itself
+    # Note: praisonaiagents main package loads qdrant/chroma via Knowledge
+    # but the db module should not add any additional heavy deps
+    heavy_libs = ['psycopg2', 'pymysql', 'redis']  # DB-specific libs
+    loaded_heavy = [lib for lib in heavy_libs if lib in sys.modules]
+    
+    if loaded_heavy:
+        print(f"  ❌ Heavy DB libs loaded by db module: {loaded_heavy}")
+        return False
+    
+    print("  ✅ No heavy DB libs loaded by praisonaiagents.db module")
+    return True
+
+
+def test_agent_accepts_db_parameter():
+    """Test that Agent accepts db and session_id parameters."""
+    print("\n=== Test: Agent Accepts db Parameter ===")
+    
+    import inspect
+    from praisonaiagents import Agent
+    
+    sig = inspect.signature(Agent.__init__)
+    params = list(sig.parameters.keys())
+    
+    if 'db' not in params:
+        print("  ❌ Agent does not accept 'db' parameter")
+        return False
+    
+    if 'session_id' not in params:
+        print("  ❌ Agent does not accept 'session_id' parameter")
+        return False
+    
+    print("  ✅ Agent accepts db and session_id parameters")
+    return True
+
+
+class MockDbAdapter:
+    """Mock DB adapter for testing hook lifecycle."""
+    
+    def __init__(self):
+        self.calls = []
+        self.history_to_return = []
+    
+    def on_agent_start(self, agent_name, session_id, user_id=None, metadata=None):
+        self.calls.append(('on_agent_start', agent_name, session_id))
+        return self.history_to_return
+    
+    def on_user_message(self, session_id, content, metadata=None):
+        self.calls.append(('on_user_message', session_id, content))
+    
+    def on_agent_message(self, session_id, content, metadata=None):
+        self.calls.append(('on_agent_message', session_id, content))
+    
+    def on_tool_call(self, session_id, tool_name, args, result, metadata=None):
+        self.calls.append(('on_tool_call', session_id, tool_name))
+    
+    def on_agent_end(self, session_id, metadata=None):
+        self.calls.append(('on_agent_end', session_id))
+    
+    def close(self):
+        self.calls.append(('close',))
+
+
+def test_db_hook_lifecycle():
+    """Test that Agent calls db hooks in correct order."""
+    print("\n=== Test: DB Hook Lifecycle ===")
+    
+    from praisonaiagents import Agent
+    
+    mock_db = MockDbAdapter()
+    
+    # Create agent with mock db
+    agent = Agent(
+        name="TestAgent",
+        instructions="Test agent",
+        db=mock_db,
+        session_id="test-session-001",
+        verbose=False
+    )
+    
+    # Verify db is stored
+    if agent._db is not mock_db:
+        print("  ❌ Agent did not store db adapter")
+        return False
+    
+    if agent._session_id != "test-session-001":
+        print("  ❌ Agent did not store session_id")
+        return False
+    
+    print("  ✅ Agent stores db and session_id correctly")
+    return True
+
+
+def test_db_none_unchanged_behavior():
+    """Test that db=None behavior is unchanged."""
+    print("\n=== Test: db=None Unchanged Behavior ===")
+    
+    from praisonaiagents import Agent
+    
+    # Create agent without db
+    agent = Agent(
+        name="TestAgent",
+        instructions="Test agent",
+        verbose=False
+    )
+    
+    # Verify db is None
+    if agent._db is not None:
+        print("  ❌ Agent._db should be None when not provided")
+        return False
+    
+    # Verify _init_db_session does nothing when db is None
+    agent._init_db_session()
+    
+    if agent._db_initialized:
+        print("  ❌ _db_initialized should remain False when db is None")
+        return False
+    
+    print("  ✅ db=None behavior unchanged")
+    return True
+
+
+def test_session_id_auto_generation():
+    """Test that session_id is auto-generated if not provided."""
+    print("\n=== Test: Session ID Auto-Generation ===")
+    
+    from praisonaiagents import Agent
+    
+    mock_db = MockDbAdapter()
+    
+    agent = Agent(
+        name="TestAgent",
+        instructions="Test agent",
+        db=mock_db,
+        # No session_id provided
+        verbose=False
+    )
+    
+    # Trigger db initialization
+    agent._init_db_session()
+    
+    if agent._session_id is None:
+        print("  ❌ session_id should be auto-generated")
+        return False
+    
+    if not agent._session_id.startswith("session-"):
+        print(f"  ❌ session_id format unexpected: {agent._session_id}")
+        return False
+    
+    print(f"  ✅ session_id auto-generated: {agent._session_id}")
+    return True
+
+
+def test_protocol_compliance():
+    """Test that DbAdapter protocol is properly defined."""
+    print("\n=== Test: DbAdapter Protocol Compliance ===")
+    
+    from praisonaiagents.db.protocol import DbAdapter, DbMessage
+    
+    # Check DbMessage fields
+    msg = DbMessage(role="user", content="test")
+    if msg.role != "user" or msg.content != "test":
+        print("  ❌ DbMessage fields not working")
+        return False
+    
+    # Check protocol methods
+    required_methods = [
+        'on_agent_start',
+        'on_user_message', 
+        'on_agent_message',
+        'on_tool_call',
+        'on_agent_end',
+        'close'
+    ]
+    
+    mock = MockDbAdapter()
+    for method in required_methods:
+        if not hasattr(mock, method):
+            print(f"  ❌ MockDbAdapter missing {method}")
+            return False
+    
+    # Check protocol isinstance
+    if not isinstance(mock, DbAdapter):
+        print("  ❌ MockDbAdapter should implement DbAdapter protocol")
+        return False
+    
+    print("  ✅ DbAdapter protocol properly defined")
+    return True
+
+
+def run_all_tests():
+    """Run all TDD tests."""
+    print("=" * 60)
+    print("PraisonAI Agents DB Integration - TDD Tests")
+    print("=" * 60)
+    
+    results = []
+    
+    results.append(("Lazy Import", test_lazy_import_no_heavy_deps()))
+    results.append(("Agent db Parameter", test_agent_accepts_db_parameter()))
+    results.append(("DB Hook Lifecycle", test_db_hook_lifecycle()))
+    results.append(("db=None Unchanged", test_db_none_unchanged_behavior()))
+    results.append(("Session ID Auto-Gen", test_session_id_auto_generation()))
+    results.append(("Protocol Compliance", test_protocol_compliance()))
+    
+    print("\n" + "=" * 60)
+    passed = sum(1 for _, r in results if r)
+    total = len(results)
+    print(f"Results: {passed}/{total} tests passed")
+    
+    for name, result in results:
+        status = "✅" if result else "❌"
+        print(f"  {status} {name}")
+    
+    print("=" * 60)
+    
+    return passed == total
+
+
+if __name__ == "__main__":
+    success = run_all_tests()
+    sys.exit(0 if success else 1)

src/praisonai-agents/praisonaiagents/eval/accuracy.py

@@ -89,11 +89,9 @@ class AccuracyEvaluator(BaseEvaluator):
             EvaluationScore with score and reasoning
         """
         try:
-            from openai import OpenAI
+            import litellm
         except ImportError:
-            raise ImportError("OpenAI package required for accuracy evaluation. Install with: pip install openai")
-        
-        client = OpenAI()
+            raise ImportError("litellm package required for accuracy evaluation. Install with: pip install litellm")
         
         prompt = f"""You are an expert evaluator. Compare the actual output against the expected output and provide a score from 1-10.
 
@@ -115,7 +113,8 @@ Respond in this exact format:
 SCORE: [number 1-10]
 REASONING: [brief explanation]"""
 
-        response = client.chat.completions.create(
+        # Use litellm for unified multi-provider support
+        response = litellm.completion(
             model=self.model,
             messages=[{"role": "user", "content": prompt}],
             temperature=0.1,

src/praisonai-agents/praisonaiagents/eval/criteria.py

@@ -99,11 +99,9 @@ class CriteriaEvaluator(BaseEvaluator):
             CriteriaScore with score, passed status, and reasoning
         """
         try:
-            from openai import OpenAI
+            import litellm
         except ImportError:
-            raise ImportError("OpenAI package required for criteria evaluation. Install with: pip install openai")
-        
-        client = OpenAI()
+            raise ImportError("litellm package required for criteria evaluation. Install with: pip install litellm")
         
         if self.scoring_type == "binary":
             prompt = f"""You are an expert evaluator. Evaluate the following output against the given criteria.
@@ -138,7 +136,8 @@ Respond in this exact format:
 SCORE: [number 1-10]
 REASONING: [brief explanation]"""
 
-        response = client.chat.completions.create(
+        # Use litellm for unified multi-provider support
+        response = litellm.completion(
             model=self.model,
             messages=[{"role": "user", "content": prompt}],
             temperature=0.1,

src/praisonai-agents/praisonaiagents/hooks/__init__.py

@@ -65,6 +65,23 @@ __all__ = [
     "AfterAgentInput",
     "SessionStartInput",
     "SessionEndInput",
+    # Middleware types
+    "InvocationContext",
+    "ModelRequest",
+    "ModelResponse",
+    "ToolRequest",
+    "ToolResponse",
+    # Middleware decorators
+    "before_model",
+    "after_model",
+    "wrap_model_call",
+    "before_tool",
+    "after_tool",
+    "wrap_tool_call",
+    # Middleware utilities
+    "MiddlewareChain",
+    "AsyncMiddlewareChain",
+    "MiddlewareManager",
 ]
 
 
@@ -94,4 +111,27 @@ def __getattr__(name: str):
         )
         return locals()[name]
     
+    # Middleware types
+    if name in ("InvocationContext", "ModelRequest", "ModelResponse", 
+                "ToolRequest", "ToolResponse"):
+        from .middleware import (
+            InvocationContext, ModelRequest, ModelResponse,
+            ToolRequest, ToolResponse
+        )
+        return locals()[name]
+    
+    # Middleware decorators
+    if name in ("before_model", "after_model", "wrap_model_call",
+                "before_tool", "after_tool", "wrap_tool_call"):
+        from .middleware import (
+            before_model, after_model, wrap_model_call,
+            before_tool, after_tool, wrap_tool_call
+        )
+        return locals()[name]
+    
+    # Middleware utilities
+    if name in ("MiddlewareChain", "AsyncMiddlewareChain", "MiddlewareManager"):
+        from .middleware import MiddlewareChain, AsyncMiddlewareChain, MiddlewareManager
+        return locals()[name]
+    
     raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

src/praisonai-agents/praisonaiagents/hooks/middleware.py

@@ -0,0 +1,439 @@
+"""
+Middleware System for PraisonAI Agents.
+
+Provides wrap_model_call and wrap_tool_call patterns for intercepting
+and modifying agent behavior at model/tool call boundaries.
+
+Zero Performance Impact:
+- All middleware is optional
+- Empty middleware chain is O(1) fast path
+- No overhead when middleware not registered
+
+Usage:
+    from praisonaiagents.hooks import before_model, wrap_model_call, wrap_tool_call
+    
+    @before_model
+    def add_context(request):
+        request.messages.append({"role": "system", "content": "Extra"})
+        return request
+    
+    @wrap_model_call
+    def retry_on_error(request, call_next):
+        for _ in range(3):
+            try:
+                return call_next(request)
+            except Exception:
+                pass
+        raise RuntimeError("All retries failed")
+    
+    agent = Agent(name="Test", hooks=[add_context, retry_on_error])
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Callable, Dict, List, Optional, TypeVar, Generic
+from functools import wraps
+
+
+@dataclass
+class InvocationContext:
+    """Context passed through middleware chain.
+    
+    Contains identifiers for multi-agent safety and optional metadata.
+    """
+    agent_id: str
+    run_id: str
+    session_id: str
+    tool_name: Optional[str] = None
+    model_name: Optional[str] = None
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class ModelRequest:
+    """Request data for model calls."""
+    messages: List[Dict[str, Any]]
+    model: str
+    temperature: float = 1.0
+    context: Optional[InvocationContext] = None
+    tools: Optional[List[Any]] = None
+    extra: Dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class ModelResponse:
+    """Response data from model calls."""
+    content: str
+    model: str
+    usage: Optional[Dict[str, int]] = None
+    context: Optional[InvocationContext] = None
+    tool_calls: Optional[List[Any]] = None
+    extra: Dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class ToolRequest:
+    """Request data for tool calls."""
+    tool_name: str
+    arguments: Dict[str, Any]
+    context: Optional[InvocationContext] = None
+    extra: Dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class ToolResponse:
+    """Response data from tool calls."""
+    tool_name: str
+    result: Any
+    error: Optional[str] = None
+    context: Optional[InvocationContext] = None
+    extra: Dict[str, Any] = field(default_factory=dict)
+
+
+# Type aliases for middleware functions
+T = TypeVar('T')
+BeforeModelFn = Callable[[ModelRequest], ModelRequest]
+AfterModelFn = Callable[[ModelResponse], ModelResponse]
+WrapModelCallFn = Callable[[ModelRequest, Callable[[ModelRequest], ModelResponse]], ModelResponse]
+BeforeToolFn = Callable[[ToolRequest], ToolRequest]
+AfterToolFn = Callable[[ToolResponse], ToolResponse]
+WrapToolCallFn = Callable[[ToolRequest, Callable[[ToolRequest], ToolResponse]], ToolResponse]
+
+
+class MiddlewareChain:
+    """Synchronous middleware chain executor.
+    
+    Composes middleware functions in registration order.
+    First registered middleware wraps all subsequent ones.
+    """
+    
+    def __init__(self, middlewares: Optional[List[Callable]] = None):
+        self._middlewares = middlewares or []
+    
+    def execute(self, request: Any, final_handler: Callable[[Any], Any]) -> Any:
+        """Execute the middleware chain with a final handler.
+        
+        Args:
+            request: The request object to pass through
+            final_handler: The actual operation to perform
+            
+        Returns:
+            The result from the chain
+        """
+        if not self._middlewares:
+            return final_handler(request)
+        
+        # Build the chain from inside out
+        # Last middleware wraps final_handler
+        # First middleware is outermost
+        def build_chain(index: int) -> Callable[[Any], Any]:
+            if index >= len(self._middlewares):
+                return final_handler
+            
+            middleware = self._middlewares[index]
+            next_handler = build_chain(index + 1)
+            
+            def handler(req: Any) -> Any:
+                return middleware(req, next_handler)
+            
+            return handler
+        
+        chain = build_chain(0)
+        return chain(request)
+    
+    def __len__(self) -> int:
+        return len(self._middlewares)
+
+
+class AsyncMiddlewareChain:
+    """Asynchronous middleware chain executor.
+    
+    Same as MiddlewareChain but for async middleware and handlers.
+    """
+    
+    def __init__(self, middlewares: Optional[List[Callable]] = None):
+        self._middlewares = middlewares or []
+    
+    async def execute(self, request: Any, final_handler: Callable) -> Any:
+        """Execute the async middleware chain.
+        
+        Args:
+            request: The request object to pass through
+            final_handler: The actual async operation to perform
+            
+        Returns:
+            The result from the chain
+        """
+        if not self._middlewares:
+            return await final_handler(request)
+        
+        # Build the chain from inside out
+        def build_chain(index: int) -> Callable:
+            if index >= len(self._middlewares):
+                return final_handler
+            
+            middleware = self._middlewares[index]
+            next_handler = build_chain(index + 1)
+            
+            async def handler(req: Any) -> Any:
+                return await middleware(req, next_handler)
+            
+            return handler
+        
+        chain = build_chain(0)
+        return await chain(request)
+    
+    def __len__(self) -> int:
+        return len(self._middlewares)
+
+
+# Decorator functions for tagging middleware
+
+def before_model(func: BeforeModelFn) -> BeforeModelFn:
+    """Decorator to mark a function as a before_model hook.
+    
+    The function receives a ModelRequest and should return a (possibly modified) ModelRequest.
+    
+    Example:
+        @before_model
+        def add_context(request):
+            request.messages.append({"role": "system", "content": "Extra"})
+            return request
+    """
+    func._hook_type = 'before_model'
+    return func
+
+
+def after_model(func: AfterModelFn) -> AfterModelFn:
+    """Decorator to mark a function as an after_model hook.
+    
+    The function receives a ModelResponse and should return a (possibly modified) ModelResponse.
+    
+    Example:
+        @after_model
+        def log_response(response):
+            print(f"Model returned: {response.content[:50]}")
+            return response
+    """
+    func._hook_type = 'after_model'
+    return func
+
+
+def wrap_model_call(func: WrapModelCallFn) -> WrapModelCallFn:
+    """Decorator to mark a function as a wrap_model_call middleware.
+    
+    The function receives (request, call_next) and should call call_next(request)
+    to continue the chain, or return early to short-circuit.
+    
+    Example:
+        @wrap_model_call
+        def retry_on_error(request, call_next):
+            for _ in range(3):
+                try:
+                    return call_next(request)
+                except Exception:
+                    pass
+            raise RuntimeError("All retries failed")
+    """
+    func._hook_type = 'wrap_model_call'
+    return func
+
+
+def before_tool(func: BeforeToolFn) -> BeforeToolFn:
+    """Decorator to mark a function as a before_tool hook.
+    
+    The function receives a ToolRequest and should return a (possibly modified) ToolRequest.
+    
+    Example:
+        @before_tool
+        def validate_args(request):
+            if 'dangerous' in request.arguments:
+                raise ValueError("Dangerous argument detected")
+            return request
+    """
+    func._hook_type = 'before_tool'
+    return func
+
+
+def after_tool(func: AfterToolFn) -> AfterToolFn:
+    """Decorator to mark a function as an after_tool hook.
+    
+    The function receives a ToolResponse and should return a (possibly modified) ToolResponse.
+    
+    Example:
+        @after_tool
+        def log_result(response):
+            print(f"Tool {response.tool_name} returned: {response.result}")
+            return response
+    """
+    func._hook_type = 'after_tool'
+    return func
+
+
+def wrap_tool_call(func: WrapToolCallFn) -> WrapToolCallFn:
+    """Decorator to mark a function as a wrap_tool_call middleware.
+    
+    The function receives (request, call_next) and should call call_next(request)
+    to continue the chain, or return early to short-circuit.
+    
+    Example:
+        @wrap_tool_call
+        def retry_flaky_tool(request, call_next):
+            last_error = None
+            for _ in range(3):
+                try:
+                    return call_next(request)
+                except Exception as e:
+                    last_error = e
+            raise last_error
+    """
+    func._hook_type = 'wrap_tool_call'
+    return func
+
+
+def get_hook_type(func: Callable) -> Optional[str]:
+    """Get the hook type of a decorated function."""
+    return getattr(func, '_hook_type', None)
+
+
+def is_middleware(func: Callable) -> bool:
+    """Check if a function is a middleware (wrap_* type)."""
+    hook_type = get_hook_type(func)
+    return hook_type in ('wrap_model_call', 'wrap_tool_call')
+
+
+def categorize_hooks(hooks: List[Callable]) -> Dict[str, List[Callable]]:
+    """Categorize hooks by their type.
+    
+    Returns:
+        Dict with keys: before_model, after_model, wrap_model_call,
+                       before_tool, after_tool, wrap_tool_call
+    """
+    categories = {
+        'before_model': [],
+        'after_model': [],
+        'wrap_model_call': [],
+        'before_tool': [],
+        'after_tool': [],
+        'wrap_tool_call': [],
+    }
+    
+    for hook in hooks or []:
+        hook_type = get_hook_type(hook)
+        if hook_type and hook_type in categories:
+            categories[hook_type].append(hook)
+    
+    return categories
+
+
+class MiddlewareManager:
+    """Manages middleware for an agent.
+    
+    Provides methods to execute before/after hooks and middleware chains
+    for both model and tool calls.
+    """
+    
+    def __init__(self, hooks: Optional[List[Callable]] = None):
+        self._hooks = hooks or []
+        self._categorized = categorize_hooks(self._hooks)
+        
+        # Pre-build chains for efficiency
+        self._model_chain = MiddlewareChain(self._categorized['wrap_model_call'])
+        self._tool_chain = MiddlewareChain(self._categorized['wrap_tool_call'])
+    
+    @property
+    def has_model_hooks(self) -> bool:
+        """Check if any model hooks are registered."""
+        return bool(
+            self._categorized['before_model'] or
+            self._categorized['after_model'] or
+            self._categorized['wrap_model_call']
+        )
+    
+    @property
+    def has_tool_hooks(self) -> bool:
+        """Check if any tool hooks are registered."""
+        return bool(
+            self._categorized['before_tool'] or
+            self._categorized['after_tool'] or
+            self._categorized['wrap_tool_call']
+        )
+    
+    def run_before_model(self, request: ModelRequest) -> ModelRequest:
+        """Run all before_model hooks."""
+        for hook in self._categorized['before_model']:
+            request = hook(request)
+        return request
+    
+    def run_after_model(self, response: ModelResponse) -> ModelResponse:
+        """Run all after_model hooks (in reverse order)."""
+        for hook in reversed(self._categorized['after_model']):
+            response = hook(response)
+        return response
+    
+    def run_before_tool(self, request: ToolRequest) -> ToolRequest:
+        """Run all before_tool hooks."""
+        for hook in self._categorized['before_tool']:
+            request = hook(request)
+        return request
+    
+    def run_after_tool(self, response: ToolResponse) -> ToolResponse:
+        """Run all after_tool hooks (in reverse order)."""
+        for hook in reversed(self._categorized['after_tool']):
+            response = hook(response)
+        return response
+    
+    def execute_model_call(
+        self,
+        request: ModelRequest,
+        model_fn: Callable[[ModelRequest], ModelResponse]
+    ) -> ModelResponse:
+        """Execute a model call with all hooks and middleware.
+        
+        Order:
+        1. before_model hooks (in order)
+        2. wrap_model_call middleware chain
+        3. actual model call
+        4. after_model hooks (in reverse order)
+        """
+        # Fast path: no hooks
+        if not self.has_model_hooks:
+            return model_fn(request)
+        
+        # Run before hooks
+        request = self.run_before_model(request)
+        
+        # Build wrapped handler that includes after hooks
+        def wrapped_model_fn(req: ModelRequest) -> ModelResponse:
+            response = model_fn(req)
+            return self.run_after_model(response)
+        
+        # Execute through middleware chain
+        return self._model_chain.execute(request, wrapped_model_fn)
+    
+    def execute_tool_call(
+        self,
+        request: ToolRequest,
+        tool_fn: Callable[[ToolRequest], ToolResponse]
+    ) -> ToolResponse:
+        """Execute a tool call with all hooks and middleware.
+        
+        Order:
+        1. before_tool hooks (in order)
+        2. wrap_tool_call middleware chain
+        3. actual tool call
+        4. after_tool hooks (in reverse order)
+        """
+        # Fast path: no hooks
+        if not self.has_tool_hooks:
+            return tool_fn(request)
+        
+        # Run before hooks
+        request = self.run_before_tool(request)
+        
+        # Build wrapped handler that includes after hooks
+        def wrapped_tool_fn(req: ToolRequest) -> ToolResponse:
+            response = tool_fn(req)
+            return self.run_after_tool(response)
+        
+        # Execute through middleware chain
+        return self._tool_chain.execute(request, wrapped_tool_fn)

src/praisonai-agents/praisonaiagents/knowledge/__init__.py

@@ -1,8 +1,130 @@
 """
-PraisonAI Knowledge - Advanced knowledge management system with configurable features
+PraisonAI Knowledge - Advanced knowledge management system with configurable features.
+
+This module provides:
+- Document readers (ReaderProtocol, ReaderRegistry)
+- Vector stores (VectorStoreProtocol, VectorStoreRegistry)
+- Retrieval strategies (RetrieverProtocol, RetrievalStrategy)
+- Rerankers (RerankerProtocol, RerankerRegistry)
+- Index types (IndexProtocol, IndexType)
+- Query engines (QueryEngineProtocol, QueryMode)
 """
 
+# Core Knowledge class (always available)
 from praisonaiagents.knowledge.knowledge import Knowledge
 from praisonaiagents.knowledge.chunking import Chunking
 
-__all__ = ["Knowledge", "Chunking"] 
+# Lazy loading for protocols and registries to avoid import overhead
+_LAZY_IMPORTS = {
+    # Readers
+    "Document": ("praisonaiagents.knowledge.readers", "Document"),
+    "ReaderProtocol": ("praisonaiagents.knowledge.readers", "ReaderProtocol"),
+    "ReaderRegistry": ("praisonaiagents.knowledge.readers", "ReaderRegistry"),
+    "get_reader_registry": ("praisonaiagents.knowledge.readers", "get_reader_registry"),
+    "detect_source_kind": ("praisonaiagents.knowledge.readers", "detect_source_kind"),
+    
+    # Vector stores
+    "VectorRecord": ("praisonaiagents.knowledge.vector_store", "VectorRecord"),
+    "VectorStoreProtocol": ("praisonaiagents.knowledge.vector_store", "VectorStoreProtocol"),
+    "VectorStoreRegistry": ("praisonaiagents.knowledge.vector_store", "VectorStoreRegistry"),
+    "get_vector_store_registry": ("praisonaiagents.knowledge.vector_store", "get_vector_store_registry"),
+    "InMemoryVectorStore": ("praisonaiagents.knowledge.vector_store", "InMemoryVectorStore"),
+    
+    # Retrieval
+    "RetrievalResult": ("praisonaiagents.knowledge.retrieval", "RetrievalResult"),
+    "RetrievalStrategy": ("praisonaiagents.knowledge.retrieval", "RetrievalStrategy"),
+    "RetrieverProtocol": ("praisonaiagents.knowledge.retrieval", "RetrieverProtocol"),
+    "RetrieverRegistry": ("praisonaiagents.knowledge.retrieval", "RetrieverRegistry"),
+    "get_retriever_registry": ("praisonaiagents.knowledge.retrieval", "get_retriever_registry"),
+    "reciprocal_rank_fusion": ("praisonaiagents.knowledge.retrieval", "reciprocal_rank_fusion"),
+    "merge_adjacent_chunks": ("praisonaiagents.knowledge.retrieval", "merge_adjacent_chunks"),
+    
+    # Rerankers
+    "RerankResult": ("praisonaiagents.knowledge.rerankers", "RerankResult"),
+    "RerankerProtocol": ("praisonaiagents.knowledge.rerankers", "RerankerProtocol"),
+    "RerankerRegistry": ("praisonaiagents.knowledge.rerankers", "RerankerRegistry"),
+    "get_reranker_registry": ("praisonaiagents.knowledge.rerankers", "get_reranker_registry"),
+    "SimpleReranker": ("praisonaiagents.knowledge.rerankers", "SimpleReranker"),
+    
+    # Index
+    "IndexType": ("praisonaiagents.knowledge.index", "IndexType"),
+    "IndexStats": ("praisonaiagents.knowledge.index", "IndexStats"),
+    "IndexProtocol": ("praisonaiagents.knowledge.index", "IndexProtocol"),
+    "IndexRegistry": ("praisonaiagents.knowledge.index", "IndexRegistry"),
+    "get_index_registry": ("praisonaiagents.knowledge.index", "get_index_registry"),
+    "KeywordIndex": ("praisonaiagents.knowledge.index", "KeywordIndex"),
+    
+    # Query engine
+    "QueryMode": ("praisonaiagents.knowledge.query_engine", "QueryMode"),
+    "QueryResult": ("praisonaiagents.knowledge.query_engine", "QueryResult"),
+    "QueryEngineProtocol": ("praisonaiagents.knowledge.query_engine", "QueryEngineProtocol"),
+    "QueryEngineRegistry": ("praisonaiagents.knowledge.query_engine", "QueryEngineRegistry"),
+    "get_query_engine_registry": ("praisonaiagents.knowledge.query_engine", "get_query_engine_registry"),
+    "decompose_question": ("praisonaiagents.knowledge.query_engine", "decompose_question"),
+    "SimpleQueryEngine": ("praisonaiagents.knowledge.query_engine", "SimpleQueryEngine"),
+    "SubQuestionEngine": ("praisonaiagents.knowledge.query_engine", "SubQuestionEngine"),
+}
+
+
+def __getattr__(name: str):
+    """Lazy load protocols and registries."""
+    if name in _LAZY_IMPORTS:
+        module_path, attr_name = _LAZY_IMPORTS[name]
+        import importlib
+        module = importlib.import_module(module_path)
+        return getattr(module, attr_name)
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
+
+
+def __dir__():
+    """List available attributes."""
+    return list(_LAZY_IMPORTS.keys()) + ["Knowledge", "Chunking"]
+
+
+__all__ = [
+    # Core
+    "Knowledge",
+    "Chunking",
+    # Readers
+    "Document",
+    "ReaderProtocol",
+    "ReaderRegistry",
+    "get_reader_registry",
+    "detect_source_kind",
+    # Vector stores
+    "VectorRecord",
+    "VectorStoreProtocol",
+    "VectorStoreRegistry",
+    "get_vector_store_registry",
+    "InMemoryVectorStore",
+    # Retrieval
+    "RetrievalResult",
+    "RetrievalStrategy",
+    "RetrieverProtocol",
+    "RetrieverRegistry",
+    "get_retriever_registry",
+    "reciprocal_rank_fusion",
+    "merge_adjacent_chunks",
+    # Rerankers
+    "RerankResult",
+    "RerankerProtocol",
+    "RerankerRegistry",
+    "get_reranker_registry",
+    "SimpleReranker",
+    # Index
+    "IndexType",
+    "IndexStats",
+    "IndexProtocol",
+    "IndexRegistry",
+    "get_index_registry",
+    "KeywordIndex",
+    # Query engine
+    "QueryMode",
+    "QueryResult",
+    "QueryEngineProtocol",
+    "QueryEngineRegistry",
+    "get_query_engine_registry",
+    "decompose_question",
+    "SimpleQueryEngine",
+    "SubQuestionEngine",
+] 

src/praisonai-agents/praisonaiagents/knowledge/index.py

@@ -0,0 +1,332 @@
+"""
+Index Type Selection for PraisonAI Agents.
+
+This module defines index types and protocols.
+NO heavy imports - only stdlib and typing.
+
+Implementations are provided by the wrapper layer.
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Callable, Dict, List, Optional, Protocol, runtime_checkable
+from enum import Enum
+import logging
+import math
+
+logger = logging.getLogger(__name__)
+
+
+class IndexType(str, Enum):
+    """Available index types."""
+    VECTOR = "vector"
+    KEYWORD = "keyword"
+    HYBRID = "hybrid"
+    GRAPH = "graph"
+
+
+@dataclass
+class IndexStats:
+    """Statistics about an index."""
+    index_type: IndexType
+    document_count: int
+    chunk_count: int
+    total_tokens: int = 0
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+
+@runtime_checkable
+class IndexProtocol(Protocol):
+    """
+    Protocol for document indices.
+    
+    Implementations must provide methods for indexing and querying.
+    """
+    
+    name: str
+    index_type: IndexType
+    
+    def add_documents(
+        self,
+        texts: List[str],
+        metadatas: Optional[List[Dict[str, Any]]] = None,
+        ids: Optional[List[str]] = None,
+        embeddings: Optional[List[List[float]]] = None
+    ) -> List[str]:
+        """
+        Add documents to the index.
+        
+        Args:
+            texts: List of text content
+            metadatas: Optional list of metadata dicts
+            ids: Optional list of IDs
+            embeddings: Optional pre-computed embeddings (for vector index)
+            
+        Returns:
+            List of document IDs
+        """
+        ...
+    
+    def query(
+        self,
+        query: str,
+        top_k: int = 10,
+        filter: Optional[Dict[str, Any]] = None,
+        query_embedding: Optional[List[float]] = None
+    ) -> List[Dict[str, Any]]:
+        """
+        Query the index.
+        
+        Args:
+            query: Search query
+            top_k: Number of results
+            filter: Optional metadata filter
+            query_embedding: Optional pre-computed query embedding
+            
+        Returns:
+            List of results with text, score, metadata
+        """
+        ...
+    
+    def delete(
+        self,
+        ids: Optional[List[str]] = None,
+        filter: Optional[Dict[str, Any]] = None,
+        delete_all: bool = False
+    ) -> int:
+        """Delete documents from the index."""
+        ...
+    
+    def stats(self) -> IndexStats:
+        """Get index statistics."""
+        ...
+
+
+class IndexRegistry:
+    """Registry for index types."""
+    
+    _instance: Optional["IndexRegistry"] = None
+    
+    def __new__(cls) -> "IndexRegistry":
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+            cls._instance._indices: Dict[str, Callable[..., IndexProtocol]] = {}
+        return cls._instance
+    
+    def register(self, name: str, factory: Callable[..., IndexProtocol]) -> None:
+        """Register an index factory."""
+        self._indices[name] = factory
+    
+    def get(self, name: str, **kwargs) -> Optional[IndexProtocol]:
+        """Get an index by name."""
+        if name in self._indices:
+            try:
+                return self._indices[name](**kwargs)
+            except Exception as e:
+                logger.warning(f"Failed to initialize index '{name}': {e}")
+                return None
+        return None
+    
+    def list_indices(self) -> List[str]:
+        """List all registered index names."""
+        return list(self._indices.keys())
+    
+    def clear(self) -> None:
+        """Clear all registered indices."""
+        self._indices.clear()
+
+
+def get_index_registry() -> IndexRegistry:
+    """Get the global index registry instance."""
+    return IndexRegistry()
+
+
+class KeywordIndex:
+    """
+    Simple BM25-like keyword index.
+    
+    Pure Python implementation with no external dependencies.
+    """
+    
+    name: str = "keyword"
+    index_type: IndexType = IndexType.KEYWORD
+    
+    def __init__(self, k1: float = 1.5, b: float = 0.75, **kwargs):
+        self.k1 = k1
+        self.b = b
+        self._documents: Dict[str, Dict[str, Any]] = {}  # id -> {text, metadata, terms}
+        self._term_doc_freq: Dict[str, int] = {}  # term -> doc count
+        self._avg_doc_len: float = 0.0
+        self._total_docs: int = 0
+    
+    def add_documents(
+        self,
+        texts: List[str],
+        metadatas: Optional[List[Dict[str, Any]]] = None,
+        ids: Optional[List[str]] = None,
+        embeddings: Optional[List[List[float]]] = None
+    ) -> List[str]:
+        """Add documents to the keyword index."""
+        import uuid
+        
+        metadatas = metadatas or [{} for _ in texts]
+        ids = ids or [str(uuid.uuid4()) for _ in texts]
+        
+        total_len = sum(len(doc["text"].split()) for doc in self._documents.values())
+        
+        for text, metadata, doc_id in zip(texts, metadatas, ids):
+            terms = self._tokenize(text)
+            term_freq = {}
+            for term in terms:
+                term_freq[term] = term_freq.get(term, 0) + 1
+            
+            self._documents[doc_id] = {
+                "text": text,
+                "metadata": metadata,
+                "terms": term_freq,
+                "length": len(terms)
+            }
+            
+            # Update document frequency
+            for term in set(terms):
+                self._term_doc_freq[term] = self._term_doc_freq.get(term, 0) + 1
+            
+            total_len += len(terms)
+        
+        self._total_docs = len(self._documents)
+        self._avg_doc_len = total_len / self._total_docs if self._total_docs > 0 else 0
+        
+        return ids
+    
+    def query(
+        self,
+        query: str,
+        top_k: int = 10,
+        filter: Optional[Dict[str, Any]] = None,
+        query_embedding: Optional[List[float]] = None
+    ) -> List[Dict[str, Any]]:
+        """Query using BM25 scoring."""
+        query_terms = self._tokenize(query)
+        
+        scores = []
+        for doc_id, doc in self._documents.items():
+            # Apply filter
+            if filter:
+                match = True
+                for key, value in filter.items():
+                    if doc["metadata"].get(key) != value:
+                        match = False
+                        break
+                if not match:
+                    continue
+            
+            score = self._bm25_score(query_terms, doc)
+            scores.append({
+                "id": doc_id,
+                "text": doc["text"],
+                "score": score,
+                "metadata": doc["metadata"]
+            })
+        
+        # Sort by score
+        scores.sort(key=lambda x: x["score"], reverse=True)
+        return scores[:top_k]
+    
+    def delete(
+        self,
+        ids: Optional[List[str]] = None,
+        filter: Optional[Dict[str, Any]] = None,
+        delete_all: bool = False
+    ) -> int:
+        """Delete documents from the index."""
+        if delete_all:
+            count = len(self._documents)
+            self._documents.clear()
+            self._term_doc_freq.clear()
+            self._total_docs = 0
+            self._avg_doc_len = 0
+            return count
+        
+        deleted = 0
+        to_delete = []
+        
+        if ids:
+            to_delete = [id_ for id_ in ids if id_ in self._documents]
+        elif filter:
+            for doc_id, doc in self._documents.items():
+                match = True
+                for key, value in filter.items():
+                    if doc["metadata"].get(key) != value:
+                        match = False
+                        break
+                if match:
+                    to_delete.append(doc_id)
+        
+        for doc_id in to_delete:
+            doc = self._documents[doc_id]
+            # Update term frequencies
+            for term in doc["terms"]:
+                if term in self._term_doc_freq:
+                    self._term_doc_freq[term] -= 1
+                    if self._term_doc_freq[term] <= 0:
+                        del self._term_doc_freq[term]
+            del self._documents[doc_id]
+            deleted += 1
+        
+        self._total_docs = len(self._documents)
+        if self._total_docs > 0:
+            self._avg_doc_len = sum(d["length"] for d in self._documents.values()) / self._total_docs
+        else:
+            self._avg_doc_len = 0
+        
+        return deleted
+    
+    def stats(self) -> IndexStats:
+        """Get index statistics."""
+        return IndexStats(
+            index_type=self.index_type,
+            document_count=self._total_docs,
+            chunk_count=self._total_docs,
+            total_tokens=sum(d["length"] for d in self._documents.values()),
+            metadata={"unique_terms": len(self._term_doc_freq)}
+        )
+    
+    def _tokenize(self, text: str) -> List[str]:
+        """Simple tokenization."""
+        import re
+        # Lowercase and split on non-alphanumeric
+        tokens = re.findall(r'\b\w+\b', text.lower())
+        # Remove very short tokens
+        return [t for t in tokens if len(t) > 1]
+    
+    def _bm25_score(self, query_terms: List[str], doc: Dict[str, Any]) -> float:
+        """Calculate BM25 score."""
+        score = 0.0
+        doc_len = doc["length"]
+        term_freq = doc["terms"]
+        
+        for term in query_terms:
+            if term not in term_freq:
+                continue
+            
+            tf = term_freq[term]
+            df = self._term_doc_freq.get(term, 0)
+            
+            # IDF component
+            idf = math.log((self._total_docs - df + 0.5) / (df + 0.5) + 1)
+            
+            # TF component with length normalization
+            tf_norm = (tf * (self.k1 + 1)) / (tf + self.k1 * (1 - self.b + self.b * doc_len / self._avg_doc_len))
+            
+            score += idf * tf_norm
+        
+        return score
+
+
+# Register the keyword index by default
+def _register_default_indices():
+    """Register default indices."""
+    registry = get_index_registry()
+    registry.register("keyword", KeywordIndex)
+
+
+_register_default_indices()

src/praisonai-agents/praisonaiagents/knowledge/knowledge.py

@@ -91,19 +91,20 @@ class MongoDBMemory:
             raise Exception(f"Failed to initialize MongoDB: {e}")
     
     def _init_embedding_model(self):
-        """Initialize embedding model from config."""
+        """Initialize embedding model from config using litellm for unified provider support."""
         try:
             # Set up embedding model based on config
             embedder_config = self.config.get("embedder", {})
-            if embedder_config.get("provider") == "openai":
-                import openai
-                self.embedding_model = openai.OpenAI()
-                self.embedding_model_name = embedder_config.get("config", {}).get("model", "text-embedding-3-small")
-            else:
-                # Default to OpenAI
-                import openai
-                self.embedding_model = openai.OpenAI()
-                self.embedding_model_name = "text-embedding-3-small"
+            provider = embedder_config.get("provider", "openai")
+            model_name = embedder_config.get("config", {}).get("model", "text-embedding-3-small")
+            
+            # Store model name for later use
+            self.embedding_model_name = model_name
+            
+            # Use litellm for embeddings - it handles all providers uniformly
+            # We'll use litellm.embedding() in _get_embeddings() instead of storing a client
+            self.embedding_provider = provider
+            self.embedding_model = None  # Will use litellm.embedding() directly
         except Exception as e:
             raise Exception(f"Failed to initialize embedding model: {e}")
     
@@ -172,13 +173,16 @@ class MongoDBMemory:
             logging.warning(f"Could not create vector search index: {e}")
     
     def _get_embedding(self, text):
-        """Get embedding for text."""
+        """Get embedding for text using litellm for unified provider support."""
         try:
-            response = self.embedding_model.embeddings.create(
-                input=text,
-                model=self.embedding_model_name
+            import litellm
+            
+            # Use litellm.embedding() for unified multi-provider support
+            response = litellm.embedding(
+                model=self.embedding_model_name,
+                input=text
             )
-            return response.data[0].embedding
+            return response.data[0]["embedding"]
         except Exception as e:
             logging.error(f"Error getting embedding: {e}")
             return None
@@ -500,6 +504,33 @@ class Knowledge:
                 base_config["graph_store"] = self._config["graph_store"]
         return base_config
 
+    def _prepare_mem0_config(self, config):
+        """Prepare config for mem0 by removing PraisonAI-specific fields.
+        
+        mem0's RerankerConfig only accepts 'provider' and 'config' fields.
+        PraisonAI adds 'enabled' and 'default_rerank' for internal use.
+        """
+        mem0_config = config.copy()
+        
+        # Strip PraisonAI-specific reranker fields that mem0 doesn't accept
+        if "reranker" in mem0_config:
+            reranker = mem0_config["reranker"]
+            if isinstance(reranker, dict):
+                # Keep only mem0-compatible fields: provider, config
+                mem0_reranker = {}
+                if "provider" in reranker:
+                    mem0_reranker["provider"] = reranker["provider"]
+                if "config" in reranker:
+                    mem0_reranker["config"] = reranker["config"]
+                
+                # If no valid mem0 fields, remove reranker entirely
+                if mem0_reranker:
+                    mem0_config["reranker"] = mem0_reranker
+                else:
+                    del mem0_config["reranker"]
+        
+        return mem0_config
+
     @cached_property
     def memory(self):
         # Check if MongoDB provider is specified
@@ -511,19 +542,22 @@ class Knowledge:
                 # Fall back to default memory
                 pass
         
+        # Prepare config for mem0 (strip PraisonAI-specific fields)
+        mem0_config = self._prepare_mem0_config(self.config)
+        
         # Default Mem0 memory
         try:
-            return CustomMemory.from_config(self.config)
+            return CustomMemory.from_config(mem0_config)
         except (NotImplementedError, ValueError) as e:
             if "list_collections" in str(e) or "Extra fields not allowed" in str(e):
                 # Keep only allowed fields
                 vector_store_config = {
-                    "collection_name": self.config["vector_store"]["config"]["collection_name"],
-                    "path": self.config["vector_store"]["config"]["path"]
+                    "collection_name": mem0_config["vector_store"]["config"]["collection_name"],
+                    "path": mem0_config["vector_store"]["config"]["path"]
                 }
-                self.config["vector_store"]["config"] = vector_store_config
+                mem0_config["vector_store"]["config"] = vector_store_config
                 from mem0 import Memory
-                return Memory.from_config(self.config)
+                return Memory.from_config(mem0_config)
             raise
 
     @cached_property

src/praisonai-agents/praisonaiagents/knowledge/query_engine.py

@@ -0,0 +1,351 @@
+"""
+Query Engine Patterns for PraisonAI Agents.
+
+This module defines query modes and protocols.
+NO heavy imports - only stdlib and typing.
+
+Implementations are provided by the wrapper layer.
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Callable, Dict, List, Optional, Protocol, runtime_checkable
+from enum import Enum
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class QueryMode(str, Enum):
+    """Available query modes."""
+    DEFAULT = "default"
+    SUB_QUESTION = "sub_question"
+    SQL = "sql"
+    ROUTER = "router"
+    SUMMARIZE = "summarize"
+
+
+@dataclass
+class QueryResult:
+    """
+    Result from a query operation.
+    
+    Attributes:
+        answer: The synthesized answer
+        sources: List of source documents/chunks used
+        sub_questions: Optional list of sub-questions (for sub_question mode)
+        metadata: Additional metadata
+    """
+    answer: str
+    sources: List[Dict[str, Any]] = field(default_factory=list)
+    sub_questions: Optional[List[str]] = None
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "answer": self.answer,
+            "sources": self.sources,
+            "sub_questions": self.sub_questions,
+            "metadata": self.metadata
+        }
+
+
+@runtime_checkable
+class QueryEngineProtocol(Protocol):
+    """
+    Protocol for query engines.
+    
+    Implementations must provide a query method that takes a question
+    and returns an answer with sources.
+    """
+    
+    name: str
+    mode: QueryMode
+    
+    def query(
+        self,
+        question: str,
+        context: Optional[List[str]] = None,
+        **kwargs
+    ) -> QueryResult:
+        """
+        Query the knowledge base.
+        
+        Args:
+            question: The user's question
+            context: Optional pre-retrieved context
+            **kwargs: Additional engine-specific options
+            
+        Returns:
+            QueryResult with answer and sources
+        """
+        ...
+    
+    async def aquery(
+        self,
+        question: str,
+        context: Optional[List[str]] = None,
+        **kwargs
+    ) -> QueryResult:
+        """Async version of query."""
+        ...
+
+
+class QueryEngineRegistry:
+    """Registry for query engines."""
+    
+    _instance: Optional["QueryEngineRegistry"] = None
+    
+    def __new__(cls) -> "QueryEngineRegistry":
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+            cls._instance._engines: Dict[str, Callable[..., QueryEngineProtocol]] = {}
+        return cls._instance
+    
+    def register(self, name: str, factory: Callable[..., QueryEngineProtocol]) -> None:
+        """Register a query engine factory."""
+        self._engines[name] = factory
+    
+    def get(self, name: str, **kwargs) -> Optional[QueryEngineProtocol]:
+        """Get a query engine by name."""
+        if name in self._engines:
+            try:
+                return self._engines[name](**kwargs)
+            except Exception as e:
+                logger.warning(f"Failed to initialize query engine '{name}': {e}")
+                return None
+        return None
+    
+    def list_engines(self) -> List[str]:
+        """List all registered engine names."""
+        return list(self._engines.keys())
+    
+    def clear(self) -> None:
+        """Clear all registered engines."""
+        self._engines.clear()
+
+
+def get_query_engine_registry() -> QueryEngineRegistry:
+    """Get the global query engine registry instance."""
+    return QueryEngineRegistry()
+
+
+def decompose_question(question: str) -> List[str]:
+    """
+    Simple deterministic question decomposition.
+    
+    Splits complex questions into sub-questions without LLM.
+    Used as fallback when LLM is not available.
+    
+    Args:
+        question: The complex question
+        
+    Returns:
+        List of sub-questions
+    """
+    sub_questions = []
+    
+    # Split on common conjunctions
+    conjunctions = [" and ", " also ", " additionally ", " furthermore ", " moreover "]
+    parts = [question]
+    
+    for conj in conjunctions:
+        new_parts = []
+        for part in parts:
+            new_parts.extend(part.split(conj))
+        parts = new_parts
+    
+    # Clean up parts
+    for part in parts:
+        part = part.strip()
+        if part:
+            # Ensure it ends with a question mark
+            if not part.endswith("?"):
+                part = part + "?"
+            # Capitalize first letter
+            part = part[0].upper() + part[1:] if len(part) > 1 else part.upper()
+            sub_questions.append(part)
+    
+    # If no decomposition happened, return original
+    if len(sub_questions) <= 1:
+        return [question]
+    
+    return sub_questions
+
+
+def synthesize_answer(
+    question: str,
+    contexts: List[str],
+    max_context_length: int = 4000
+) -> str:
+    """
+    Simple answer synthesis without LLM.
+    
+    Concatenates relevant context with the question.
+    Used as fallback when LLM is not available.
+    
+    Args:
+        question: The original question
+        contexts: List of context strings
+        max_context_length: Maximum total context length
+        
+    Returns:
+        Synthesized answer string
+    """
+    if not contexts:
+        return "No relevant information found."
+    
+    # Truncate contexts to fit
+    total_length = 0
+    selected_contexts = []
+    
+    for ctx in contexts:
+        if total_length + len(ctx) > max_context_length:
+            remaining = max_context_length - total_length
+            if remaining > 100:
+                selected_contexts.append(ctx[:remaining] + "...")
+            break
+        selected_contexts.append(ctx)
+        total_length += len(ctx)
+    
+    # Format as a simple answer
+    context_text = "\n\n".join(selected_contexts)
+    
+    return f"Based on the available information:\n\n{context_text}"
+
+
+class SimpleQueryEngine:
+    """
+    Simple query engine for testing and fallback.
+    
+    Uses basic context concatenation without LLM synthesis.
+    """
+    
+    name: str = "simple"
+    mode: QueryMode = QueryMode.DEFAULT
+    
+    def __init__(self, **kwargs):
+        pass
+    
+    def query(
+        self,
+        question: str,
+        context: Optional[List[str]] = None,
+        **kwargs
+    ) -> QueryResult:
+        """Simple query that returns context as answer."""
+        contexts = context or []
+        answer = synthesize_answer(question, contexts)
+        
+        return QueryResult(
+            answer=answer,
+            sources=[{"text": ctx} for ctx in contexts],
+            metadata={"mode": self.mode.value}
+        )
+    
+    async def aquery(
+        self,
+        question: str,
+        context: Optional[List[str]] = None,
+        **kwargs
+    ) -> QueryResult:
+        """Async version (just calls sync)."""
+        return self.query(question, context, **kwargs)
+
+
+class SubQuestionEngine:
+    """
+    Sub-question decomposition query engine.
+    
+    Decomposes complex questions into simpler sub-questions.
+    Uses deterministic fallback when LLM is not available.
+    """
+    
+    name: str = "sub_question"
+    mode: QueryMode = QueryMode.SUB_QUESTION
+    
+    def __init__(self, llm: Optional[Any] = None, **kwargs):
+        self.llm = llm
+    
+    def query(
+        self,
+        question: str,
+        context: Optional[List[str]] = None,
+        retriever: Optional[Any] = None,
+        **kwargs
+    ) -> QueryResult:
+        """Query with sub-question decomposition."""
+        # Decompose question
+        if self.llm:
+            sub_questions = self._decompose_with_llm(question)
+        else:
+            sub_questions = decompose_question(question)
+        
+        # Gather context for each sub-question
+        all_contexts = list(context or [])
+        sub_answers = []
+        
+        for sub_q in sub_questions:
+            if retriever:
+                # Retrieve for each sub-question
+                try:
+                    results = retriever.retrieve(sub_q, top_k=3)
+                    for r in results:
+                        if hasattr(r, 'text'):
+                            all_contexts.append(r.text)
+                        elif isinstance(r, dict):
+                            all_contexts.append(r.get('text', str(r)))
+                except Exception as e:
+                    logger.warning(f"Retrieval failed for sub-question: {e}")
+            
+            # Simple sub-answer (would use LLM in full implementation)
+            sub_answers.append(f"Sub-question: {sub_q}")
+        
+        # Synthesize final answer
+        answer = synthesize_answer(question, all_contexts)
+        
+        return QueryResult(
+            answer=answer,
+            sources=[{"text": ctx} for ctx in all_contexts[:10]],
+            sub_questions=sub_questions,
+            metadata={"mode": self.mode.value, "sub_question_count": len(sub_questions)}
+        )
+    
+    async def aquery(
+        self,
+        question: str,
+        context: Optional[List[str]] = None,
+        **kwargs
+    ) -> QueryResult:
+        """Async version (just calls sync)."""
+        return self.query(question, context, **kwargs)
+    
+    def _decompose_with_llm(self, question: str) -> List[str]:
+        """Decompose question using LLM."""
+        try:
+            prompt = f"""Break down this complex question into simpler sub-questions that can be answered independently.
+            
+Question: {question}
+
+Return only the sub-questions, one per line, without numbering or bullets."""
+            
+            response = self.llm.chat(prompt)
+            if response:
+                lines = response.strip().split("\n")
+                return [line.strip() for line in lines if line.strip()]
+        except Exception as e:
+            logger.warning(f"LLM decomposition failed: {e}")
+        
+        return decompose_question(question)
+
+
+# Register default engines
+def _register_default_engines():
+    """Register default query engines."""
+    registry = get_query_engine_registry()
+    registry.register("default", SimpleQueryEngine)
+    registry.register("simple", SimpleQueryEngine)
+    registry.register("sub_question", SubQuestionEngine)
+
+
+_register_default_engines()

src/praisonai-agents/praisonaiagents/knowledge/readers.py

@@ -0,0 +1,286 @@
+"""
+Data Reader Adapter Protocol for PraisonAI Agents.
+
+This module defines the protocol for document readers and a registry system.
+NO heavy imports - only stdlib and typing.
+
+Implementations are provided by the wrapper layer (praisonai.adapters.readers).
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Callable, Dict, List, Optional, Protocol, runtime_checkable
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class Document:
+    """
+    Lightweight document representation.
+    
+    Attributes:
+        content: The text content of the document
+        metadata: Optional metadata dict (source, filename, page, etc.)
+        doc_id: Optional unique identifier
+        embedding: Optional pre-computed embedding
+    """
+    content: str
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    doc_id: Optional[str] = None
+    embedding: Optional[List[float]] = None
+    
+    def __post_init__(self):
+        if self.doc_id is None:
+            import uuid
+            self.doc_id = str(uuid.uuid4())
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "content": self.content,
+            "metadata": self.metadata,
+            "doc_id": self.doc_id,
+            "embedding": self.embedding
+        }
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "Document":
+        """Create from dictionary."""
+        return cls(
+            content=data.get("content", ""),
+            metadata=data.get("metadata", {}),
+            doc_id=data.get("doc_id"),
+            embedding=data.get("embedding")
+        )
+
+
+@runtime_checkable
+class ReaderProtocol(Protocol):
+    """
+    Protocol for document readers.
+    
+    Implementations must provide:
+    - name: Identifier for the reader
+    - supported_extensions: List of file extensions this reader handles
+    - load(): Method to load documents from a source
+    """
+    
+    name: str
+    supported_extensions: List[str]
+    
+    def load(
+        self, 
+        source: str, 
+        *, 
+        metadata: Optional[Dict[str, Any]] = None
+    ) -> List[Document]:
+        """
+        Load documents from a source.
+        
+        Args:
+            source: File path, URL, or other source identifier
+            metadata: Optional metadata to attach to all documents
+            
+        Returns:
+            List of Document objects
+        """
+        ...
+    
+    def can_handle(self, source: str) -> bool:
+        """
+        Check if this reader can handle the given source.
+        
+        Args:
+            source: File path, URL, or other source identifier
+            
+        Returns:
+            True if this reader can handle the source
+        """
+        ...
+
+
+class ReaderRegistry:
+    """
+    Registry for document readers.
+    
+    Provides lazy loading and automatic reader selection based on source type.
+    """
+    
+    _instance: Optional["ReaderRegistry"] = None
+    
+    def __new__(cls) -> "ReaderRegistry":
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+            cls._instance._readers: Dict[str, Callable[[], ReaderProtocol]] = {}
+            cls._instance._extension_map: Dict[str, str] = {}
+            cls._instance._initialized_readers: Dict[str, ReaderProtocol] = {}
+        return cls._instance
+    
+    def register(
+        self, 
+        name: str, 
+        factory: Callable[[], ReaderProtocol],
+        extensions: Optional[List[str]] = None
+    ) -> None:
+        """
+        Register a reader factory.
+        
+        Args:
+            name: Unique name for the reader
+            factory: Callable that returns a ReaderProtocol instance
+            extensions: List of file extensions this reader handles
+        """
+        self._readers[name] = factory
+        if extensions:
+            for ext in extensions:
+                ext_lower = ext.lower().lstrip(".")
+                self._extension_map[ext_lower] = name
+    
+    def get(self, name: str) -> Optional[ReaderProtocol]:
+        """
+        Get a reader by name (lazy initialization).
+        
+        Args:
+            name: Reader name
+            
+        Returns:
+            ReaderProtocol instance or None
+        """
+        if name in self._initialized_readers:
+            return self._initialized_readers[name]
+        
+        if name in self._readers:
+            try:
+                reader = self._readers[name]()
+                self._initialized_readers[name] = reader
+                return reader
+            except Exception as e:
+                logger.warning(f"Failed to initialize reader '{name}': {e}")
+                return None
+        return None
+    
+    def get_for_source(self, source: str) -> Optional[ReaderProtocol]:
+        """
+        Get the best reader for a given source.
+        
+        Args:
+            source: File path, URL, or other source identifier
+            
+        Returns:
+            Best matching ReaderProtocol or None
+        """
+        source_kind = detect_source_kind(source)
+        
+        if source_kind == "file":
+            ext = os.path.splitext(source)[1].lower().lstrip(".")
+            if ext in self._extension_map:
+                return self.get(self._extension_map[ext])
+        elif source_kind == "url":
+            # Try URL reader
+            if "url" in self._readers:
+                return self.get("url")
+        elif source_kind == "directory":
+            if "directory" in self._readers:
+                return self.get("directory")
+        
+        # Fallback to auto reader if available
+        if "auto" in self._readers:
+            return self.get("auto")
+        
+        return None
+    
+    def list_readers(self) -> List[str]:
+        """List all registered reader names."""
+        return list(self._readers.keys())
+    
+    def list_extensions(self) -> Dict[str, str]:
+        """List all registered extensions and their readers."""
+        return dict(self._extension_map)
+    
+    def clear(self) -> None:
+        """Clear all registered readers (mainly for testing)."""
+        self._readers.clear()
+        self._extension_map.clear()
+        self._initialized_readers.clear()
+
+
+def get_reader_registry() -> ReaderRegistry:
+    """Get the global reader registry instance."""
+    return ReaderRegistry()
+
+
+def detect_source_kind(source: str) -> str:
+    """
+    Detect the kind of source without importing heavy libraries.
+    
+    Args:
+        source: File path, URL, directory, or glob pattern
+        
+    Returns:
+        One of: "file", "url", "directory", "glob", "unknown"
+    """
+    if not isinstance(source, str):
+        return "unknown"
+    
+    source = source.strip()
+    
+    # Check for URL
+    if source.startswith(("http://", "https://", "ftp://", "s3://", "gs://")):
+        return "url"
+    
+    # Check for glob pattern
+    if "*" in source or "?" in source or "[" in source:
+        return "glob"
+    
+    # Check if it's a directory
+    if os.path.isdir(source):
+        return "directory"
+    
+    # Check if it's a file
+    if os.path.isfile(source):
+        return "file"
+    
+    # Check if it looks like a file path (has extension)
+    if os.path.splitext(source)[1]:
+        return "file"
+    
+    return "unknown"
+
+
+def get_file_extension(source: str) -> str:
+    """
+    Get the file extension from a source path.
+    
+    Args:
+        source: File path or URL
+        
+    Returns:
+        Lowercase extension without dot, or empty string
+    """
+    # Handle URLs by extracting path
+    if source.startswith(("http://", "https://")):
+        from urllib.parse import urlparse
+        path = urlparse(source).path
+        return os.path.splitext(path)[1].lower().lstrip(".")
+    
+    return os.path.splitext(source)[1].lower().lstrip(".")
+
+
+# Default extension mappings (reader name -> extensions)
+DEFAULT_EXTENSION_MAPPINGS = {
+    "text": ["txt", "text"],
+    "markdown": ["md", "markdown"],
+    "json": ["json", "jsonl"],
+    "csv": ["csv", "tsv"],
+    "html": ["html", "htm"],
+    "xml": ["xml"],
+    "pdf": ["pdf"],
+    "docx": ["docx", "doc"],
+    "xlsx": ["xlsx", "xls"],
+    "pptx": ["pptx", "ppt"],
+    "image": ["jpg", "jpeg", "png", "gif", "bmp", "tiff", "webp"],
+    "audio": ["mp3", "wav", "ogg", "m4a", "flac"],
+    "video": ["mp4", "avi", "mov", "mkv", "webm"],
+}

src/praisonai-agents/praisonaiagents/knowledge/rerankers.py

@@ -0,0 +1,196 @@
+"""
+Reranker Protocol for PraisonAI Agents.
+
+This module defines the protocol for rerankers and a registry system.
+NO heavy imports - only stdlib and typing.
+
+Implementations are provided by praisonai-tools or wrapper layer.
+"""
+
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Protocol, runtime_checkable
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class RerankResult:
+    """
+    Result from a reranking operation.
+    
+    Attributes:
+        text: The text content
+        score: Reranking score (higher is more relevant)
+        original_index: Original position before reranking
+        metadata: Optional metadata
+    """
+    text: str
+    score: float
+    original_index: int
+    metadata: Dict[str, Any] = None
+    
+    def __post_init__(self):
+        if self.metadata is None:
+            self.metadata = {}
+
+
+@runtime_checkable
+class RerankerProtocol(Protocol):
+    """
+    Protocol for rerankers.
+    
+    Implementations must provide a rerank method that takes a query
+    and a list of documents and returns reranked results.
+    """
+    
+    name: str
+    
+    def rerank(
+        self,
+        query: str,
+        documents: List[str],
+        top_k: Optional[int] = None,
+        **kwargs
+    ) -> List[RerankResult]:
+        """
+        Rerank documents based on relevance to query.
+        
+        Args:
+            query: The search query
+            documents: List of document texts to rerank
+            top_k: Optional limit on results (None = return all)
+            **kwargs: Additional reranker-specific options
+            
+        Returns:
+            List of RerankResult objects sorted by relevance
+        """
+        ...
+    
+    async def arerank(
+        self,
+        query: str,
+        documents: List[str],
+        top_k: Optional[int] = None,
+        **kwargs
+    ) -> List[RerankResult]:
+        """Async version of rerank."""
+        ...
+
+
+class RerankerRegistry:
+    """
+    Registry for rerankers.
+    """
+    
+    _instance: Optional["RerankerRegistry"] = None
+    
+    def __new__(cls) -> "RerankerRegistry":
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+            cls._instance._rerankers: Dict[str, Callable[..., RerankerProtocol]] = {}
+        return cls._instance
+    
+    def register(
+        self, 
+        name: str, 
+        factory: Callable[..., RerankerProtocol]
+    ) -> None:
+        """Register a reranker factory."""
+        self._rerankers[name] = factory
+    
+    def get(
+        self, 
+        name: str, 
+        **kwargs
+    ) -> Optional[RerankerProtocol]:
+        """Get a reranker by name."""
+        if name in self._rerankers:
+            try:
+                return self._rerankers[name](**kwargs)
+            except Exception as e:
+                logger.warning(f"Failed to initialize reranker '{name}': {e}")
+                return None
+        return None
+    
+    def list_rerankers(self) -> List[str]:
+        """List all registered reranker names."""
+        return list(self._rerankers.keys())
+    
+    def clear(self) -> None:
+        """Clear all registered rerankers."""
+        self._rerankers.clear()
+
+
+def get_reranker_registry() -> RerankerRegistry:
+    """Get the global reranker registry instance."""
+    return RerankerRegistry()
+
+
+class SimpleReranker:
+    """
+    Simple keyword-based reranker for testing and fallback.
+    
+    Uses term frequency overlap for scoring. No external dependencies.
+    """
+    
+    name: str = "simple"
+    
+    def __init__(self, **kwargs):
+        pass
+    
+    def rerank(
+        self,
+        query: str,
+        documents: List[str],
+        top_k: Optional[int] = None,
+        **kwargs
+    ) -> List[RerankResult]:
+        """Rerank using simple term overlap."""
+        query_terms = set(query.lower().split())
+        
+        results = []
+        for i, doc in enumerate(documents):
+            doc_terms = set(doc.lower().split())
+            
+            # Calculate Jaccard similarity
+            if not query_terms or not doc_terms:
+                score = 0.0
+            else:
+                intersection = len(query_terms & doc_terms)
+                union = len(query_terms | doc_terms)
+                score = intersection / union if union > 0 else 0.0
+            
+            results.append(RerankResult(
+                text=doc,
+                score=score,
+                original_index=i
+            ))
+        
+        # Sort by score descending
+        results.sort(key=lambda x: x.score, reverse=True)
+        
+        if top_k is not None:
+            results = results[:top_k]
+        
+        return results
+    
+    async def arerank(
+        self,
+        query: str,
+        documents: List[str],
+        top_k: Optional[int] = None,
+        **kwargs
+    ) -> List[RerankResult]:
+        """Async version (just calls sync)."""
+        return self.rerank(query, documents, top_k, **kwargs)
+
+
+# Register the simple reranker by default
+def _register_default_rerankers():
+    """Register default rerankers."""
+    registry = get_reranker_registry()
+    registry.register("simple", SimpleReranker)
+
+
+_register_default_rerankers()

src/praisonai-agents/praisonaiagents/knowledge/retrieval.py

@@ -0,0 +1,285 @@
+"""
+Retrieval Strategy Patterns for PraisonAI Agents.
+
+This module defines protocols and strategies for document retrieval.
+NO heavy imports - only stdlib and typing.
+
+Implementations are provided by the wrapper layer.
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Callable, Dict, List, Optional, Protocol, runtime_checkable
+from enum import Enum
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class RetrievalStrategy(str, Enum):
+    """Available retrieval strategies."""
+    BASIC = "basic"
+    FUSION = "fusion"
+    RECURSIVE = "recursive"
+    AUTO_MERGE = "auto_merge"
+    HYBRID = "hybrid"
+
+
+@dataclass
+class RetrievalResult:
+    """
+    Result from a retrieval operation.
+    
+    Attributes:
+        text: The retrieved text content
+        score: Relevance score (0-1)
+        metadata: Source metadata
+        doc_id: Document identifier
+        chunk_index: Index of chunk within document
+    """
+    text: str
+    score: float
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    doc_id: Optional[str] = None
+    chunk_index: Optional[int] = None
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "text": self.text,
+            "score": self.score,
+            "metadata": self.metadata,
+            "doc_id": self.doc_id,
+            "chunk_index": self.chunk_index
+        }
+
+
+@runtime_checkable
+class RetrieverProtocol(Protocol):
+    """
+    Protocol for document retrievers.
+    
+    Implementations must provide a retrieve method that takes a query
+    and returns relevant documents.
+    """
+    
+    name: str
+    strategy: RetrievalStrategy
+    
+    def retrieve(
+        self,
+        query: str,
+        top_k: int = 10,
+        filter: Optional[Dict[str, Any]] = None,
+        **kwargs
+    ) -> List[RetrievalResult]:
+        """
+        Retrieve relevant documents for a query.
+        
+        Args:
+            query: The search query
+            top_k: Maximum number of results
+            filter: Optional metadata filter
+            **kwargs: Additional retriever-specific options
+            
+        Returns:
+            List of RetrievalResult objects sorted by relevance
+        """
+        ...
+    
+    async def aretrieve(
+        self,
+        query: str,
+        top_k: int = 10,
+        filter: Optional[Dict[str, Any]] = None,
+        **kwargs
+    ) -> List[RetrievalResult]:
+        """Async version of retrieve."""
+        ...
+
+
+class RetrieverRegistry:
+    """
+    Registry for retrieval strategies.
+    """
+    
+    _instance: Optional["RetrieverRegistry"] = None
+    
+    def __new__(cls) -> "RetrieverRegistry":
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+            cls._instance._retrievers: Dict[str, Callable[..., RetrieverProtocol]] = {}
+        return cls._instance
+    
+    def register(
+        self, 
+        name: str, 
+        factory: Callable[..., RetrieverProtocol]
+    ) -> None:
+        """Register a retriever factory."""
+        self._retrievers[name] = factory
+    
+    def get(
+        self, 
+        name: str, 
+        **kwargs
+    ) -> Optional[RetrieverProtocol]:
+        """Get a retriever by name."""
+        if name in self._retrievers:
+            try:
+                return self._retrievers[name](**kwargs)
+            except Exception as e:
+                logger.warning(f"Failed to initialize retriever '{name}': {e}")
+                return None
+        return None
+    
+    def list_retrievers(self) -> List[str]:
+        """List all registered retriever names."""
+        return list(self._retrievers.keys())
+    
+    def clear(self) -> None:
+        """Clear all registered retrievers."""
+        self._retrievers.clear()
+
+
+def get_retriever_registry() -> RetrieverRegistry:
+    """Get the global retriever registry instance."""
+    return RetrieverRegistry()
+
+
+def reciprocal_rank_fusion(
+    result_lists: List[List[RetrievalResult]],
+    k: int = 60
+) -> List[RetrievalResult]:
+    """
+    Reciprocal Rank Fusion (RRF) algorithm for combining multiple result lists.
+    
+    Pure Python implementation with no external dependencies.
+    
+    Args:
+        result_lists: List of result lists from different retrievers
+        k: RRF constant (default 60)
+        
+    Returns:
+        Fused and re-ranked list of results
+    """
+    # Track scores by document ID
+    fused_scores: Dict[str, float] = {}
+    doc_map: Dict[str, RetrievalResult] = {}
+    
+    for results in result_lists:
+        for rank, result in enumerate(results):
+            doc_key = result.doc_id or result.text[:100]  # Use text prefix as fallback key
+            
+            # RRF score: 1 / (k + rank)
+            rrf_score = 1.0 / (k + rank + 1)
+            
+            if doc_key in fused_scores:
+                fused_scores[doc_key] += rrf_score
+            else:
+                fused_scores[doc_key] = rrf_score
+                doc_map[doc_key] = result
+    
+    # Sort by fused score
+    sorted_keys = sorted(fused_scores.keys(), key=lambda x: fused_scores[x], reverse=True)
+    
+    # Build result list with updated scores
+    fused_results = []
+    for key in sorted_keys:
+        result = doc_map[key]
+        fused_results.append(RetrievalResult(
+            text=result.text,
+            score=fused_scores[key],
+            metadata=result.metadata,
+            doc_id=result.doc_id,
+            chunk_index=result.chunk_index
+        ))
+    
+    return fused_results
+
+
+def merge_adjacent_chunks(
+    results: List[RetrievalResult],
+    max_gap: int = 1
+) -> List[RetrievalResult]:
+    """
+    Merge adjacent chunks from the same document.
+    
+    Args:
+        results: List of retrieval results
+        max_gap: Maximum gap between chunk indices to merge
+        
+    Returns:
+        Merged results
+    """
+    if not results:
+        return results
+    
+    # Group by document ID
+    doc_groups: Dict[str, List[RetrievalResult]] = {}
+    for result in results:
+        doc_id = result.doc_id or "unknown"
+        if doc_id not in doc_groups:
+            doc_groups[doc_id] = []
+        doc_groups[doc_id].append(result)
+    
+    merged_results = []
+    
+    for doc_id, chunks in doc_groups.items():
+        # Sort by chunk index
+        chunks_with_index = [c for c in chunks if c.chunk_index is not None]
+        chunks_without_index = [c for c in chunks if c.chunk_index is None]
+        
+        if not chunks_with_index:
+            merged_results.extend(chunks_without_index)
+            continue
+        
+        chunks_with_index.sort(key=lambda x: x.chunk_index or 0)
+        
+        # Merge adjacent chunks
+        current_group = [chunks_with_index[0]]
+        
+        for chunk in chunks_with_index[1:]:
+            last_chunk = current_group[-1]
+            last_idx = last_chunk.chunk_index or 0
+            curr_idx = chunk.chunk_index or 0
+            
+            if curr_idx - last_idx <= max_gap:
+                current_group.append(chunk)
+            else:
+                # Merge current group and start new one
+                merged_results.append(_merge_chunk_group(current_group))
+                current_group = [chunk]
+        
+        # Don't forget the last group
+        if current_group:
+            merged_results.append(_merge_chunk_group(current_group))
+        
+        merged_results.extend(chunks_without_index)
+    
+    # Sort by score
+    merged_results.sort(key=lambda x: x.score, reverse=True)
+    return merged_results
+
+
+def _merge_chunk_group(chunks: List[RetrievalResult]) -> RetrievalResult:
+    """Merge a group of chunks into a single result."""
+    if len(chunks) == 1:
+        return chunks[0]
+    
+    # Combine text
+    merged_text = "\n".join(c.text for c in chunks)
+    
+    # Average score
+    avg_score = sum(c.score for c in chunks) / len(chunks)
+    
+    # Use first chunk's metadata
+    metadata = chunks[0].metadata.copy()
+    metadata["merged_chunks"] = len(chunks)
+    
+    return RetrievalResult(
+        text=merged_text,
+        score=avg_score,
+        metadata=metadata,
+        doc_id=chunks[0].doc_id,
+        chunk_index=chunks[0].chunk_index
+    )

src/praisonai-agents/praisonaiagents/knowledge/vector_store.py

@@ -0,0 +1,422 @@
+"""
+Vector Store Adapter Protocol for PraisonAI Agents.
+
+This module defines the protocol for vector stores and a registry system.
+NO heavy imports - only stdlib and typing.
+
+Implementations are provided by praisonai-tools or wrapper layer.
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Callable, Dict, List, Optional, Protocol, runtime_checkable
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class VectorRecord:
+    """
+    A record in a vector store.
+    
+    Attributes:
+        id: Unique identifier
+        text: The text content
+        embedding: The vector embedding
+        metadata: Optional metadata dict
+        score: Optional similarity score (for query results)
+    """
+    id: str
+    text: str
+    embedding: List[float]
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    score: Optional[float] = None
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary."""
+        return {
+            "id": self.id,
+            "text": self.text,
+            "embedding": self.embedding,
+            "metadata": self.metadata,
+            "score": self.score
+        }
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> "VectorRecord":
+        """Create from dictionary."""
+        return cls(
+            id=data["id"],
+            text=data["text"],
+            embedding=data["embedding"],
+            metadata=data.get("metadata", {}),
+            score=data.get("score")
+        )
+
+
+@runtime_checkable
+class VectorStoreProtocol(Protocol):
+    """
+    Protocol for vector stores.
+    
+    Implementations must provide methods for:
+    - Adding vectors
+    - Querying by vector similarity
+    - Deleting vectors
+    - Getting count
+    """
+    
+    name: str
+    
+    def add(
+        self,
+        texts: List[str],
+        embeddings: List[List[float]],
+        metadatas: Optional[List[Dict[str, Any]]] = None,
+        ids: Optional[List[str]] = None,
+        namespace: Optional[str] = None
+    ) -> List[str]:
+        """
+        Add vectors to the store.
+        
+        Args:
+            texts: List of text content
+            embeddings: List of embedding vectors
+            metadatas: Optional list of metadata dicts
+            ids: Optional list of IDs (auto-generated if not provided)
+            namespace: Optional namespace for multi-tenant isolation
+            
+        Returns:
+            List of IDs for the added vectors
+        """
+        ...
+    
+    def query(
+        self,
+        embedding: List[float],
+        top_k: int = 10,
+        namespace: Optional[str] = None,
+        filter: Optional[Dict[str, Any]] = None
+    ) -> List[VectorRecord]:
+        """
+        Query vectors by similarity.
+        
+        Args:
+            embedding: Query embedding vector
+            top_k: Number of results to return
+            namespace: Optional namespace filter
+            filter: Optional metadata filter
+            
+        Returns:
+            List of VectorRecord objects sorted by similarity
+        """
+        ...
+    
+    def delete(
+        self,
+        ids: Optional[List[str]] = None,
+        namespace: Optional[str] = None,
+        filter: Optional[Dict[str, Any]] = None,
+        delete_all: bool = False
+    ) -> int:
+        """
+        Delete vectors from the store.
+        
+        Args:
+            ids: Optional list of IDs to delete
+            namespace: Optional namespace filter
+            filter: Optional metadata filter
+            delete_all: If True, delete all vectors
+            
+        Returns:
+            Number of vectors deleted
+        """
+        ...
+    
+    def count(self, namespace: Optional[str] = None) -> int:
+        """
+        Get the number of vectors in the store.
+        
+        Args:
+            namespace: Optional namespace filter
+            
+        Returns:
+            Number of vectors
+        """
+        ...
+    
+    def get(
+        self,
+        ids: List[str],
+        namespace: Optional[str] = None
+    ) -> List[VectorRecord]:
+        """
+        Get vectors by ID.
+        
+        Args:
+            ids: List of IDs to retrieve
+            namespace: Optional namespace filter
+            
+        Returns:
+            List of VectorRecord objects
+        """
+        ...
+
+
+class VectorStoreRegistry:
+    """
+    Registry for vector stores.
+    
+    Provides lazy loading and configuration-based initialization.
+    """
+    
+    _instance: Optional["VectorStoreRegistry"] = None
+    
+    def __new__(cls) -> "VectorStoreRegistry":
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+            cls._instance._stores: Dict[str, Callable[..., VectorStoreProtocol]] = {}
+            cls._instance._initialized_stores: Dict[str, VectorStoreProtocol] = {}
+        return cls._instance
+    
+    def register(
+        self, 
+        name: str, 
+        factory: Callable[..., VectorStoreProtocol]
+    ) -> None:
+        """
+        Register a vector store factory.
+        
+        Args:
+            name: Unique name for the store (e.g., "chroma", "pinecone")
+            factory: Callable that returns a VectorStoreProtocol instance
+        """
+        self._stores[name] = factory
+    
+    def get(
+        self, 
+        name: str, 
+        config: Optional[Dict[str, Any]] = None,
+        namespace: Optional[str] = None
+    ) -> Optional[VectorStoreProtocol]:
+        """
+        Get a vector store by name (lazy initialization).
+        
+        Args:
+            name: Store name
+            config: Optional configuration dict
+            namespace: Optional namespace for multi-tenant isolation
+            
+        Returns:
+            VectorStoreProtocol instance or None
+        """
+        cache_key = f"{name}:{namespace or 'default'}"
+        
+        if cache_key in self._initialized_stores:
+            return self._initialized_stores[cache_key]
+        
+        if name in self._stores:
+            try:
+                store = self._stores[name](config=config, namespace=namespace)
+                self._initialized_stores[cache_key] = store
+                return store
+            except Exception as e:
+                logger.warning(f"Failed to initialize vector store '{name}': {e}")
+                return None
+        return None
+    
+    def list_stores(self) -> List[str]:
+        """List all registered store names."""
+        return list(self._stores.keys())
+    
+    def clear(self) -> None:
+        """Clear all registered stores (mainly for testing)."""
+        self._stores.clear()
+        self._initialized_stores.clear()
+
+
+def get_vector_store_registry() -> VectorStoreRegistry:
+    """Get the global vector store registry instance."""
+    return VectorStoreRegistry()
+
+
+class InMemoryVectorStore:
+    """
+    Pure Python in-memory vector store for testing and default usage.
+    
+    Uses cosine similarity for queries. No external dependencies.
+    """
+    
+    name: str = "memory"
+    
+    def __init__(
+        self, 
+        config: Optional[Dict[str, Any]] = None,
+        namespace: Optional[str] = None
+    ):
+        self.config = config or {}
+        self.default_namespace = namespace or "default"
+        self._data: Dict[str, Dict[str, VectorRecord]] = {}  # namespace -> id -> record
+    
+    def _get_namespace_data(self, namespace: Optional[str] = None) -> Dict[str, VectorRecord]:
+        """Get or create namespace data."""
+        ns = namespace or self.default_namespace
+        if ns not in self._data:
+            self._data[ns] = {}
+        return self._data[ns]
+    
+    def add(
+        self,
+        texts: List[str],
+        embeddings: List[List[float]],
+        metadatas: Optional[List[Dict[str, Any]]] = None,
+        ids: Optional[List[str]] = None,
+        namespace: Optional[str] = None
+    ) -> List[str]:
+        """Add vectors to the store."""
+        import uuid
+        
+        ns_data = self._get_namespace_data(namespace)
+        metadatas = metadatas or [{} for _ in texts]
+        ids = ids or [str(uuid.uuid4()) for _ in texts]
+        
+        result_ids = []
+        for i, (text, embedding, metadata, id_) in enumerate(zip(texts, embeddings, metadatas, ids)):
+            record = VectorRecord(
+                id=id_,
+                text=text,
+                embedding=embedding,
+                metadata=metadata
+            )
+            ns_data[id_] = record
+            result_ids.append(id_)
+        
+        return result_ids
+    
+    def query(
+        self,
+        embedding: List[float],
+        top_k: int = 10,
+        namespace: Optional[str] = None,
+        filter: Optional[Dict[str, Any]] = None
+    ) -> List[VectorRecord]:
+        """Query vectors by cosine similarity."""
+        ns_data = self._get_namespace_data(namespace)
+        
+        if not ns_data:
+            return []
+        
+        # Calculate cosine similarity for all records
+        results = []
+        for record in ns_data.values():
+            # Apply metadata filter if provided
+            if filter:
+                match = True
+                for key, value in filter.items():
+                    if record.metadata.get(key) != value:
+                        match = False
+                        break
+                if not match:
+                    continue
+            
+            score = self._cosine_similarity(embedding, record.embedding)
+            results.append(VectorRecord(
+                id=record.id,
+                text=record.text,
+                embedding=record.embedding,
+                metadata=record.metadata,
+                score=score
+            ))
+        
+        # Sort by score descending and return top_k
+        results.sort(key=lambda x: x.score or 0, reverse=True)
+        return results[:top_k]
+    
+    def delete(
+        self,
+        ids: Optional[List[str]] = None,
+        namespace: Optional[str] = None,
+        filter: Optional[Dict[str, Any]] = None,
+        delete_all: bool = False
+    ) -> int:
+        """Delete vectors from the store."""
+        ns = namespace or self.default_namespace
+        
+        if delete_all:
+            if ns in self._data:
+                count = len(self._data[ns])
+                self._data[ns] = {}
+                return count
+            return 0
+        
+        if ns not in self._data:
+            return 0
+        
+        ns_data = self._data[ns]
+        deleted = 0
+        
+        if ids:
+            for id_ in ids:
+                if id_ in ns_data:
+                    del ns_data[id_]
+                    deleted += 1
+        elif filter:
+            to_delete = []
+            for id_, record in ns_data.items():
+                match = True
+                for key, value in filter.items():
+                    if record.metadata.get(key) != value:
+                        match = False
+                        break
+                if match:
+                    to_delete.append(id_)
+            for id_ in to_delete:
+                del ns_data[id_]
+                deleted += 1
+        
+        return deleted
+    
+    def count(self, namespace: Optional[str] = None) -> int:
+        """Get the number of vectors in the store."""
+        ns = namespace or self.default_namespace
+        if ns in self._data:
+            return len(self._data[ns])
+        return 0
+    
+    def get(
+        self,
+        ids: List[str],
+        namespace: Optional[str] = None
+    ) -> List[VectorRecord]:
+        """Get vectors by ID."""
+        ns_data = self._get_namespace_data(namespace)
+        results = []
+        for id_ in ids:
+            if id_ in ns_data:
+                results.append(ns_data[id_])
+        return results
+    
+    @staticmethod
+    def _cosine_similarity(a: List[float], b: List[float]) -> float:
+        """Calculate cosine similarity between two vectors."""
+        if len(a) != len(b):
+            return 0.0
+        
+        dot_product = sum(x * y for x, y in zip(a, b))
+        norm_a = sum(x * x for x in a) ** 0.5
+        norm_b = sum(x * x for x in b) ** 0.5
+        
+        if norm_a == 0 or norm_b == 0:
+            return 0.0
+        
+        return dot_product / (norm_a * norm_b)
+
+
+# Register the in-memory store by default
+def _register_default_stores():
+    """Register default vector stores."""
+    registry = get_vector_store_registry()
+    registry.register("memory", InMemoryVectorStore)
+
+
+_register_default_stores()

src/praisonai-agents/praisonaiagents/llm/__init__.py

@@ -27,6 +27,7 @@ from .model_router import (
     TaskComplexity,
     create_routing_agent
 )
+from .rate_limiter import RateLimiter
 
 __all__ = [
     "LLM", 
@@ -46,5 +47,6 @@ __all__ = [
     "ModelRouter",
     "ModelProfile",
     "TaskComplexity",
-    "create_routing_agent"
+    "create_routing_agent",
+    "RateLimiter"
 ]

src/praisonai-agents/praisonaiagents/llm/rate_limiter.py

@@ -0,0 +1,147 @@
+"""
+Rate Limiter for PraisonAI Agents.
+
+Provides token bucket rate limiting for LLM API calls.
+
+Zero Performance Impact:
+- When rate_limiter is None, no overhead
+- Uses monotonic clock for accuracy
+- Injectable clock/sleep for deterministic testing
+
+Usage:
+    from praisonaiagents.llm import RateLimiter
+    from praisonaiagents import Agent
+    
+    limiter = RateLimiter(requests_per_minute=60)
+    agent = Agent(name="Test", rate_limiter=limiter)
+"""
+
+import time
+import asyncio
+from dataclasses import dataclass, field
+from typing import Optional, Callable
+
+
+@dataclass
+class RateLimiter:
+    """Token bucket rate limiter for API calls.
+    
+    Uses the token bucket algorithm:
+    - Tokens are added at a fixed rate (requests_per_minute / 60 per second)
+    - Each request consumes one token
+    - If no tokens available, wait until one is available
+    - Burst allows multiple requests in quick succession
+    
+    Args:
+        requests_per_minute: Maximum requests per minute
+        tokens_per_minute: Optional token-based rate limiting (for future use)
+        burst: Maximum burst size (default: 1)
+    
+    Example:
+        limiter = RateLimiter(requests_per_minute=60)  # 1 req/sec
+        limiter.acquire()  # Blocks if rate exceeded
+        
+        # Async usage
+        await limiter.acquire_async()
+    """
+    requests_per_minute: int
+    tokens_per_minute: Optional[int] = None
+    burst: int = 1
+    
+    # Internal state
+    _tokens: float = field(default=None, init=False, repr=False)
+    _last_update: float = field(default=None, init=False, repr=False)
+    _lock: asyncio.Lock = field(default=None, init=False, repr=False)
+    
+    # Injectable functions for testing
+    _get_time: Callable[[], float] = field(default=None, init=False, repr=False)
+    _sleep: Callable[[float], None] = field(default=None, init=False, repr=False)
+    _async_sleep: Callable[[float], None] = field(default=None, init=False, repr=False)
+    
+    def __post_init__(self):
+        """Initialize internal state."""
+        self._tokens = float(self.burst)
+        self._last_update = time.monotonic()
+        self._lock = asyncio.Lock()
+        
+        # Default implementations (can be overridden for testing)
+        if self._get_time is None:
+            self._get_time = time.monotonic
+        if self._sleep is None:
+            self._sleep = time.sleep
+        if self._async_sleep is None:
+            self._async_sleep = asyncio.sleep
+    
+    @property
+    def _rate(self) -> float:
+        """Tokens per second."""
+        return self.requests_per_minute / 60.0
+    
+    def _refill(self) -> None:
+        """Refill tokens based on elapsed time."""
+        now = self._get_time()
+        elapsed = now - self._last_update
+        self._tokens = min(self.burst, self._tokens + elapsed * self._rate)
+        self._last_update = now
+    
+    def _wait_time(self) -> float:
+        """Calculate time to wait for next token."""
+        if self._tokens >= 1.0:
+            return 0.0
+        return (1.0 - self._tokens) / self._rate
+    
+    def acquire(self) -> None:
+        """Acquire a token, blocking if necessary.
+        
+        This is the synchronous version. Use acquire_async() for async code.
+        """
+        self._refill()
+        
+        wait = self._wait_time()
+        if wait > 0:
+            self._sleep(wait)
+            self._refill()
+        
+        self._tokens -= 1.0
+    
+    async def acquire_async(self) -> None:
+        """Acquire a token asynchronously, waiting if necessary.
+        
+        This is the async version. Use acquire() for sync code.
+        """
+        async with self._lock:
+            self._refill()
+            
+            wait = self._wait_time()
+            if wait > 0:
+                await self._async_sleep(wait)
+                self._refill()
+            
+            self._tokens -= 1.0
+    
+    def try_acquire(self) -> bool:
+        """Try to acquire a token without blocking.
+        
+        Returns:
+            True if token acquired, False if rate limit exceeded
+        """
+        self._refill()
+        
+        if self._tokens >= 1.0:
+            self._tokens -= 1.0
+            return True
+        return False
+    
+    def reset(self) -> None:
+        """Reset the rate limiter to initial state."""
+        self._tokens = float(self.burst)
+        self._last_update = self._get_time()
+    
+    @property
+    def available_tokens(self) -> float:
+        """Get current available tokens (for monitoring)."""
+        self._refill()
+        return self._tokens
+    
+    def __repr__(self) -> str:
+        return f"RateLimiter(rpm={self.requests_per_minute}, burst={self.burst})"

src/praisonai-agents/praisonaiagents/obs/__init__.py

@@ -0,0 +1,240 @@
+"""
+Observability interface for PraisonAI Agents.
+
+This module provides the protocol and types for observability/tracing.
+Implementations are provided by the wrapper layer (praisonai_tools.observability).
+
+Usage (simplest - recommended):
+    from praisonaiagents import Agent, obs
+    
+    agent = Agent(
+        name="Assistant",
+        observability=obs.auto(),  # Auto-detect from env vars
+    )
+    agent.chat("Hello!")  # Auto-traces to configured provider
+
+Alternative (explicit provider):
+    observability=obs.langfuse()     # Langfuse
+    observability=obs.langsmith()    # LangSmith
+    observability=obs.agentops()     # AgentOps
+    observability=obs.arize()        # Arize Phoenix
+    observability=obs.datadog()      # Datadog
+"""
+
+from typing import Any, Dict, List, Optional, Protocol, runtime_checkable
+from dataclasses import dataclass, field
+import time
+
+
+@dataclass
+class Span:
+    """Represents a span in a trace."""
+    span_id: str
+    trace_id: str
+    name: str
+    parent_span_id: Optional[str] = None
+    start_time: float = field(default_factory=time.time)
+    end_time: Optional[float] = None
+    attributes: Dict[str, Any] = field(default_factory=dict)
+    status: str = "ok"
+
+
+@dataclass
+class Trace:
+    """Represents a complete trace."""
+    trace_id: str
+    session_id: Optional[str] = None
+    agent_name: Optional[str] = None
+    user_id: Optional[str] = None
+    start_time: float = field(default_factory=time.time)
+    end_time: Optional[float] = None
+    spans: List[Span] = field(default_factory=list)
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    status: str = "ok"
+
+
+@runtime_checkable
+class ObservabilityAdapter(Protocol):
+    """Protocol for observability adapters."""
+    
+    def on_trace_start(
+        self,
+        trace_id: str,
+        session_id: Optional[str] = None,
+        agent_name: Optional[str] = None,
+        user_id: Optional[str] = None,
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Called when a trace starts."""
+        ...
+    
+    def on_trace_end(
+        self,
+        trace_id: str,
+        status: str = "ok",
+        metadata: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Called when a trace ends."""
+        ...
+    
+    def on_span_start(
+        self,
+        span_id: str,
+        trace_id: str,
+        name: str,
+        parent_span_id: Optional[str] = None,
+        attributes: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Called when a span starts."""
+        ...
+    
+    def on_span_end(
+        self,
+        span_id: str,
+        status: str = "ok",
+        attributes: Optional[Dict[str, Any]] = None,
+    ) -> None:
+        """Called when a span ends."""
+        ...
+    
+    def on_llm_call(
+        self,
+        span_id: str,
+        model: str,
+        messages: List[Dict[str, Any]],
+        response: Optional[str] = None,
+        tokens: Optional[Dict[str, int]] = None,
+        latency_ms: Optional[float] = None,
+    ) -> None:
+        """Called when an LLM call is made."""
+        ...
+    
+    def on_tool_call(
+        self,
+        span_id: str,
+        tool_name: str,
+        args: Dict[str, Any],
+        result: Any,
+        latency_ms: Optional[float] = None,
+    ) -> None:
+        """Called when a tool is invoked."""
+        ...
+    
+    def flush(self) -> None:
+        """Flush any pending data."""
+        ...
+    
+    def close(self) -> None:
+        """Close the adapter and release resources."""
+        ...
+
+
+class _LazyObsModule:
+    """
+    Lazy proxy for observability backends.
+    
+    Allows `from praisonaiagents import obs` without importing heavy deps.
+    Actual provider classes are loaded only when accessed.
+    """
+    
+    _PROVIDERS = {
+        "langfuse": "praisonai_tools.observability.providers.langfuse_provider",
+        "langsmith": "praisonai_tools.observability.providers.langsmith_provider",
+        "agentops": "praisonai_tools.observability.providers.agentops_provider",
+        "arize": "praisonai_tools.observability.providers.arize_phoenix_provider",
+        "phoenix": "praisonai_tools.observability.providers.arize_phoenix_provider",
+        "datadog": "praisonai_tools.observability.providers.datadog_provider",
+        "mlflow": "praisonai_tools.observability.providers.mlflow_provider",
+        "openlit": "praisonai_tools.observability.providers.openlit_provider",
+        "opik": "praisonai_tools.observability.providers.opik_provider",
+        "portkey": "praisonai_tools.observability.providers.portkey_provider",
+        "traceloop": "praisonai_tools.observability.providers.traceloop_provider",
+        "weave": "praisonai_tools.observability.providers.weave_provider",
+        "braintrust": "praisonai_tools.observability.providers.braintrust_provider",
+        "langtrace": "praisonai_tools.observability.providers.langtrace_provider",
+        "langwatch": "praisonai_tools.observability.providers.langwatch_provider",
+        "maxim": "praisonai_tools.observability.providers.maxim_provider",
+    }
+    
+    def __getattr__(self, name: str):
+        if name in self._PROVIDERS:
+            return self._create_provider_factory(name)
+        raise AttributeError(f"module 'obs' has no attribute {name!r}")
+    
+    def _create_provider_factory(self, provider_name: str):
+        """Create a factory function for a provider."""
+        def factory(**kwargs):
+            try:
+                import importlib
+                module = importlib.import_module(self._PROVIDERS[provider_name])
+                # Get the provider class (naming convention: XxxProvider)
+                class_name = f"{provider_name.title().replace('_', '')}Provider"
+                provider_cls = getattr(module, class_name, None)
+                if provider_cls:
+                    return provider_cls(**kwargs)
+                # Fallback: try to find any class ending with Provider
+                for attr_name in dir(module):
+                    if attr_name.endswith("Provider"):
+                        return getattr(module, attr_name)(**kwargs)
+                raise AttributeError(f"No provider class found in {module}")
+            except ImportError as e:
+                raise ImportError(
+                    f"Observability provider '{provider_name}' requires praisonai-tools. "
+                    f"Install with: pip install praisonai-tools\n"
+                    f"Original error: {e}"
+                ) from e
+        return factory
+    
+    def auto(self, **kwargs):
+        """
+        Auto-detect observability provider from environment variables.
+        
+        Checks for common env vars like LANGFUSE_PUBLIC_KEY, LANGSMITH_API_KEY, etc.
+        Returns None if no provider is detected.
+        """
+        import os
+        
+        # Check for common provider env vars
+        provider_env_map = {
+            "LANGFUSE_PUBLIC_KEY": "langfuse",
+            "LANGSMITH_API_KEY": "langsmith",
+            "AGENTOPS_API_KEY": "agentops",
+            "ARIZE_API_KEY": "arize",
+            "DATADOG_API_KEY": "datadog",
+            "MLFLOW_TRACKING_URI": "mlflow",
+            "OPENLIT_API_KEY": "openlit",
+            "OPIK_API_KEY": "opik",
+            "PORTKEY_API_KEY": "portkey",
+            "TRACELOOP_API_KEY": "traceloop",
+            "WANDB_API_KEY": "weave",
+            "BRAINTRUST_API_KEY": "braintrust",
+        }
+        
+        for env_var, provider in provider_env_map.items():
+            if os.getenv(env_var):
+                return self._create_provider_factory(provider)(**kwargs)
+        
+        return None
+    
+    def __repr__(self):
+        return "<module 'praisonaiagents.obs' (lazy observability backends)>"
+    
+    def __call__(self, provider: str = "auto", **kwargs):
+        """
+        Shortcut: obs("langfuse") or obs() for auto-detection.
+        """
+        if provider == "auto":
+            return self.auto(**kwargs)
+        return self._create_provider_factory(provider)(**kwargs)
+
+
+# Singleton lazy module instance
+obs = _LazyObsModule()
+
+
+__all__ = [
+    "ObservabilityAdapter",
+    "Span",
+    "Trace",
+    "obs",
+]

src/praisonai-agents/praisonaiagents/session.py

@@ -4,16 +4,20 @@ Session Management for PraisonAI Agents
 A simple wrapper around existing stateful capabilities to provide a unified
 session API for developers building stateful agent applications.
 """
+from __future__ import annotations
 
 import os
 import uuid
 import requests
 import json
 import time
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Optional, TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from .memory.memory import Memory
+    from .knowledge.knowledge import Knowledge
+
 from .agent import Agent
-from .memory import Memory
-from .knowledge import Knowledge
 
 
 class Session:
@@ -112,6 +116,7 @@ class Session:
         if self.is_remote:
             raise ValueError("Memory operations are not available for remote agent sessions")
         if self._memory is None:
+            from .memory.memory import Memory
             self._memory = Memory(config=self.memory_config)
         return self._memory
 
@@ -121,6 +126,7 @@ class Session:
         if self.is_remote:
             raise ValueError("Knowledge operations are not available for remote agent sessions")
         if self._knowledge is None:
+            from .knowledge.knowledge import Knowledge
             self._knowledge = Knowledge(config=self.knowledge_config)
         return self._knowledge
 

src/praisonai-agents/praisonaiagents/tools/__init__.py

@@ -2,6 +2,9 @@
 from importlib import import_module
 from typing import Any
 
+# Export Injected type for tool state injection
+from .injected import Injected
+
 # Map of function names to their module and class (if any)
 TOOL_MAPPINGS = {
     # Direct functions
@@ -286,4 +289,4 @@ def __getattr__(name: str) -> Any:
         method = getattr(_instances[class_name], name)
         return method
 
-__all__ = list(TOOL_MAPPINGS.keys())
+__all__ = list(TOOL_MAPPINGS.keys()) + ['Injected']

src/praisonai-agents/praisonaiagents/tools/decorator.py

@@ -14,6 +14,14 @@ Usage:
     @tool(name="web_search", description="Search the internet")
     def search(query: str, max_results: int = 5) -> list:
         return [...]
+    
+    # With injected state:
+    from praisonaiagents.tools import Injected
+    
+    @tool
+    def my_tool(query: str, state: Injected[dict]) -> str:
+        '''Tool with injected state.'''
+        return f"session={state.get('session_id')}"
 """
 
 import inspect
@@ -22,12 +30,19 @@ import logging
 from typing import Any, Callable, Dict, Optional, Union, get_type_hints
 
 from .base import BaseTool
+from .injected import (
+    is_injected_type, 
+    get_injected_params, 
+    inject_state_into_kwargs,
+    filter_injected_from_schema
+)
 
 
 class FunctionTool(BaseTool):
     """A BaseTool wrapper for plain functions.
     
     Created automatically by the @tool decorator.
+    Supports Injected[T] parameters for state injection.
     """
     
     def __init__(
@@ -42,6 +57,9 @@ class FunctionTool(BaseTool):
         self.description = description or func.__doc__ or f"Tool: {self.name}"
         self.version = version
         
+        # Detect injected parameters
+        self._injected_params = get_injected_params(func)
+        
         # Generate schema from the original function, not run()
         self.parameters = self._generate_schema_from_func(func)
         
@@ -55,8 +73,16 @@ class FunctionTool(BaseTool):
         if not self.description:
             self.description = self.__class__.__doc__ or f"Tool: {self.name}"
     
+    @property
+    def injected_params(self) -> Dict[str, Any]:
+        """Get the injected parameters for this tool."""
+        return self._injected_params
+    
     def _generate_schema_from_func(self, func: Callable) -> Dict[str, Any]:
-        """Generate JSON Schema from the wrapped function's signature."""
+        """Generate JSON Schema from the wrapped function's signature.
+        
+        Injected parameters are excluded from the schema.
+        """
         schema = {
             "type": "object",
             "properties": {},
@@ -71,8 +97,17 @@ class FunctionTool(BaseTool):
                 if param_name in ('self', 'cls'):
                     continue
                 
+                # Skip injected parameters - they don't go in schema
+                if param_name in self._injected_params:
+                    continue
+                
                 # Get type hint
                 param_type = hints.get(param_name, Any)
+                
+                # Double-check it's not an Injected type
+                if is_injected_type(param_type):
+                    continue
+                
                 json_type = BaseTool._python_type_to_json(param_type)
                 
                 schema["properties"][param_name] = {"type": json_type}
@@ -86,11 +121,18 @@ class FunctionTool(BaseTool):
         return schema
     
     def run(self, **kwargs) -> Any:
-        """Execute the wrapped function."""
+        """Execute the wrapped function with injected state."""
+        # Inject state for any Injected parameters
+        kwargs = inject_state_into_kwargs(kwargs, self._injected_params)
         return self._func(**kwargs)
     
     def __call__(self, *args, **kwargs) -> Any:
-        """Allow calling with positional args like the original function."""
+        """Allow calling with positional args like the original function.
+        
+        Injects state for Injected parameters.
+        """
+        # Inject state for any Injected parameters
+        kwargs = inject_state_into_kwargs(kwargs, self._injected_params)
         return self._func(*args, **kwargs)
 
 

src/praisonai-agents/praisonaiagents/tools/injected.py

@@ -0,0 +1,213 @@
+"""
+Injected State for PraisonAI Tools.
+
+Provides a mechanism for tools to receive injected state (agent context)
+without exposing it in the tool's public schema.
+
+Zero Performance Impact:
+- When no Injected params, no overhead
+- Context is thread-local for safety
+- Lazy evaluation of state
+
+Usage:
+    from praisonaiagents import tool
+    from praisonaiagents.tools import Injected
+    
+    @tool
+    def my_tool(query: str, state: Injected[dict]) -> str:
+        '''Tool with injected state.'''
+        session_id = state.get('session_id')
+        return f"Query: {query}, Session: {session_id}"
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, Generic, List, Optional, TypeVar, get_type_hints, get_origin, get_args
+from contextvars import ContextVar
+from contextlib import contextmanager
+
+
+T = TypeVar('T')
+
+
+class Injected(Generic[T]):
+    """Type marker for injected parameters.
+    
+    Parameters annotated with Injected[T] will:
+    - Not appear in the tool's public schema
+    - Be automatically injected at runtime with the current agent state
+    
+    Supported types:
+    - Injected[dict] - Receives state as a dictionary
+    - Injected[AgentState] - Receives state as AgentState dataclass
+    
+    Example:
+        @tool
+        def my_tool(query: str, state: Injected[dict]) -> str:
+            return f"session={state.get('session_id')}"
+    """
+    pass
+
+
+@dataclass
+class AgentState:
+    """State object injected into tools.
+    
+    Contains context about the current agent execution.
+    """
+    agent_id: str
+    run_id: str
+    session_id: str
+    last_user_message: Optional[str] = None
+    last_agent_message: Optional[str] = None
+    memory: Optional[Any] = None
+    previous_tool_results: List[Any] = field(default_factory=list)
+    metadata: Dict[str, Any] = field(default_factory=dict)
+    
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert to dictionary for Injected[dict] usage."""
+        return {
+            'agent_id': self.agent_id,
+            'run_id': self.run_id,
+            'session_id': self.session_id,
+            'last_user_message': self.last_user_message,
+            'last_agent_message': self.last_agent_message,
+            'previous_tool_results': self.previous_tool_results,
+            'metadata': self.metadata,
+        }
+
+
+# Thread-local context for current state
+_current_state: ContextVar[Optional[AgentState]] = ContextVar('current_state', default=None)
+
+
+def get_current_state() -> Optional[AgentState]:
+    """Get the current injection state."""
+    return _current_state.get()
+
+
+def set_current_state(state: Optional[AgentState]) -> None:
+    """Set the current injection state."""
+    _current_state.set(state)
+
+
+@contextmanager
+def with_injection_context(state: AgentState):
+    """Context manager for setting injection state.
+    
+    Usage:
+        with with_injection_context(state):
+            result = my_tool(query="test")
+    """
+    token = _current_state.set(state)
+    try:
+        yield
+    finally:
+        _current_state.reset(token)
+
+
+def is_injected_type(annotation: Any) -> bool:
+    """Check if a type annotation is Injected[T]."""
+    origin = get_origin(annotation)
+    if origin is Injected:
+        return True
+    # Handle string annotations
+    if isinstance(annotation, str) and 'Injected' in annotation:
+        return True
+    return False
+
+
+def get_injected_type(annotation: Any) -> Optional[type]:
+    """Get the inner type T from Injected[T]."""
+    origin = get_origin(annotation)
+    if origin is Injected:
+        args = get_args(annotation)
+        if args:
+            return args[0]
+    return None
+
+
+def resolve_injected_value(annotation: Any, state: Optional[AgentState]) -> Any:
+    """Resolve the value to inject based on annotation type.
+    
+    Args:
+        annotation: The Injected[T] annotation
+        state: The current agent state
+        
+    Returns:
+        The resolved value (dict or AgentState)
+    """
+    if state is None:
+        # Return empty dict if no state available
+        return {}
+    
+    inner_type = get_injected_type(annotation)
+    
+    if inner_type is dict or inner_type is Dict:
+        return state.to_dict()
+    elif inner_type is AgentState:
+        return state
+    else:
+        # Default to dict
+        return state.to_dict()
+
+
+def get_injected_params(func) -> Dict[str, Any]:
+    """Get all injected parameters from a function.
+    
+    Returns:
+        Dict mapping param name to its Injected[T] annotation
+    """
+    try:
+        hints = get_type_hints(func)
+    except Exception:
+        # If we can't get hints, return empty
+        return {}
+    
+    injected = {}
+    for name, hint in hints.items():
+        if name == 'return':
+            continue
+        if is_injected_type(hint):
+            injected[name] = hint
+    
+    return injected
+
+
+def filter_injected_from_schema(properties: Dict[str, Any], required: List[str], 
+                                 injected_params: Dict[str, Any]) -> tuple:
+    """Remove injected parameters from schema.
+    
+    Args:
+        properties: Schema properties dict
+        required: List of required param names
+        injected_params: Dict of injected param names
+        
+    Returns:
+        Tuple of (filtered_properties, filtered_required)
+    """
+    filtered_props = {k: v for k, v in properties.items() if k not in injected_params}
+    filtered_required = [r for r in required if r not in injected_params]
+    return filtered_props, filtered_required
+
+
+def inject_state_into_kwargs(kwargs: Dict[str, Any], injected_params: Dict[str, Any]) -> Dict[str, Any]:
+    """Inject state values into kwargs for injected parameters.
+    
+    Args:
+        kwargs: The original kwargs
+        injected_params: Dict of injected param names to annotations
+        
+    Returns:
+        Updated kwargs with injected values
+    """
+    if not injected_params:
+        return kwargs
+    
+    state = get_current_state()
+    result = kwargs.copy()
+    
+    for param_name, annotation in injected_params.items():
+        if param_name not in result:
+            result[param_name] = resolve_injected_value(annotation, state)
+    
+    return result

src/praisonai-agents/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "praisonaiagents"
-version = "0.2.0"
+version = "0.3.4"
 description = "Praison AI agents for completing complex tasks with Self Reflection Agents"
 readme = "README.md"
 requires-python = ">=3.10"

src/praisonai-agents/tests/test_planning.py

@@ -672,16 +672,16 @@ class TestPlanningAgent:
         
         planner = PlanningAgent()
         
-        assert planner.llm_model == "gpt-4o-mini"
+        assert planner.llm_model in ["gpt-4o-mini", "gpt-5-nano"]
         assert planner.read_only is True
         
     def test_planning_agent_custom_llm(self):
         """Test PlanningAgent with custom LLM."""
         from praisonaiagents.planning import PlanningAgent
         
-        planner = PlanningAgent(llm="gpt-4o")
+        planner = PlanningAgent(llm="gpt-5-nano")
         
-        assert planner.llm_model == "gpt-4o"
+        assert planner.llm_model == "gpt-5-nano"
         
     def test_planning_agent_allowed_tools(self):
         """Test PlanningAgent has only read-only tools."""
@@ -840,10 +840,10 @@ class TestPraisonAIAgentsPlanningIntegration:
             agents=[agent],
             tasks=[task],
             planning=True,
-            planning_llm="gpt-4o"
+            planning_llm="gpt-5-nano"
         )
         
-        assert agents.planning_llm == "gpt-4o"
+        assert agents.planning_llm == "gpt-5-nano"
         
     def test_agents_auto_approve_plan(self):
         """Test auto-approve plan option."""
@@ -867,7 +867,7 @@ class TestPraisonAIAgentsPlanningIntegration:
         from praisonaiagents import Agent, Task, PraisonAIAgents
         from praisonaiagents.planning import Plan, PlanStep
         
-        agent = Agent(name="Test", role="Tester", llm="gpt-4o-mini")
+        agent = Agent(name="Test", role="Tester", llm="gpt-5-nano")
         task = Task(description="Test task", agent=agent)
         
         agents = PraisonAIAgents(
@@ -880,7 +880,7 @@ class TestPraisonAIAgentsPlanningIntegration:
         # Test that planning is enabled and configured correctly
         assert agents.planning is True
         assert agents.auto_approve_plan is True
-        assert agents.planning_llm == "gpt-4o-mini"
+        assert agents.planning_llm == "gpt-5-nano"
         assert agents._current_plan is None  # No plan created yet
         assert agents._todo_list is None
                 

src/praisonai-agents/tests/test_workflow_patterns.py

@@ -1189,16 +1189,16 @@ class TestWorkflowConfiguration:
         """Test that workflow planning setting is stored."""
         def step1(ctx): return StepResult(output="Done")
         
-        workflow = Workflow(steps=[step1], planning=True, planning_llm="gpt-4o")
+        workflow = Workflow(steps=[step1], planning=True, planning_llm="gpt-5-nano")
         assert workflow.planning == True
-        assert workflow.planning_llm == "gpt-4o"
+        assert workflow.planning_llm == "gpt-5-nano"
     
     def test_workflow_default_llm(self):
         """Test default LLM setting."""
         def step1(ctx): return StepResult(output="Done")
         
-        workflow = Workflow(steps=[step1], default_llm="gpt-4o-mini")
-        assert workflow.default_llm == "gpt-4o-mini"
+        workflow = Workflow(steps=[step1], default_llm="gpt-5-nano")
+        assert workflow.default_llm == "gpt-5-nano"
     
     def test_workflow_default_agent_config(self):
         """Test default agent config."""
@@ -1393,10 +1393,10 @@ class TestPlanningAndReasoning:
         workflow = Workflow(
             steps=[lambda ctx: StepResult(output="test")],
             planning=True,
-            planning_llm="gpt-4o"
+            planning_llm="gpt-5-nano"
         )
         assert workflow.planning == True
-        assert workflow.planning_llm == "gpt-4o"
+        assert workflow.planning_llm == "gpt-5-nano"
     
     def test_workflow_reasoning_parameter_exists(self):
         """Test that Workflow accepts reasoning parameter."""

src/praisonai-agents/tests/unit/eval/test_evaluators.py

@@ -91,14 +91,11 @@ class TestAccuracyEvaluator:
     
     def test_judge_output(self):
         """Test LLM judging of output."""
-        with patch('openai.OpenAI') as mock_openai_class:
-            mock_client = MagicMock()
-            mock_openai_class.return_value = mock_client
-            
+        with patch('litellm.completion') as mock_completion:
             mock_response = MagicMock()
             mock_response.choices = [MagicMock()]
             mock_response.choices[0].message.content = "SCORE: 8\nREASONING: Good match"
-            mock_client.chat.completions.create.return_value = mock_response
+            mock_completion.return_value = mock_response
             
             def test_func(x):
                 return "4"
@@ -115,14 +112,11 @@ class TestAccuracyEvaluator:
     
     def test_run_returns_accuracy_result(self):
         """Test that run returns AccuracyResult."""
-        with patch('openai.OpenAI') as mock_openai_class:
-            mock_client = MagicMock()
-            mock_openai_class.return_value = mock_client
-            
+        with patch('litellm.completion') as mock_completion:
             mock_response = MagicMock()
             mock_response.choices = [MagicMock()]
             mock_response.choices[0].message.content = "SCORE: 9\nREASONING: Excellent"
-            mock_client.chat.completions.create.return_value = mock_response
+            mock_completion.return_value = mock_response
             
             def test_func(x):
                 return "4"
@@ -140,14 +134,11 @@ class TestAccuracyEvaluator:
     
     def test_evaluate_output(self):
         """Test evaluate_output method for pre-generated outputs."""
-        with patch('openai.OpenAI') as mock_openai_class:
-            mock_client = MagicMock()
-            mock_openai_class.return_value = mock_client
-            
+        with patch('litellm.completion') as mock_completion:
             mock_response = MagicMock()
             mock_response.choices = [MagicMock()]
             mock_response.choices[0].message.content = "SCORE: 10\nREASONING: Perfect"
-            mock_client.chat.completions.create.return_value = mock_response
+            mock_completion.return_value = mock_response
             
             def test_func(x):
                 return "unused"
@@ -325,14 +316,11 @@ class TestCriteriaEvaluator:
     
     def test_judge_output_numeric(self):
         """Test numeric scoring."""
-        with patch('openai.OpenAI') as mock_openai_class:
-            mock_client = MagicMock()
-            mock_openai_class.return_value = mock_client
-            
+        with patch('litellm.completion') as mock_completion:
             mock_response = MagicMock()
             mock_response.choices = [MagicMock()]
             mock_response.choices[0].message.content = "SCORE: 8\nREASONING: Good"
-            mock_client.chat.completions.create.return_value = mock_response
+            mock_completion.return_value = mock_response
             
             def test_func(x):
                 return "test"
@@ -350,14 +338,11 @@ class TestCriteriaEvaluator:
     
     def test_judge_output_binary(self):
         """Test binary scoring."""
-        with patch('openai.OpenAI') as mock_openai_class:
-            mock_client = MagicMock()
-            mock_openai_class.return_value = mock_client
-            
+        with patch('litellm.completion') as mock_completion:
             mock_response = MagicMock()
             mock_response.choices = [MagicMock()]
             mock_response.choices[0].message.content = "RESULT: PASS\nREASONING: Meets criteria"
-            mock_client.chat.completions.create.return_value = mock_response
+            mock_completion.return_value = mock_response
             
             def test_func(x):
                 return "test"
@@ -374,14 +359,11 @@ class TestCriteriaEvaluator:
     
     def test_on_fail_callback(self):
         """Test on_fail callback is called."""
-        with patch('openai.OpenAI') as mock_openai_class:
-            mock_client = MagicMock()
-            mock_openai_class.return_value = mock_client
-            
+        with patch('litellm.completion') as mock_completion:
             mock_response = MagicMock()
             mock_response.choices = [MagicMock()]
             mock_response.choices[0].message.content = "SCORE: 3\nREASONING: Poor"
-            mock_client.chat.completions.create.return_value = mock_response
+            mock_completion.return_value = mock_response
             
             on_fail_called = []
             def on_fail(score):
@@ -403,14 +385,11 @@ class TestCriteriaEvaluator:
     
     def test_evaluate_output_criteria(self):
         """Test evaluate_output method."""
-        with patch('openai.OpenAI') as mock_openai_class:
-            mock_client = MagicMock()
-            mock_openai_class.return_value = mock_client
-            
+        with patch('litellm.completion') as mock_completion:
             mock_response = MagicMock()
             mock_response.choices = [MagicMock()]
             mock_response.choices[0].message.content = "SCORE: 9\nREASONING: Great"
-            mock_client.chat.completions.create.return_value = mock_response
+            mock_completion.return_value = mock_response
             
             def test_func(x):
                 return "unused"

src/praisonai-agents/tests/unit/test_configurable_model.py

@@ -0,0 +1,234 @@
+"""
+TDD Tests for Configurable Model (runtime model switching).
+
+These tests are written FIRST before implementation.
+"""
+
+import pytest
+from unittest.mock import Mock, patch
+
+
+class TestConfigurableModelBasic:
+    """Test basic configurable model functionality."""
+    
+    def test_agent_accepts_llm_config_configurable(self):
+        """Agent should accept llm_config with configurable=True."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o",
+            llm_config={"configurable": True}
+        )
+        assert agent is not None
+        assert getattr(agent, '_llm_configurable', False) is True
+    
+    def test_agent_chat_accepts_config_parameter(self):
+        """Agent.chat() should accept config parameter for model override."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o",
+            llm_config={"configurable": True}
+        )
+        
+        # Should not raise - config parameter should be accepted
+        # We'll mock the actual LLM call
+        with patch.object(agent, '_chat_completion', return_value="mocked"):
+            result = agent.chat(
+                "Hello",
+                config={"model": "claude-3-5-sonnet"}
+            )
+    
+    def test_config_model_override(self):
+        """config.model should override default model for that call."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o",
+            llm_config={"configurable": True}
+        )
+        
+        # Track which model was used
+        used_model = [None]
+        
+        original_chat = agent._chat_completion
+        def mock_chat(*args, **kwargs):
+            # The model should be accessible somehow
+            used_model[0] = getattr(agent, '_current_model', agent.llm)
+            return "response"
+        
+        with patch.object(agent, '_chat_completion', side_effect=mock_chat):
+            agent.chat("Hello", config={"model": "claude-3-5-sonnet"})
+        
+        # Model should have been switched for this call
+        # Implementation will set _current_model or similar
+    
+    def test_config_does_not_mutate_default(self):
+        """Per-call config should not change agent's default model."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o",
+            llm_config={"configurable": True}
+        )
+        
+        original_llm = agent.llm
+        
+        with patch.object(agent, '_chat_completion', return_value="mocked"):
+            agent.chat("Hello", config={"model": "claude-3-5-sonnet"})
+        
+        # Default should be unchanged
+        assert agent.llm == original_llm
+    
+    def test_config_temperature_override(self):
+        """config.temperature should override for that call."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o",
+            llm_config={"configurable": True}
+        )
+        
+        # Should accept temperature in config
+        with patch.object(agent, '_chat_completion', return_value="mocked") as mock:
+            agent.chat("Hello", config={"temperature": 0.2})
+            # Temperature should be passed through
+
+
+class TestConfigurableModelCaching:
+    """Test model client caching for configurable models."""
+    
+    def test_model_client_cached(self):
+        """Model clients should be cached to avoid recreation."""
+        from praisonaiagents.llm import LLM
+        
+        llm = LLM(model="gpt-4o", configurable=True)
+        
+        # First call with model A
+        # Second call with model A should reuse client
+        # This is internal behavior - we test via call count or similar
+    
+    def test_cache_size_limited(self):
+        """Model client cache should have a size limit."""
+        from praisonaiagents.llm import LLM
+        
+        llm = LLM(model="gpt-4o", configurable=True, cache_size=3)
+        
+        # Should not grow unbounded
+
+
+class TestConfigurableModelThreadSafety:
+    """Test thread safety of configurable models."""
+    
+    def test_concurrent_calls_different_models(self):
+        """Concurrent calls with different models should not interfere."""
+        import threading
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o",
+            llm_config={"configurable": True}
+        )
+        
+        results = {}
+        errors = []
+        
+        def call_with_model(model_name, thread_id):
+            try:
+                with patch.object(agent, '_chat_completion', return_value=f"response_{model_name}"):
+                    result = agent.chat("Hello", config={"model": model_name})
+                    results[thread_id] = result
+            except Exception as e:
+                errors.append(e)
+        
+        threads = [
+            threading.Thread(target=call_with_model, args=("model_a", 1)),
+            threading.Thread(target=call_with_model, args=("model_b", 2)),
+        ]
+        
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join()
+        
+        assert len(errors) == 0
+
+
+class TestConfigurableModelValidation:
+    """Test validation of configurable model settings."""
+    
+    def test_config_without_configurable_flag_ignored(self):
+        """If llm_config.configurable is False, config param should be ignored."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o"
+            # No llm_config or configurable=False
+        )
+        
+        # Should still work, just ignore the config
+        with patch.object(agent, '_chat_completion', return_value="mocked"):
+            result = agent.chat("Hello", config={"model": "other"})
+            # Should use default model, not "other"
+    
+    def test_invalid_config_keys_ignored(self):
+        """Unknown keys in config should be ignored, not error."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o",
+            llm_config={"configurable": True}
+        )
+        
+        # Should not raise for unknown keys
+        with patch.object(agent, '_chat_completion', return_value="mocked"):
+            result = agent.chat("Hello", config={"unknown_key": "value"})
+
+
+class TestConfigurableModelProvider:
+    """Test provider switching in configurable models."""
+    
+    def test_config_provider_override(self):
+        """config.provider should allow switching providers."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o",
+            llm_config={"configurable": True}
+        )
+        
+        # Should accept provider in config
+        with patch.object(agent, '_chat_completion', return_value="mocked"):
+            agent.chat("Hello", config={"provider": "anthropic", "model": "claude-3-5-sonnet"})
+    
+    def test_model_string_with_provider_prefix(self):
+        """Model string like 'anthropic/claude-3' should work."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            llm="gpt-4o",
+            llm_config={"configurable": True}
+        )
+        
+        with patch.object(agent, '_chat_completion', return_value="mocked"):
+            agent.chat("Hello", config={"model": "anthropic/claude-3-5-sonnet"})

src/praisonai-agents/tests/unit/test_injected_state.py

@@ -0,0 +1,367 @@
+"""
+TDD Tests for InjectedState for Tools.
+
+These tests are written FIRST before implementation.
+"""
+
+import pytest
+from unittest.mock import Mock, patch
+
+
+class TestInjectedTypeExists:
+    """Test that Injected type marker exists."""
+    
+    def test_injected_type_exists(self):
+        """Injected type should exist in praisonaiagents.tools."""
+        from praisonaiagents.tools import Injected
+        
+        # Should be a generic type marker
+        assert Injected is not None
+    
+    def test_injected_generic_usage(self):
+        """Injected should work as a generic type annotation."""
+        from praisonaiagents.tools import Injected
+        
+        # Should be usable as Injected[dict], Injected[SomeType], etc.
+        def my_tool(query: str, state: Injected[dict]) -> str:
+            return query
+        
+        # Should have annotations
+        annotations = my_tool.__annotations__
+        assert 'state' in annotations
+    
+    def test_injected_with_agent_state(self):
+        """Injected[AgentState] should work."""
+        from praisonaiagents.tools import Injected
+        from praisonaiagents.tools.injected import AgentState
+        
+        def my_tool(query: str, state: Injected[AgentState]) -> str:
+            return query
+        
+        annotations = my_tool.__annotations__
+        assert 'state' in annotations
+
+
+class TestAgentStateType:
+    """Test AgentState type for injection."""
+    
+    def test_agent_state_exists(self):
+        """AgentState type should exist."""
+        from praisonaiagents.tools.injected import AgentState
+        
+        assert AgentState is not None
+    
+    def test_agent_state_has_required_fields(self):
+        """AgentState should have agent_id, run_id, session_id."""
+        from praisonaiagents.tools.injected import AgentState
+        
+        state = AgentState(
+            agent_id="agent-123",
+            run_id="run-456",
+            session_id="session-789"
+        )
+        
+        assert state.agent_id == "agent-123"
+        assert state.run_id == "run-456"
+        assert state.session_id == "session-789"
+    
+    def test_agent_state_optional_fields(self):
+        """AgentState should have optional fields."""
+        from praisonaiagents.tools.injected import AgentState
+        
+        state = AgentState(
+            agent_id="a",
+            run_id="r",
+            session_id="s",
+            last_user_message="Hello",
+            last_agent_message="Hi there",
+            memory=Mock(),
+            previous_tool_results=["result1", "result2"]
+        )
+        
+        assert state.last_user_message == "Hello"
+        assert state.last_agent_message == "Hi there"
+        assert state.memory is not None
+        assert len(state.previous_tool_results) == 2
+
+
+class TestToolDecoratorInjection:
+    """Test @tool decorator handles Injected parameters."""
+    
+    def test_tool_decorator_detects_injected(self):
+        """@tool should detect Injected parameters."""
+        from praisonaiagents import tool
+        from praisonaiagents.tools import Injected
+        
+        @tool
+        def my_tool(query: str, state: Injected[dict]) -> str:
+            """A tool with injected state."""
+            return f"query={query}, session={state.get('session_id')}"
+        
+        # Should have metadata about injected params
+        assert hasattr(my_tool, '_injected_params') or hasattr(my_tool, 'injected_params')
+    
+    def test_injected_params_not_in_schema(self):
+        """Injected parameters should not appear in tool schema."""
+        from praisonaiagents import tool
+        from praisonaiagents.tools import Injected
+        
+        @tool
+        def my_tool(query: str, state: Injected[dict]) -> str:
+            """A tool with injected state."""
+            return query
+        
+        schema = my_tool.get_schema()
+        params = schema['function']['parameters']['properties']
+        
+        # 'query' should be in schema
+        assert 'query' in params
+        # 'state' should NOT be in schema (it's injected)
+        assert 'state' not in params
+    
+    def test_injected_params_not_required(self):
+        """Injected parameters should not be in required list."""
+        from praisonaiagents import tool
+        from praisonaiagents.tools import Injected
+        
+        @tool
+        def my_tool(query: str, state: Injected[dict]) -> str:
+            """A tool with injected state."""
+            return query
+        
+        schema = my_tool.get_schema()
+        required = schema['function']['parameters'].get('required', [])
+        
+        assert 'state' not in required
+        assert 'query' in required
+
+
+class TestToolExecutionInjection:
+    """Test that injected state is provided at runtime."""
+    
+    def test_injected_state_provided_at_runtime(self):
+        """When tool is called, injected state should be provided."""
+        from praisonaiagents import tool
+        from praisonaiagents.tools import Injected
+        from praisonaiagents.tools.injected import AgentState
+        
+        received_state = [None]
+        
+        @tool
+        def my_tool(query: str, state: Injected[AgentState]) -> str:
+            """A tool with injected state."""
+            received_state[0] = state
+            return f"session={state.session_id}"
+        
+        # Simulate execution with injection context
+        from praisonaiagents.tools.injected import with_injection_context
+        
+        mock_state = AgentState(
+            agent_id="a",
+            run_id="r",
+            session_id="test-session"
+        )
+        
+        with with_injection_context(mock_state):
+            result = my_tool(query="hello")
+        
+        assert received_state[0] is not None
+        assert received_state[0].session_id == "test-session"
+    
+    def test_injected_dict_state(self):
+        """Injected[dict] should provide dict with state info."""
+        from praisonaiagents import tool
+        from praisonaiagents.tools import Injected
+        from praisonaiagents.tools.injected import with_injection_context, AgentState
+        
+        received_state = [None]
+        
+        @tool
+        def my_tool(query: str, state: Injected[dict]) -> str:
+            """A tool with injected dict state."""
+            received_state[0] = state
+            return "ok"
+        
+        mock_state = AgentState(
+            agent_id="a",
+            run_id="r",
+            session_id="s"
+        )
+        
+        with with_injection_context(mock_state):
+            my_tool(query="test")
+        
+        assert isinstance(received_state[0], dict)
+        assert 'session_id' in received_state[0]
+
+
+class TestAgentToolExecution:
+    """Test Agent executes tools with injection."""
+    
+    def test_agent_injects_state_to_tools(self):
+        """Agent should inject state when executing tools."""
+        from praisonaiagents import Agent, tool
+        from praisonaiagents.tools import Injected
+        
+        received_session = [None]
+        
+        @tool
+        def show_state(query: str, state: Injected[dict]) -> str:
+            """Show the injected state."""
+            received_session[0] = state.get('session_id')
+            return f"session={state.get('session_id')}"
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            tools=[show_state],
+            session_id="my-session-123"
+        )
+        
+        # Execute the tool directly through agent
+        result = agent.execute_tool("show_state", {"query": "test"})
+        
+        # State should have been injected
+        assert received_session[0] == "my-session-123"
+
+
+class TestInjectionContextManager:
+    """Test injection context manager."""
+    
+    def test_with_injection_context_exists(self):
+        """with_injection_context should exist."""
+        from praisonaiagents.tools.injected import with_injection_context
+        
+        assert with_injection_context is not None
+    
+    def test_context_manager_sets_state(self):
+        """Context manager should set current state."""
+        from praisonaiagents.tools.injected import (
+            with_injection_context, 
+            get_current_state,
+            AgentState
+        )
+        
+        state = AgentState(agent_id="a", run_id="r", session_id="s")
+        
+        with with_injection_context(state):
+            current = get_current_state()
+            assert current is state
+        
+        # After context, should be None
+        assert get_current_state() is None
+    
+    def test_nested_contexts(self):
+        """Nested contexts should work correctly."""
+        from praisonaiagents.tools.injected import (
+            with_injection_context,
+            get_current_state,
+            AgentState
+        )
+        
+        state1 = AgentState(agent_id="a1", run_id="r1", session_id="s1")
+        state2 = AgentState(agent_id="a2", run_id="r2", session_id="s2")
+        
+        with with_injection_context(state1):
+            assert get_current_state().agent_id == "a1"
+            
+            with with_injection_context(state2):
+                assert get_current_state().agent_id == "a2"
+            
+            # Back to state1
+            assert get_current_state().agent_id == "a1"
+
+
+class TestAsyncInjection:
+    """Test async tool injection."""
+    
+    @pytest.mark.asyncio
+    async def test_async_tool_injection(self):
+        """Async tools should also receive injected state."""
+        from praisonaiagents import tool
+        from praisonaiagents.tools import Injected
+        from praisonaiagents.tools.injected import with_injection_context, AgentState
+        
+        received_state = [None]
+        
+        @tool
+        async def async_tool(query: str, state: Injected[dict]) -> str:
+            """Async tool with injected state."""
+            received_state[0] = state
+            return "ok"
+        
+        state = AgentState(agent_id="a", run_id="r", session_id="async-session")
+        
+        with with_injection_context(state):
+            await async_tool(query="test")
+        
+        assert received_state[0] is not None
+        assert received_state[0].get('session_id') == "async-session"
+
+
+class TestInjectionZeroOverhead:
+    """Test that injection has zero overhead when not used."""
+    
+    def test_no_injected_params_no_overhead(self):
+        """Tools without Injected params should have no overhead."""
+        from praisonaiagents import tool
+        
+        @tool
+        def simple_tool(query: str) -> str:
+            """Simple tool without injection."""
+            return query
+        
+        # Should work normally without any injection context
+        result = simple_tool(query="hello")
+        assert result == "hello"
+    
+    def test_schema_generation_fast(self):
+        """Schema generation should be fast even with Injected params."""
+        from praisonaiagents import tool
+        from praisonaiagents.tools import Injected
+        import time
+        
+        @tool
+        def my_tool(a: str, b: int, c: Injected[dict], d: float) -> str:
+            """Tool with mixed params."""
+            return a
+        
+        start = time.monotonic()
+        for _ in range(100):
+            my_tool.get_schema()
+        elapsed = time.monotonic() - start
+        
+        # Should be very fast
+        assert elapsed < 0.1  # 100 calls in < 100ms
+
+
+class TestBackwardCompatibility:
+    """Test backward compatibility with existing tools."""
+    
+    def test_existing_tools_still_work(self):
+        """Existing tools without Injected should continue to work."""
+        from praisonaiagents import tool
+        
+        @tool
+        def legacy_tool(query: str, count: int = 5) -> str:
+            """Legacy tool."""
+            return f"{query} x {count}"
+        
+        result = legacy_tool(query="test", count=3)
+        assert result == "test x 3"
+    
+    def test_existing_tool_schema_unchanged(self):
+        """Existing tool schemas should be unchanged."""
+        from praisonaiagents import tool
+        
+        @tool
+        def legacy_tool(query: str, count: int = 5) -> str:
+            """Legacy tool."""
+            return f"{query} x {count}"
+        
+        schema = legacy_tool.get_schema()
+        params = schema['function']['parameters']['properties']
+        
+        assert 'query' in params
+        assert 'count' in params
+        assert len(params) == 2

src/praisonai-agents/tests/unit/test_middleware.py

@@ -0,0 +1,403 @@
+"""
+TDD Tests for Middleware System (wrap_model_call, wrap_tool_call).
+
+These tests are written FIRST before implementation.
+"""
+
+import pytest
+from unittest.mock import Mock, MagicMock, patch, AsyncMock
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+
+class TestMiddlewareTypes:
+    """Test middleware type definitions."""
+    
+    def test_invocation_context_exists(self):
+        """InvocationContext dataclass should exist with required fields."""
+        from praisonaiagents.hooks.middleware import InvocationContext
+        
+        ctx = InvocationContext(
+            agent_id="agent-123",
+            run_id="run-456",
+            session_id="session-789"
+        )
+        assert ctx.agent_id == "agent-123"
+        assert ctx.run_id == "run-456"
+        assert ctx.session_id == "session-789"
+    
+    def test_invocation_context_optional_fields(self):
+        """InvocationContext should have optional tool_name, model_name, metadata."""
+        from praisonaiagents.hooks.middleware import InvocationContext
+        
+        ctx = InvocationContext(
+            agent_id="a",
+            run_id="r",
+            session_id="s",
+            tool_name="my_tool",
+            model_name="gpt-4o",
+            metadata={"key": "value"}
+        )
+        assert ctx.tool_name == "my_tool"
+        assert ctx.model_name == "gpt-4o"
+        assert ctx.metadata == {"key": "value"}
+    
+    def test_model_request_exists(self):
+        """ModelRequest dataclass for before_model hooks."""
+        from praisonaiagents.hooks.middleware import ModelRequest
+        
+        req = ModelRequest(
+            messages=[{"role": "user", "content": "Hello"}],
+            model="gpt-4o",
+            temperature=0.7,
+            context=Mock()
+        )
+        assert req.messages == [{"role": "user", "content": "Hello"}]
+        assert req.model == "gpt-4o"
+        assert req.temperature == 0.7
+    
+    def test_model_response_exists(self):
+        """ModelResponse dataclass for after_model hooks."""
+        from praisonaiagents.hooks.middleware import ModelResponse
+        
+        resp = ModelResponse(
+            content="Hello back!",
+            model="gpt-4o",
+            usage={"prompt_tokens": 10, "completion_tokens": 5},
+            context=Mock()
+        )
+        assert resp.content == "Hello back!"
+        assert resp.model == "gpt-4o"
+    
+    def test_tool_request_exists(self):
+        """ToolRequest dataclass for before_tool hooks."""
+        from praisonaiagents.hooks.middleware import ToolRequest
+        
+        req = ToolRequest(
+            tool_name="get_weather",
+            arguments={"city": "Paris"},
+            context=Mock()
+        )
+        assert req.tool_name == "get_weather"
+        assert req.arguments == {"city": "Paris"}
+    
+    def test_tool_response_exists(self):
+        """ToolResponse dataclass for after_tool hooks."""
+        from praisonaiagents.hooks.middleware import ToolResponse
+        
+        resp = ToolResponse(
+            tool_name="get_weather",
+            result="Sunny, 22°C",
+            error=None,
+            context=Mock()
+        )
+        assert resp.result == "Sunny, 22°C"
+        assert resp.error is None
+
+
+class TestMiddlewareDecorators:
+    """Test middleware decorator functions."""
+    
+    def test_before_model_decorator_exists(self):
+        """before_model decorator should exist and tag functions."""
+        from praisonaiagents.hooks import before_model
+        
+        @before_model
+        def my_hook(request):
+            return request
+        
+        assert hasattr(my_hook, '_hook_type')
+        assert my_hook._hook_type == 'before_model'
+    
+    def test_after_model_decorator_exists(self):
+        """after_model decorator should exist and tag functions."""
+        from praisonaiagents.hooks import after_model
+        
+        @after_model
+        def my_hook(response):
+            return response
+        
+        assert hasattr(my_hook, '_hook_type')
+        assert my_hook._hook_type == 'after_model'
+    
+    def test_wrap_model_call_decorator_exists(self):
+        """wrap_model_call decorator should exist and tag functions."""
+        from praisonaiagents.hooks import wrap_model_call
+        
+        @wrap_model_call
+        def my_middleware(request, call_next):
+            return call_next(request)
+        
+        assert hasattr(my_middleware, '_hook_type')
+        assert my_middleware._hook_type == 'wrap_model_call'
+    
+    def test_before_tool_decorator_exists(self):
+        """before_tool decorator should exist and tag functions."""
+        from praisonaiagents.hooks import before_tool
+        
+        @before_tool
+        def my_hook(request):
+            return request
+        
+        assert hasattr(my_hook, '_hook_type')
+        assert my_hook._hook_type == 'before_tool'
+    
+    def test_after_tool_decorator_exists(self):
+        """after_tool decorator should exist and tag functions."""
+        from praisonaiagents.hooks import after_tool
+        
+        @after_tool
+        def my_hook(response):
+            return response
+        
+        assert hasattr(my_hook, '_hook_type')
+        assert my_hook._hook_type == 'after_tool'
+    
+    def test_wrap_tool_call_decorator_exists(self):
+        """wrap_tool_call decorator should exist and tag functions."""
+        from praisonaiagents.hooks import wrap_tool_call
+        
+        @wrap_tool_call
+        def my_middleware(request, call_next):
+            return call_next(request)
+        
+        assert hasattr(my_middleware, '_hook_type')
+        assert my_middleware._hook_type == 'wrap_tool_call'
+
+
+class TestMiddlewareChaining:
+    """Test middleware chain composition."""
+    
+    def test_wrap_model_call_chain_order(self):
+        """wrap_model_call hooks should compose in registration order."""
+        from praisonaiagents.hooks.middleware import MiddlewareChain
+        
+        calls = []
+        
+        def mw1(req, call_next):
+            calls.append("mw1_before")
+            result = call_next(req)
+            calls.append("mw1_after")
+            return result
+        
+        def mw2(req, call_next):
+            calls.append("mw2_before")
+            result = call_next(req)
+            calls.append("mw2_after")
+            return result
+        
+        def final_handler(req):
+            calls.append("final")
+            return "response"
+        
+        chain = MiddlewareChain([mw1, mw2])
+        result = chain.execute({"test": True}, final_handler)
+        
+        assert calls == ["mw1_before", "mw2_before", "final", "mw2_after", "mw1_after"]
+        assert result == "response"
+    
+    def test_empty_middleware_chain(self):
+        """Empty middleware chain should just call final handler."""
+        from praisonaiagents.hooks.middleware import MiddlewareChain
+        
+        def final_handler(req):
+            return "direct"
+        
+        chain = MiddlewareChain([])
+        result = chain.execute({"test": True}, final_handler)
+        assert result == "direct"
+    
+    def test_middleware_can_modify_request(self):
+        """Middleware should be able to modify request before passing on."""
+        from praisonaiagents.hooks.middleware import MiddlewareChain
+        
+        def add_context(req, call_next):
+            req["extra"] = "added"
+            return call_next(req)
+        
+        def final_handler(req):
+            return req.get("extra")
+        
+        chain = MiddlewareChain([add_context])
+        result = chain.execute({"original": True}, final_handler)
+        assert result == "added"
+    
+    def test_middleware_can_modify_response(self):
+        """Middleware should be able to modify response after call_next."""
+        from praisonaiagents.hooks.middleware import MiddlewareChain
+        
+        def uppercase_response(req, call_next):
+            result = call_next(req)
+            return result.upper()
+        
+        def final_handler(req):
+            return "hello"
+        
+        chain = MiddlewareChain([uppercase_response])
+        result = chain.execute({}, final_handler)
+        assert result == "HELLO"
+    
+    def test_middleware_can_short_circuit(self):
+        """Middleware can return early without calling call_next."""
+        from praisonaiagents.hooks.middleware import MiddlewareChain
+        
+        def block_middleware(req, call_next):
+            if req.get("block"):
+                return "blocked"
+            return call_next(req)
+        
+        def final_handler(req):
+            return "reached"
+        
+        chain = MiddlewareChain([block_middleware])
+        
+        # Should be blocked
+        result = chain.execute({"block": True}, final_handler)
+        assert result == "blocked"
+        
+        # Should pass through
+        result = chain.execute({"block": False}, final_handler)
+        assert result == "reached"
+
+
+class TestMiddlewareRetry:
+    """Test retry logic in middleware."""
+    
+    def test_wrap_tool_call_retry_on_error(self):
+        """wrap_tool_call should support retry patterns."""
+        from praisonaiagents.hooks.middleware import MiddlewareChain
+        
+        attempt_count = [0]
+        
+        def retry_middleware(req, call_next):
+            last_error = None
+            for i in range(3):
+                try:
+                    return call_next(req)
+                except Exception as e:
+                    last_error = e
+            raise last_error
+        
+        def flaky_handler(req):
+            attempt_count[0] += 1
+            if attempt_count[0] < 3:
+                raise RuntimeError("flaky")
+            return "success"
+        
+        chain = MiddlewareChain([retry_middleware])
+        result = chain.execute({}, flaky_handler)
+        assert result == "success"
+        assert attempt_count[0] == 3
+
+
+class TestAsyncMiddleware:
+    """Test async middleware support."""
+    
+    @pytest.mark.asyncio
+    async def test_async_wrap_model_call(self):
+        """Async middleware should work with async handlers."""
+        from praisonaiagents.hooks.middleware import AsyncMiddlewareChain
+        
+        calls = []
+        
+        async def async_mw(req, call_next):
+            calls.append("before")
+            result = await call_next(req)
+            calls.append("after")
+            return result
+        
+        async def async_handler(req):
+            calls.append("handler")
+            return "async_result"
+        
+        chain = AsyncMiddlewareChain([async_mw])
+        result = await chain.execute({}, async_handler)
+        
+        assert calls == ["before", "handler", "after"]
+        assert result == "async_result"
+    
+    @pytest.mark.asyncio
+    async def test_async_middleware_chain_order(self):
+        """Async middleware should maintain order."""
+        from praisonaiagents.hooks.middleware import AsyncMiddlewareChain
+        
+        order = []
+        
+        async def mw1(req, call_next):
+            order.append(1)
+            result = await call_next(req)
+            order.append(-1)
+            return result
+        
+        async def mw2(req, call_next):
+            order.append(2)
+            result = await call_next(req)
+            order.append(-2)
+            return result
+        
+        async def handler(req):
+            order.append(0)
+            return "done"
+        
+        chain = AsyncMiddlewareChain([mw1, mw2])
+        await chain.execute({}, handler)
+        
+        assert order == [1, 2, 0, -2, -1]
+
+
+class TestAgentMiddlewareIntegration:
+    """Test middleware integration with Agent class."""
+    
+    def test_agent_accepts_hooks_parameter(self):
+        """Agent should accept hooks parameter."""
+        from praisonaiagents import Agent
+        from praisonaiagents.hooks import before_model
+        
+        @before_model
+        def my_hook(req):
+            return req
+        
+        # Should not raise
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            hooks=[my_hook]
+        )
+        assert agent is not None
+    
+    def test_agent_hooks_empty_by_default(self):
+        """Agent should have empty hooks by default."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(name="Test", instructions="Test")
+        # Should have no hooks or empty hooks list
+        hooks = getattr(agent, '_hooks', None) or getattr(agent, 'hooks', [])
+        assert hooks is None or len(hooks) == 0
+
+
+class TestMiddlewareZeroOverhead:
+    """Test that middleware has zero overhead when not used."""
+    
+    def test_no_hooks_fast_path(self):
+        """When no hooks registered, should take fast path."""
+        from praisonaiagents.hooks.middleware import MiddlewareChain
+        
+        # Empty chain should be essentially a no-op
+        chain = MiddlewareChain([])
+        
+        call_count = [0]
+        def handler(req):
+            call_count[0] += 1
+            return "result"
+        
+        result = chain.execute({}, handler)
+        assert result == "result"
+        assert call_count[0] == 1
+    
+    def test_middleware_chain_is_none_check(self):
+        """Should handle None middleware list gracefully."""
+        from praisonaiagents.hooks.middleware import MiddlewareChain
+        
+        chain = MiddlewareChain(None)
+        result = chain.execute({}, lambda r: "ok")
+        assert result == "ok"

src/praisonai-agents/tests/unit/test_rate_limiter.py

@@ -0,0 +1,260 @@
+"""
+TDD Tests for Rate Limiter.
+
+These tests are written FIRST before implementation.
+"""
+
+import pytest
+import time
+from unittest.mock import Mock, patch
+
+
+class TestRateLimiterBasic:
+    """Test basic rate limiter functionality."""
+    
+    def test_rate_limiter_exists(self):
+        """RateLimiter class should exist in praisonaiagents.llm."""
+        from praisonaiagents.llm import RateLimiter
+        
+        limiter = RateLimiter(requests_per_minute=60)
+        assert limiter is not None
+    
+    def test_rate_limiter_rpm_parameter(self):
+        """RateLimiter should accept requests_per_minute."""
+        from praisonaiagents.llm import RateLimiter
+        
+        limiter = RateLimiter(requests_per_minute=60)
+        assert limiter.requests_per_minute == 60
+    
+    def test_rate_limiter_acquire_sync(self):
+        """RateLimiter.acquire() should block if rate exceeded."""
+        from praisonaiagents.llm import RateLimiter
+        
+        # 1 request per second = 60 per minute
+        limiter = RateLimiter(requests_per_minute=60)
+        
+        # First acquire should be instant
+        start = time.monotonic()
+        limiter.acquire()
+        elapsed = time.monotonic() - start
+        assert elapsed < 0.1  # Should be nearly instant
+    
+    def test_rate_limiter_blocks_when_exceeded(self):
+        """RateLimiter should block when rate is exceeded."""
+        from praisonaiagents.llm import RateLimiter
+        
+        # Very low rate: 6 per minute = 1 per 10 seconds
+        # But we'll use a mock clock to test without waiting
+        limiter = RateLimiter(requests_per_minute=6)
+        
+        # Inject a mock clock for testing
+        mock_time = [0.0]
+        limiter._get_time = lambda: mock_time[0]
+        limiter._sleep = lambda s: mock_time.__setitem__(0, mock_time[0] + s)
+        
+        # First request should pass
+        limiter.acquire()
+        
+        # Second request should need to wait ~10 seconds
+        # (but our mock sleep just advances time)
+        limiter.acquire()
+        
+        # Time should have advanced
+        assert mock_time[0] >= 10.0
+    
+    def test_rate_limiter_token_bucket(self):
+        """RateLimiter should use token bucket algorithm."""
+        from praisonaiagents.llm import RateLimiter
+        
+        limiter = RateLimiter(requests_per_minute=60, burst=5)
+        
+        # Reset internal state with mock time
+        mock_time = [0.0]
+        limiter._get_time = lambda: mock_time[0]
+        limiter._sleep = lambda s: mock_time.__setitem__(0, mock_time[0] + s)
+        limiter._tokens = 5.0  # Reset tokens to burst size
+        limiter._last_update = 0.0  # Reset last update
+        
+        for _ in range(5):
+            limiter.acquire()
+        
+        # Should not have waited much (burst allows 5 instant requests)
+        assert mock_time[0] < 1.0
+
+
+class TestRateLimiterAsync:
+    """Test async rate limiter functionality."""
+    
+    @pytest.mark.asyncio
+    async def test_rate_limiter_acquire_async(self):
+        """RateLimiter.acquire_async() should work with async."""
+        from praisonaiagents.llm import RateLimiter
+        
+        limiter = RateLimiter(requests_per_minute=60)
+        
+        # Should have async acquire method
+        await limiter.acquire_async()
+    
+    @pytest.mark.asyncio
+    async def test_rate_limiter_async_blocks(self):
+        """Async acquire should also respect rate limits."""
+        from praisonaiagents.llm import RateLimiter
+        import asyncio
+        
+        limiter = RateLimiter(requests_per_minute=6)
+        
+        # Mock async sleep
+        sleep_times = []
+        original_sleep = asyncio.sleep
+        
+        async def mock_sleep(s):
+            sleep_times.append(s)
+            # Don't actually sleep in tests
+        
+        with patch('asyncio.sleep', mock_sleep):
+            await limiter.acquire_async()
+            await limiter.acquire_async()
+        
+        # Should have tried to sleep
+
+
+class TestRateLimiterAgentIntegration:
+    """Test rate limiter integration with Agent."""
+    
+    def test_agent_accepts_rate_limiter(self):
+        """Agent should accept rate_limiter parameter."""
+        from praisonaiagents import Agent
+        from praisonaiagents.llm import RateLimiter
+        
+        limiter = RateLimiter(requests_per_minute=60)
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            rate_limiter=limiter
+        )
+        assert agent is not None
+        assert getattr(agent, '_rate_limiter', None) is limiter or \
+               getattr(agent, 'rate_limiter', None) is limiter
+    
+    def test_agent_rate_limiter_applied_to_llm_calls(self):
+        """Rate limiter should be applied before LLM calls."""
+        from praisonaiagents import Agent
+        from praisonaiagents.llm import RateLimiter
+        
+        limiter = RateLimiter(requests_per_minute=60)
+        acquire_count = [0]
+        original_acquire = limiter.acquire
+        
+        def counting_acquire():
+            acquire_count[0] += 1
+            return original_acquire()
+        
+        limiter.acquire = counting_acquire
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent",
+            rate_limiter=limiter
+        )
+        
+        with patch.object(agent, '_chat_completion', return_value="mocked"):
+            agent.chat("Hello")
+        
+        # Rate limiter should have been called
+        assert acquire_count[0] >= 1
+
+
+class TestRateLimiterZeroOverhead:
+    """Test that rate limiter has zero overhead when not used."""
+    
+    def test_no_rate_limiter_no_overhead(self):
+        """When no rate limiter, should not add any overhead."""
+        from praisonaiagents import Agent
+        
+        agent = Agent(
+            name="Test",
+            instructions="Test agent"
+            # No rate_limiter
+        )
+        
+        # Should not have rate limiter
+        assert getattr(agent, '_rate_limiter', None) is None
+    
+    def test_rate_limiter_none_check_fast(self):
+        """Checking for None rate limiter should be O(1)."""
+        from praisonaiagents.llm import RateLimiter
+        
+        # This is more of a design requirement than a test
+        # The implementation should do: if self._rate_limiter: self._rate_limiter.acquire()
+
+
+class TestRateLimiterConfiguration:
+    """Test rate limiter configuration options."""
+    
+    def test_rate_limiter_default_burst(self):
+        """RateLimiter should have sensible default burst."""
+        from praisonaiagents.llm import RateLimiter
+        
+        limiter = RateLimiter(requests_per_minute=60)
+        # Default burst should be 1 or based on rpm
+        assert hasattr(limiter, 'burst') or hasattr(limiter, '_burst')
+    
+    def test_rate_limiter_custom_burst(self):
+        """RateLimiter should accept custom burst parameter."""
+        from praisonaiagents.llm import RateLimiter
+        
+        limiter = RateLimiter(requests_per_minute=60, burst=10)
+        burst = getattr(limiter, 'burst', None) or getattr(limiter, '_burst', None)
+        assert burst == 10
+    
+    def test_rate_limiter_tpm_optional(self):
+        """RateLimiter should optionally accept tokens_per_minute."""
+        from praisonaiagents.llm import RateLimiter
+        
+        # Should not raise
+        limiter = RateLimiter(
+            requests_per_minute=60,
+            tokens_per_minute=100000
+        )
+        assert limiter is not None
+
+
+class TestRateLimiterDeterministic:
+    """Test that rate limiter is deterministic for testing."""
+    
+    def test_injectable_clock(self):
+        """RateLimiter should allow injecting clock for testing."""
+        from praisonaiagents.llm import RateLimiter
+        
+        mock_time = [0.0]
+        
+        limiter = RateLimiter(requests_per_minute=60)
+        limiter._get_time = lambda: mock_time[0]
+        
+        # Should use our mock time
+        t1 = limiter._get_time()
+        mock_time[0] = 100.0
+        t2 = limiter._get_time()
+        
+        assert t1 == 0.0
+        assert t2 == 100.0
+    
+    def test_injectable_sleep(self):
+        """RateLimiter should allow injecting sleep for testing."""
+        from praisonaiagents.llm import RateLimiter
+        
+        sleep_calls = []
+        
+        limiter = RateLimiter(requests_per_minute=60)
+        limiter._sleep = lambda s: sleep_calls.append(s)
+        
+        # Force a wait scenario
+        limiter._get_time = lambda: 0.0
+        limiter._tokens = 0  # No tokens available
+        limiter._last_update = 0.0
+        
+        limiter.acquire()
+        
+        # Should have called our mock sleep
+        # (actual behavior depends on implementation)

src/praisonai-agents/uv.lock

@@ -2422,7 +2422,7 @@ wheels = [
 
 [[package]]
 name = "praisonaiagents"
-version = "0.2.0"
+version = "0.3.4"
 source = { editable = "." }
 dependencies = [
     { name = "aiohttp" },

src/praisonai-ts/FEATURE_PARITY_TRACKER.json

@@ -0,0 +1,206 @@
+{
+  "version": "1.0.0",
+  "lastUpdated": "2024-12-26",
+  "implementationStatus": "Phase 3 Complete - Full Documentation",
+  "summary": {
+    "pythonCoreFeatures": 85,
+    "pythonWrapperFeatures": 45,
+    "typescriptFeatures": 48,
+    "gapCount": 25,
+    "priorityP0": 12,
+    "priorityP1": 15,
+    "priorityP2": 20,
+    "priorityP3": 20
+  },
+  "pythonCoreSDK": {
+    "path": "/Users/praison/praisonai-package/src/praisonai-agents/praisonaiagents",
+    "exports": [
+      "Agent", "ImageAgent", "ContextAgent", "DeepResearchAgent", "QueryRewriterAgent", "PromptExpanderAgent",
+      "PraisonAIAgents", "Agents", "Task", "Tools", "BaseTool", "ToolResult", "tool", "FunctionTool",
+      "ToolRegistry", "get_registry", "register_tool", "get_tool", "AutoAgents", "Session", "Memory",
+      "Knowledge", "Chunking", "db", "obs", "MCP", "Workflow", "Pipeline", "WorkflowStep", "WorkflowContext",
+      "StepResult", "Route", "Parallel", "Loop", "Repeat", "GuardrailResult", "LLMGuardrail",
+      "Handoff", "handoff", "FlowDisplay", "FastContext", "SkillManager", "SkillProperties",
+      "get_telemetry", "enable_telemetry", "MinimalTelemetry", "TelemetryCollector", "AGUI", "A2A"
+    ],
+    "modules": {
+      "agent": ["Agent", "ImageAgent", "ContextAgent", "DeepResearchAgent", "QueryRewriterAgent", "PromptExpanderAgent", "RouterAgent", "Handoff"],
+      "agents": ["PraisonAIAgents", "AutoAgents"],
+      "task": ["Task"],
+      "tools": ["Tools", "BaseTool", "ToolResult", "tool", "FunctionTool", "ToolRegistry", "40+ built-in tools"],
+      "llm": ["LLM", "OpenAIClient", "ModelRouter", "RateLimiter", "ModelCapabilities"],
+      "memory": ["Memory", "AutoMemory", "FileMemory", "RulesManager", "DocsManager", "Hooks"],
+      "knowledge": ["Knowledge", "Chunking", "VectorStore", "QueryEngine", "Readers", "Rerankers"],
+      "db": ["DbAdapter", "AsyncDbAdapter", "DbMessage", "DbToolCall", "DbRun", "DbSpan", "DbTrace"],
+      "obs": ["ObservabilityAdapter", "Span", "Trace", "16 provider integrations"],
+      "mcp": ["MCP", "MCPServer", "MCPSession", "MCPTransport", "MCPWebSocket", "MCPSSE", "MCPHTTPStream"],
+      "workflows": ["Workflow", "Pipeline", "WorkflowStep", "WorkflowContext", "StepResult", "Route", "Parallel", "Loop", "Repeat", "YAMLWorkflowParser"],
+      "guardrails": ["GuardrailResult", "LLMGuardrail"],
+      "eval": ["AccuracyEval", "PerformanceEval", "ReliabilityEval", "CriteriaEval"],
+      "skills": ["SkillManager", "SkillProperties", "SkillLoader", "SkillDiscovery", "SkillValidator"],
+      "telemetry": ["MinimalTelemetry", "TelemetryCollector", "PerformanceMonitor", "TokenCollector"],
+      "context": ["FastContext", "FastContextResult"],
+      "session": ["Session"],
+      "ui": ["AGUI", "A2A"]
+    }
+  },
+  "pythonWrapper": {
+    "path": "/Users/praison/praisonai-package/src/praisonai/praisonai",
+    "exports": ["PraisonAI"],
+    "cliFeatures": [
+      "slash_commands", "autonomy_mode", "cost_tracker", "repo_map", "interactive_tui",
+      "git_integration", "sandbox_executor", "agent_scheduler", "agents", "at_mentions",
+      "auto_memory", "background", "checkpoints", "compare", "eval", "external_agents",
+      "fast_context", "flow_display", "guardrail", "handoff", "hooks", "image", "jobs",
+      "knowledge", "mcp", "message_queue", "metrics", "n8n", "observability", "persistence",
+      "profile", "profiling", "router", "session", "skills", "telemetry", "todo", "tools", "workflow"
+    ]
+  },
+  "typescriptSDK": {
+    "path": "/Users/praison/praisonai-package/src/praisonai-ts/src",
+    "exports": ["Agent", "PraisonAIAgents", "Task", "Tool", "BaseTool", "Memory", "BaseMemoryStore", "Process", "BaseProcess", "OpenAIService", "Logger"],
+    "modules": {
+      "agent": ["Agent (simple)", "Agent (types/proxy)", "PraisonAIAgents"],
+      "task": ["Task"],
+      "tools": ["Tool interface", "BaseTool", "arxivTools", "mcpSse"],
+      "llm": ["OpenAIService (OpenAI only)"],
+      "memory": ["Memory interface", "BaseMemoryStore (in-memory only)"],
+      "process": ["Process interface", "BaseProcess"],
+      "utils": ["Logger"]
+    }
+  },
+  "gapMatrix": {
+    "P0_CoreParity": [
+      {"feature": "Multi-provider LLM support (LiteLLM equivalent)", "python": true, "typescript": false, "priority": "P0", "effort": "high"},
+      {"feature": "Streaming responses", "python": true, "typescript": "partial", "priority": "P0", "effort": "medium"},
+      {"feature": "Structured outputs (JSON schema)", "python": true, "typescript": false, "priority": "P0", "effort": "medium"},
+      {"feature": "Tool/function calling with validation", "python": true, "typescript": "partial", "priority": "P0", "effort": "medium"},
+      {"feature": "Session management (session_id, run_id)", "python": true, "typescript": false, "priority": "P0", "effort": "medium"},
+      {"feature": "Multi-agent orchestration patterns", "python": true, "typescript": "partial", "priority": "P0", "effort": "high"},
+      {"feature": "Agent handoff", "python": true, "typescript": false, "priority": "P0", "effort": "medium"},
+      {"feature": "Tool decorator (@tool)", "python": true, "typescript": false, "priority": "P0", "effort": "low"},
+      {"feature": "Tool registry", "python": true, "typescript": false, "priority": "P0", "effort": "medium"},
+      {"feature": "Async agent execution", "python": true, "typescript": "partial", "priority": "P0", "effort": "low"},
+      {"feature": "Error handling and retries", "python": true, "typescript": "partial", "priority": "P0", "effort": "medium"},
+      {"feature": "Verbose/debug logging", "python": true, "typescript": "partial", "priority": "P0", "effort": "low"}
+    ],
+    "P1_Persistence": [
+      {"feature": "Database adapter protocol", "python": true, "typescript": false, "priority": "P1", "effort": "high"},
+      {"feature": "SQLite adapter", "python": true, "typescript": false, "priority": "P1", "effort": "medium"},
+      {"feature": "PostgreSQL adapter", "python": true, "typescript": false, "priority": "P1", "effort": "medium"},
+      {"feature": "Redis adapter", "python": true, "typescript": false, "priority": "P1", "effort": "medium"},
+      {"feature": "Message persistence", "python": true, "typescript": false, "priority": "P1", "effort": "medium"},
+      {"feature": "Run/trace persistence", "python": true, "typescript": false, "priority": "P1", "effort": "medium"},
+      {"feature": "Memory persistence (vector)", "python": true, "typescript": false, "priority": "P1", "effort": "high"},
+      {"feature": "Knowledge base (RAG)", "python": true, "typescript": false, "priority": "P1", "effort": "high"},
+      {"feature": "Chunking strategies", "python": true, "typescript": false, "priority": "P1", "effort": "medium"},
+      {"feature": "Vector store integration", "python": true, "typescript": false, "priority": "P1", "effort": "high"},
+      {"feature": "Semantic search/recall", "python": true, "typescript": false, "priority": "P1", "effort": "high"},
+      {"feature": "Working memory", "python": true, "typescript": false, "priority": "P1", "effort": "medium"},
+      {"feature": "File-based memory", "python": true, "typescript": false, "priority": "P1", "effort": "medium"},
+      {"feature": "Auto-memory", "python": true, "typescript": false, "priority": "P1", "effort": "medium"},
+      {"feature": "Session export/import", "python": true, "typescript": false, "priority": "P1", "effort": "medium"}
+    ],
+    "P2_CLI": [
+      {"feature": "CLI entrypoint", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Slash commands", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Interactive TUI", "python": true, "typescript": false, "priority": "P2", "effort": "high"},
+      {"feature": "Cost tracking", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Repo map", "python": true, "typescript": false, "priority": "P2", "effort": "high"},
+      {"feature": "Git integration", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Sandbox executor", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Autonomy modes", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Agent scheduler", "python": true, "typescript": false, "priority": "P2", "effort": "high"},
+      {"feature": "Background jobs", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Checkpoints", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Flow display", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Profiling", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Metrics", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "External agent integrations", "python": true, "typescript": false, "priority": "P2", "effort": "high"},
+      {"feature": "N8N integration", "python": true, "typescript": false, "priority": "P2", "effort": "high"},
+      {"feature": "MCP CLI support", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Skills CLI", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Eval CLI", "python": true, "typescript": false, "priority": "P2", "effort": "medium"},
+      {"feature": "Workflow CLI", "python": true, "typescript": false, "priority": "P2", "effort": "medium"}
+    ],
+    "P3_Advanced": [
+      {"feature": "Workflows/Pipelines", "python": true, "typescript": false, "priority": "P3", "effort": "high"},
+      {"feature": "Route pattern", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Parallel pattern", "python": true, "typescript": "partial", "priority": "P3", "effort": "low"},
+      {"feature": "Loop pattern", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Repeat pattern", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "YAML workflow parser", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Guardrails", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "LLM Guardrail", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Evaluation framework", "python": true, "typescript": false, "priority": "P3", "effort": "high"},
+      {"feature": "Accuracy eval", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Performance eval", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Reliability eval", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Criteria eval", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Observability protocol", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Telemetry", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Performance monitoring", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Token tracking", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Skills system", "python": true, "typescript": false, "priority": "P3", "effort": "high"},
+      {"feature": "Fast context", "python": true, "typescript": false, "priority": "P3", "effort": "medium"},
+      {"feature": "Image agent", "python": true, "typescript": false, "priority": "P3", "effort": "medium"}
+    ]
+  },
+  "mastraCapabilities": {
+    "implemented": [
+      "Agent core", "Tool calling", "Streaming (partial)", "Multi-agent (partial)"
+    ],
+    "missing": [
+      "Memory with semantic recall", "Working memory", "Thread management", "Resource scoping",
+      "Storage abstraction", "Vector store integration", "Embeddings", "Evaluation scorers",
+      "Workflow engine", "Step execution", "Branching logic", "Event-driven workflows",
+      "MCP integration", "Voice/TTS", "Observability hooks", "Request context",
+      "Tool validation", "Tool suspend/resume", "Provider options per tool"
+    ]
+  },
+  "implementationPlan": {
+    "phase1_P0": {
+      "description": "Core agent parity - streaming, structured outputs, sessions, tool system",
+      "tasks": [
+        "Implement AI SDK provider abstraction (OpenAI, Anthropic, Google)",
+        "Add streaming response support with proper typing",
+        "Implement structured output with JSON schema/Zod",
+        "Add session management (session_id, run_id, trace_id)",
+        "Implement tool decorator and registry",
+        "Add agent handoff support",
+        "Improve error handling and retries"
+      ]
+    },
+    "phase2_P1": {
+      "description": "Persistence layer - db adapters, memory, knowledge",
+      "tasks": [
+        "Implement DbAdapter protocol",
+        "Add SQLite adapter",
+        "Add message/run/trace persistence",
+        "Implement Memory with persistence",
+        "Add Knowledge base (RAG) support",
+        "Implement vector store integration"
+      ]
+    },
+    "phase3_P2": {
+      "description": "CLI parity",
+      "tasks": [
+        "Create CLI entrypoint",
+        "Implement slash commands",
+        "Add cost tracking",
+        "Implement interactive mode"
+      ]
+    },
+    "phase4_P3": {
+      "description": "Advanced features",
+      "tasks": [
+        "Implement Workflow engine",
+        "Add workflow patterns (Route, Parallel, Loop, Repeat)",
+        "Implement Guardrails",
+        "Add Evaluation framework",
+        "Implement Observability hooks",
+        "Add Skills system"
+      ]
+    }
+  }
+}

src/praisonai-ts/jest.config.ts

@@ -4,7 +4,7 @@ const config: Config.InitialOptions = {
   preset: 'ts-jest',
   testEnvironment: 'node',
   roots: ['<rootDir>/tests'],
-  testMatch: ['**/*.main.test.ts', '**/*.spec.ts'],
+  testMatch: ['**/*.test.ts', '**/*.spec.ts'],
   moduleFileExtensions: ['ts', 'js', 'json'],
   coverageDirectory: 'coverage',
   collectCoverageFrom: [

src/praisonai-ts/package.json

@@ -1,9 +1,12 @@
 {
   "name": "praisonai",
-  "version": "1.0.19",
+  "version": "1.1.0",
   "description": "PraisonAI TypeScript AI Agents Framework - Node.js, npm, and Javascript AI Agents Framework",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
+  "bin": {
+    "praisonai-ts": "./dist/cli/index.js"
+  },
   "scripts": {
     "build": "tsc",
     "test": "jest",
@@ -62,7 +65,6 @@
     "fast-xml-parser": "^4.5.1",
     "node-fetch": "^2.6.9",
     "openai": "^4.81.0",
-    "praisonai": "^1.0.19",
     "@modelcontextprotocol/sdk": "^1.12.1"
   },
   "optionalDependencies": {

src/praisonai-ts/src/agent/context.ts

@@ -0,0 +1,159 @@
+/**
+ * Context Agent - Agent with enhanced context management
+ */
+
+import { createProvider, type LLMProvider, type Message } from '../llm/providers';
+import { Session } from '../session';
+import { KnowledgeBase, type SearchResult } from '../knowledge/rag';
+
+export interface ContextAgentConfig {
+  name?: string;
+  instructions: string;
+  llm?: string;
+  knowledgeBase?: KnowledgeBase;
+  contextWindow?: number;
+  maxContextTokens?: number;
+  verbose?: boolean;
+}
+
+export interface ContextMessage {
+  role: 'system' | 'user' | 'assistant';
+  content: string;
+  timestamp: number;
+  metadata?: Record<string, any>;
+}
+
+/**
+ * Context Agent - Agent with RAG and context management
+ */
+export class ContextAgent {
+  readonly name: string;
+  readonly instructions: string;
+  
+  private provider: LLMProvider;
+  private knowledgeBase?: KnowledgeBase;
+  private contextWindow: number;
+  private maxContextTokens: number;
+  private messages: ContextMessage[] = [];
+  private verbose: boolean;
+
+  constructor(config: ContextAgentConfig) {
+    this.name = config.name || `ContextAgent_${Math.random().toString(36).substr(2, 9)}`;
+    this.instructions = config.instructions;
+    this.provider = createProvider(config.llm || 'gpt-4o-mini');
+    this.knowledgeBase = config.knowledgeBase;
+    this.contextWindow = config.contextWindow ?? 10;
+    this.maxContextTokens = config.maxContextTokens ?? 4000;
+    this.verbose = config.verbose ?? false;
+  }
+
+  /**
+   * Chat with context awareness
+   */
+  async chat(prompt: string): Promise<{ text: string; context?: SearchResult[] }> {
+    // Add user message
+    this.messages.push({
+      role: 'user',
+      content: prompt,
+      timestamp: Date.now()
+    });
+
+    // Build context from knowledge base
+    let ragContext: SearchResult[] = [];
+    let contextString = '';
+    
+    if (this.knowledgeBase) {
+      ragContext = await this.knowledgeBase.search(prompt);
+      if (ragContext.length > 0) {
+        contextString = this.knowledgeBase.buildContext(ragContext);
+        if (this.verbose) {
+          console.log(`[ContextAgent] Found ${ragContext.length} relevant documents`);
+        }
+      }
+    }
+
+    // Build messages array with context
+    const systemPrompt = contextString
+      ? `${this.instructions}\n\nRelevant context:\n${contextString}`
+      : this.instructions;
+
+    const messages: Message[] = [
+      { role: 'system', content: systemPrompt },
+      ...this.getRecentMessages()
+    ];
+
+    // Generate response
+    const result = await this.provider.generateText({ messages });
+
+    // Add assistant message
+    this.messages.push({
+      role: 'assistant',
+      content: result.text,
+      timestamp: Date.now()
+    });
+
+    return {
+      text: result.text,
+      context: ragContext.length > 0 ? ragContext : undefined
+    };
+  }
+
+  /**
+   * Get recent messages within context window
+   */
+  private getRecentMessages(): Message[] {
+    const recent = this.messages.slice(-this.contextWindow);
+    return recent.map(m => ({
+      role: m.role,
+      content: m.content
+    }));
+  }
+
+  /**
+   * Add document to knowledge base
+   */
+  async addDocument(id: string, content: string, metadata?: Record<string, any>): Promise<void> {
+    if (!this.knowledgeBase) {
+      throw new Error('No knowledge base configured');
+    }
+    await this.knowledgeBase.add({ id, content, metadata });
+  }
+
+  /**
+   * Search knowledge base
+   */
+  async searchKnowledge(query: string, limit?: number): Promise<SearchResult[]> {
+    if (!this.knowledgeBase) {
+      return [];
+    }
+    return this.knowledgeBase.search(query, limit);
+  }
+
+  /**
+   * Clear conversation history
+   */
+  clearHistory(): void {
+    this.messages = [];
+  }
+
+  /**
+   * Get conversation history
+   */
+  getHistory(): ContextMessage[] {
+    return [...this.messages];
+  }
+
+  /**
+   * Set knowledge base
+   */
+  setKnowledgeBase(kb: KnowledgeBase): void {
+    this.knowledgeBase = kb;
+  }
+}
+
+/**
+ * Create a context agent
+ */
+export function createContextAgent(config: ContextAgentConfig): ContextAgent {
+  return new ContextAgent(config);
+}

src/praisonai-ts/src/agent/enhanced.ts

@@ -0,0 +1,337 @@
+/**
+ * Enhanced Agent - Agent with session management, provider abstraction, and tool support
+ */
+
+import { createProvider, type LLMProvider, type Message, type ToolCall, type GenerateTextResult } from '../llm/providers';
+import { Session, Run, Trace, getSessionManager } from '../session';
+import { FunctionTool, ToolRegistry, tool } from '../tools/decorator';
+import { Logger } from '../utils/logger';
+
+export interface EnhancedAgentConfig {
+  name?: string;
+  instructions: string;
+  llm?: string;
+  tools?: Array<FunctionTool | Function | { name: string; description?: string; parameters?: any; execute: Function }>;
+  session?: Session;
+  sessionId?: string;
+  verbose?: boolean;
+  stream?: boolean;
+  maxToolCalls?: number;
+  temperature?: number;
+  maxTokens?: number;
+  outputSchema?: any;
+}
+
+export interface ChatOptions {
+  stream?: boolean;
+  temperature?: number;
+  maxTokens?: number;
+  outputSchema?: any;
+  onToken?: (token: string) => void;
+}
+
+export interface ChatResult {
+  text: string;
+  structured?: any;
+  toolCalls?: ToolCall[];
+  usage: {
+    promptTokens: number;
+    completionTokens: number;
+    totalTokens: number;
+  };
+  runId: string;
+  sessionId: string;
+}
+
+/**
+ * Enhanced Agent with full feature support
+ */
+export class EnhancedAgent {
+  readonly name: string;
+  readonly instructions: string;
+  readonly sessionId: string;
+  
+  private provider: LLMProvider;
+  private session: Session;
+  private toolRegistry: ToolRegistry;
+  private verbose: boolean;
+  private stream: boolean;
+  private maxToolCalls: number;
+  private temperature: number;
+  private maxTokens?: number;
+  private outputSchema?: any;
+
+  constructor(config: EnhancedAgentConfig) {
+    this.name = config.name || `Agent_${Math.random().toString(36).substr(2, 9)}`;
+    this.instructions = config.instructions;
+    this.verbose = config.verbose ?? true;
+    this.stream = config.stream ?? true;
+    this.maxToolCalls = config.maxToolCalls ?? 10;
+    this.temperature = config.temperature ?? 0.7;
+    this.maxTokens = config.maxTokens;
+    this.outputSchema = config.outputSchema;
+
+    // Initialize provider
+    this.provider = createProvider(config.llm || 'gpt-4o-mini');
+
+    // Initialize session
+    if (config.session) {
+      this.session = config.session;
+    } else if (config.sessionId) {
+      this.session = getSessionManager().getOrCreate(config.sessionId);
+    } else {
+      this.session = getSessionManager().create();
+    }
+    this.sessionId = this.session.id;
+
+    // Initialize tool registry
+    this.toolRegistry = new ToolRegistry();
+    if (config.tools) {
+      this.registerTools(config.tools);
+    }
+
+    Logger.setVerbose(this.verbose);
+  }
+
+  private registerTools(tools: EnhancedAgentConfig['tools']): void {
+    if (!tools) return;
+
+    for (const t of tools) {
+      if (t instanceof FunctionTool) {
+        this.toolRegistry.register(t);
+      } else if (typeof t === 'function') {
+        // Convert function to tool
+        const funcName = t.name || `function_${Math.random().toString(36).substr(2, 9)}`;
+        this.toolRegistry.register(tool({
+          name: funcName,
+          description: `Function ${funcName}`,
+          execute: t as any,
+        }));
+      } else if (typeof t === 'object' && 'execute' in t) {
+        // Tool config object
+        this.toolRegistry.register(tool({
+          name: t.name,
+          description: t.description,
+          parameters: t.parameters,
+          execute: t.execute as any,
+        }));
+      }
+    }
+  }
+
+  /**
+   * Add a tool to the agent
+   */
+  addTool(t: FunctionTool | { name: string; description?: string; parameters?: any; execute: Function }): this {
+    if (t instanceof FunctionTool) {
+      this.toolRegistry.register(t);
+    } else {
+      this.toolRegistry.register(tool({
+        name: t.name,
+        description: t.description,
+        parameters: t.parameters,
+        execute: t.execute as any,
+      }));
+    }
+    return this;
+  }
+
+  /**
+   * Get all registered tools
+   */
+  getTools(): FunctionTool[] {
+    return this.toolRegistry.list();
+  }
+
+  /**
+   * Chat with the agent
+   */
+  async chat(prompt: string, options: ChatOptions = {}): Promise<ChatResult> {
+    const run = this.session.createRun().start();
+    const trace = run.createTrace({ name: 'chat' }).start();
+
+    try {
+      // Add user message to session
+      this.session.addMessage({ role: 'user', content: prompt });
+
+      // Build messages array
+      const messages: Message[] = [
+        { role: 'system', content: this.instructions },
+        ...this.session.getMessagesForLLM() as Message[],
+      ];
+
+      // Get tool definitions
+      const tools = this.toolRegistry.list().length > 0
+        ? this.toolRegistry.getDefinitions()
+        : undefined;
+
+      let result: GenerateTextResult;
+      let fullText = '';
+      let allToolCalls: ToolCall[] = [];
+      let iterations = 0;
+
+      // Tool calling loop
+      while (iterations < this.maxToolCalls) {
+        iterations++;
+
+        if (options.stream && !tools && options.onToken) {
+          // Streaming without tools
+          const stream = await this.provider.streamText({
+            messages,
+            temperature: options.temperature ?? this.temperature,
+            maxTokens: options.maxTokens ?? this.maxTokens,
+            onToken: options.onToken,
+          });
+
+          let usage = { promptTokens: 0, completionTokens: 0, totalTokens: 0 };
+          for await (const chunk of stream) {
+            if (chunk.text) fullText += chunk.text;
+            if (chunk.usage) usage = { ...usage, ...chunk.usage };
+          }
+
+          result = {
+            text: fullText,
+            usage,
+            finishReason: 'stop',
+          };
+          break;
+        } else if (options.outputSchema) {
+          // Structured output
+          const objResult = await this.provider.generateObject({
+            messages,
+            schema: options.outputSchema ?? this.outputSchema,
+            temperature: options.temperature ?? this.temperature,
+            maxTokens: options.maxTokens ?? this.maxTokens,
+          });
+
+          result = {
+            text: JSON.stringify(objResult.object),
+            usage: objResult.usage,
+            finishReason: 'stop',
+          };
+          fullText = result.text;
+          break;
+        } else {
+          // Regular generation with potential tool calls
+          result = await this.provider.generateText({
+            messages,
+            temperature: options.temperature ?? this.temperature,
+            maxTokens: options.maxTokens ?? this.maxTokens,
+            tools,
+          });
+
+          if (result.toolCalls && result.toolCalls.length > 0) {
+            allToolCalls.push(...result.toolCalls);
+
+            // Add assistant message with tool calls
+            messages.push({
+              role: 'assistant',
+              content: result.text || null,
+              tool_calls: result.toolCalls,
+            });
+
+            // Execute tool calls
+            for (const tc of result.toolCalls) {
+              const toolTrace = trace.createChild({ name: `tool:${tc.function.name}` }).start();
+              
+              try {
+                const toolFn = this.toolRegistry.get(tc.function.name);
+                if (!toolFn) {
+                  throw new Error(`Tool '${tc.function.name}' not found`);
+                }
+
+                const args = JSON.parse(tc.function.arguments);
+                const toolResult = await toolFn.execute(args, {
+                  agentName: this.name,
+                  sessionId: this.sessionId,
+                  runId: run.id,
+                });
+
+                const resultStr = typeof toolResult === 'string' 
+                  ? toolResult 
+                  : JSON.stringify(toolResult);
+
+                messages.push({
+                  role: 'tool',
+                  content: resultStr,
+                  tool_call_id: tc.id,
+                });
+
+                toolTrace.complete({ result: resultStr });
+              } catch (error: any) {
+                messages.push({
+                  role: 'tool',
+                  content: `Error: ${error.message}`,
+                  tool_call_id: tc.id,
+                });
+                toolTrace.fail(error);
+              }
+            }
+
+            // Continue loop to get final response
+            continue;
+          }
+
+          // No tool calls, we have our final response
+          fullText = result.text;
+          break;
+        }
+      }
+
+      // Add assistant message to session
+      this.session.addMessage({
+        role: 'assistant',
+        content: fullText,
+        tool_calls: allToolCalls.length > 0 ? allToolCalls : undefined,
+      });
+
+      trace.complete({ responseLength: fullText.length });
+      run.complete();
+
+      // Parse structured output if schema was provided
+      let structured: any;
+      if (options.outputSchema) {
+        try {
+          structured = JSON.parse(fullText);
+        } catch {
+          // Not valid JSON
+        }
+      }
+
+      return {
+        text: fullText,
+        structured,
+        toolCalls: allToolCalls.length > 0 ? allToolCalls : undefined,
+        usage: result!.usage,
+        runId: run.id,
+        sessionId: this.sessionId,
+      };
+    } catch (error: any) {
+      trace.fail(error);
+      run.fail(error);
+      throw error;
+    }
+  }
+
+  /**
+   * Start method for compatibility
+   */
+  async start(prompt: string, options?: ChatOptions): Promise<string> {
+    const result = await this.chat(prompt, options);
+    return result.text;
+  }
+
+  /**
+   * Get the session
+   */
+  getSession(): Session {
+    return this.session;
+  }
+
+  /**
+   * Clear conversation history
+   */
+  clearHistory(): void {
+    this.session.clearMessages();
+  }
+}

src/praisonai-ts/src/agent/handoff.ts

@@ -0,0 +1,162 @@
+/**
+ * Agent Handoff - Transfer conversations between agents
+ */
+
+import type { EnhancedAgent } from './enhanced';
+
+export interface HandoffConfig {
+  agent: EnhancedAgent;
+  name?: string;
+  description?: string;
+  condition?: (context: HandoffContext) => boolean;
+  transformContext?: (messages: any[]) => any[];
+}
+
+export interface HandoffContext {
+  messages: any[];
+  lastMessage: string;
+  topic?: string;
+  metadata?: Record<string, any>;
+}
+
+export interface HandoffResult {
+  handedOffTo: string;
+  response: string;
+  context: HandoffContext;
+}
+
+/**
+ * Handoff class - Represents a handoff target
+ */
+export class Handoff {
+  readonly targetAgent: EnhancedAgent;
+  readonly name: string;
+  readonly description: string;
+  readonly condition?: (context: HandoffContext) => boolean;
+  readonly transformContext?: (messages: any[]) => any[];
+
+  constructor(config: HandoffConfig) {
+    this.targetAgent = config.agent;
+    this.name = config.name || `handoff_to_${config.agent.name}`;
+    this.description = config.description || `Transfer conversation to ${config.agent.name}`;
+    this.condition = config.condition;
+    this.transformContext = config.transformContext;
+  }
+
+  /**
+   * Check if handoff should be triggered
+   */
+  shouldTrigger(context: HandoffContext): boolean {
+    if (this.condition) {
+      return this.condition(context);
+    }
+    return true;
+  }
+
+  /**
+   * Execute the handoff
+   */
+  async execute(context: HandoffContext): Promise<HandoffResult> {
+    let messages = context.messages;
+    
+    if (this.transformContext) {
+      messages = this.transformContext(messages);
+    }
+
+    // Transfer context to target agent
+    const response = await this.targetAgent.chat(context.lastMessage);
+
+    return {
+      handedOffTo: this.targetAgent.name,
+      response: response.text,
+      context: {
+        ...context,
+        messages,
+      },
+    };
+  }
+
+  /**
+   * Get tool definition for LLM
+   */
+  getToolDefinition(): { name: string; description: string; parameters: any } {
+    return {
+      name: this.name,
+      description: this.description,
+      parameters: {
+        type: 'object',
+        properties: {
+          reason: {
+            type: 'string',
+            description: 'Reason for the handoff',
+          },
+        },
+        required: [],
+      },
+    };
+  }
+}
+
+/**
+ * Create a handoff configuration
+ */
+export function handoff(config: HandoffConfig): Handoff {
+  return new Handoff(config);
+}
+
+/**
+ * Handoff filters - Common condition functions
+ */
+export const handoffFilters = {
+  /**
+   * Trigger handoff based on topic keywords
+   */
+  topic: (keywords: string | string[]) => {
+    const keywordList = Array.isArray(keywords) ? keywords : [keywords];
+    return (context: HandoffContext): boolean => {
+      const lastMessage = context.lastMessage.toLowerCase();
+      return keywordList.some(kw => lastMessage.includes(kw.toLowerCase()));
+    };
+  },
+
+  /**
+   * Trigger handoff based on metadata
+   */
+  metadata: (key: string, value: any) => {
+    return (context: HandoffContext): boolean => {
+      return context.metadata?.[key] === value;
+    };
+  },
+
+  /**
+   * Always trigger
+   */
+  always: () => {
+    return (): boolean => true;
+  },
+
+  /**
+   * Never trigger (manual only)
+   */
+  never: () => {
+    return (): boolean => false;
+  },
+
+  /**
+   * Combine multiple conditions with AND
+   */
+  and: (...conditions: Array<(context: HandoffContext) => boolean>) => {
+    return (context: HandoffContext): boolean => {
+      return conditions.every(cond => cond(context));
+    };
+  },
+
+  /**
+   * Combine multiple conditions with OR
+   */
+  or: (...conditions: Array<(context: HandoffContext) => boolean>) => {
+    return (context: HandoffContext): boolean => {
+      return conditions.some(cond => cond(context));
+    };
+  },
+};

src/praisonai-ts/src/agent/router.ts

@@ -0,0 +1,146 @@
+/**
+ * Router Agent - Route requests to specialized agents
+ */
+
+import type { EnhancedAgent } from './enhanced';
+
+export interface RouteConfig {
+  agent: EnhancedAgent;
+  condition: (input: string, context?: RouteContext) => boolean | Promise<boolean>;
+  priority?: number;
+}
+
+export interface RouteContext {
+  history?: string[];
+  metadata?: Record<string, any>;
+}
+
+export interface RouterConfig {
+  name?: string;
+  routes: RouteConfig[];
+  defaultAgent?: EnhancedAgent;
+  verbose?: boolean;
+}
+
+/**
+ * Router Agent - Routes requests to the most appropriate agent
+ */
+export class RouterAgent {
+  readonly name: string;
+  private routes: RouteConfig[];
+  private defaultAgent?: EnhancedAgent;
+  private verbose: boolean;
+
+  constructor(config: RouterConfig) {
+    this.name = config.name || 'Router';
+    this.routes = config.routes.sort((a, b) => (b.priority || 0) - (a.priority || 0));
+    this.defaultAgent = config.defaultAgent;
+    this.verbose = config.verbose ?? false;
+  }
+
+  /**
+   * Route a request to the appropriate agent
+   */
+  async route(input: string, context?: RouteContext): Promise<{ agent: EnhancedAgent; response: string } | null> {
+    for (const route of this.routes) {
+      const matches = await route.condition(input, context);
+      if (matches) {
+        if (this.verbose) {
+          console.log(`[Router] Routing to: ${route.agent.name}`);
+        }
+        const response = await route.agent.chat(input);
+        return { agent: route.agent, response: response.text };
+      }
+    }
+
+    if (this.defaultAgent) {
+      if (this.verbose) {
+        console.log(`[Router] Using default agent: ${this.defaultAgent.name}`);
+      }
+      const response = await this.defaultAgent.chat(input);
+      return { agent: this.defaultAgent, response: response.text };
+    }
+
+    return null;
+  }
+
+  /**
+   * Add a route
+   */
+  addRoute(config: RouteConfig): this {
+    this.routes.push(config);
+    this.routes.sort((a, b) => (b.priority || 0) - (a.priority || 0));
+    return this;
+  }
+
+  /**
+   * Get all routes
+   */
+  getRoutes(): RouteConfig[] {
+    return [...this.routes];
+  }
+}
+
+/**
+ * Route condition helpers
+ */
+export const routeConditions = {
+  /**
+   * Match by keywords
+   */
+  keywords: (keywords: string | string[]) => {
+    const keywordList = Array.isArray(keywords) ? keywords : [keywords];
+    return (input: string): boolean => {
+      const lower = input.toLowerCase();
+      return keywordList.some(kw => lower.includes(kw.toLowerCase()));
+    };
+  },
+
+  /**
+   * Match by regex
+   */
+  pattern: (regex: RegExp) => {
+    return (input: string): boolean => regex.test(input);
+  },
+
+  /**
+   * Match by metadata
+   */
+  metadata: (key: string, value: any) => {
+    return (_input: string, context?: RouteContext): boolean => {
+      return context?.metadata?.[key] === value;
+    };
+  },
+
+  /**
+   * Always match (for default routes)
+   */
+  always: () => {
+    return (): boolean => true;
+  },
+
+  /**
+   * Combine conditions with AND
+   */
+  and: (...conditions: Array<(input: string, context?: RouteContext) => boolean>) => {
+    return (input: string, context?: RouteContext): boolean => {
+      return conditions.every(c => c(input, context));
+    };
+  },
+
+  /**
+   * Combine conditions with OR
+   */
+  or: (...conditions: Array<(input: string, context?: RouteContext) => boolean>) => {
+    return (input: string, context?: RouteContext): boolean => {
+      return conditions.some(c => c(input, context));
+    };
+  }
+};
+
+/**
+ * Create a router agent
+ */
+export function createRouter(config: RouterConfig): RouterAgent {
+  return new RouterAgent(config);
+}

src/praisonai-ts/src/cli/index.ts

@@ -0,0 +1,163 @@
+#!/usr/bin/env node
+/**
+ * PraisonAI TypeScript CLI
+ */
+
+import { createProvider, isProviderAvailable, getAvailableProviders } from '../llm/providers';
+import { Session } from '../session';
+import { tool, ToolRegistry } from '../tools/decorator';
+
+interface CLIOptions {
+  model?: string;
+  stream?: boolean;
+  verbose?: boolean;
+  sessionId?: string;
+}
+
+async function chat(prompt: string, options: CLIOptions = {}): Promise<string> {
+  const model = options.model || process.env.PRAISONAI_MODEL || 'openai/gpt-4o-mini';
+  const provider = createProvider(model);
+  
+  const session = new Session({ id: options.sessionId });
+  session.addMessage({ role: 'user', content: prompt });
+
+  if (options.stream) {
+    let fullText = '';
+    const stream = await provider.streamText({
+      messages: [{ role: 'user', content: prompt }],
+      onToken: (token) => {
+        process.stdout.write(token);
+        fullText += token;
+      }
+    });
+    
+    for await (const chunk of stream) {
+      // Stream is consumed by onToken
+    }
+    console.log(); // New line after streaming
+    return fullText;
+  } else {
+    const result = await provider.generateText({
+      messages: [{ role: 'user', content: prompt }]
+    });
+    console.log(result.text);
+    return result.text;
+  }
+}
+
+async function listProviders(): Promise<void> {
+  console.log('Available Providers:');
+  const available = getAvailableProviders();
+  
+  const providers = ['openai', 'anthropic', 'google'];
+  for (const p of providers) {
+    const status = available.includes(p) ? '✅' : '❌';
+    console.log(`  ${status} ${p}`);
+  }
+}
+
+async function version(): Promise<void> {
+  const pkg = require('../../package.json');
+  console.log(`praisonai v${pkg.version}`);
+}
+
+async function help(): Promise<void> {
+  console.log(`
+PraisonAI TypeScript CLI
+
+Usage:
+  praisonai-ts chat <prompt>     Chat with an AI agent
+  praisonai-ts providers         List available providers
+  praisonai-ts version           Show version
+  praisonai-ts help              Show this help
+
+Options:
+  --model, -m <model>     Model to use (e.g., openai/gpt-4o-mini)
+  --stream, -s            Enable streaming output
+  --verbose, -v           Verbose output
+  --session <id>          Session ID for conversation continuity
+
+Environment Variables:
+  OPENAI_API_KEY          OpenAI API key
+  ANTHROPIC_API_KEY       Anthropic API key
+  GOOGLE_API_KEY          Google API key
+  PRAISONAI_MODEL         Default model
+
+Examples:
+  praisonai-ts chat "Hello, how are you?"
+  praisonai-ts chat "Write a poem" --stream
+  praisonai-ts chat "Explain AI" -m anthropic/claude-sonnet-4-20250514
+  praisonai-ts providers
+`);
+}
+
+function parseArgs(args: string[]): { command: string; prompt: string; options: CLIOptions } {
+  const options: CLIOptions = {};
+  let command = 'help';
+  let prompt = '';
+  
+  let i = 0;
+  while (i < args.length) {
+    const arg = args[i];
+    
+    if (arg === '--model' || arg === '-m') {
+      options.model = args[++i];
+    } else if (arg === '--stream' || arg === '-s') {
+      options.stream = true;
+    } else if (arg === '--verbose' || arg === '-v') {
+      options.verbose = true;
+    } else if (arg === '--session') {
+      options.sessionId = args[++i];
+    } else if (!arg.startsWith('-')) {
+      if (!command || command === 'help') {
+        command = arg;
+      } else {
+        prompt = arg;
+      }
+    }
+    i++;
+  }
+  
+  return { command, prompt, options };
+}
+
+async function main(): Promise<void> {
+  const args = process.argv.slice(2);
+  const { command, prompt, options } = parseArgs(args);
+  
+  try {
+    switch (command) {
+      case 'chat':
+        if (!prompt) {
+          console.error('Error: Please provide a prompt');
+          process.exit(1);
+        }
+        await chat(prompt, options);
+        break;
+      case 'providers':
+        await listProviders();
+        break;
+      case 'version':
+        await version();
+        break;
+      case 'help':
+      default:
+        await help();
+        break;
+    }
+  } catch (error: any) {
+    console.error('Error:', error.message);
+    if (options.verbose) {
+      console.error(error.stack);
+    }
+    process.exit(1);
+  }
+}
+
+// Export for programmatic use
+export { chat, listProviders, version, help, parseArgs };
+
+// Run CLI if executed directly
+if (require.main === module) {
+  main();
+}

src/praisonai-ts/src/eval/index.ts

@@ -0,0 +1,207 @@
+/**
+ * Evaluation Framework - Accuracy, Performance, and Reliability evaluation
+ */
+
+export interface EvalResult {
+  passed: boolean;
+  score: number;
+  message?: string;
+  details?: Record<string, any>;
+  duration: number;
+}
+
+export interface AccuracyEvalConfig {
+  input: string;
+  expectedOutput: string;
+  actualOutput: string;
+  threshold?: number;
+}
+
+export interface PerformanceEvalConfig {
+  func: () => Promise<any>;
+  iterations?: number;
+  warmupRuns?: number;
+}
+
+export interface PerformanceResult extends EvalResult {
+  avgTime: number;
+  minTime: number;
+  maxTime: number;
+  p95Time: number;
+  times: number[];
+}
+
+export interface ReliabilityEvalConfig {
+  expectedToolCalls: string[];
+  actualToolCalls: string[];
+}
+
+/**
+ * Accuracy Evaluation - Compare actual output to expected
+ */
+export async function accuracyEval(config: AccuracyEvalConfig): Promise<EvalResult> {
+  const start = Date.now();
+  const threshold = config.threshold ?? 0.8;
+  
+  const similarity = calculateSimilarity(config.expectedOutput, config.actualOutput);
+  const passed = similarity >= threshold;
+  
+  return {
+    passed,
+    score: similarity,
+    message: passed ? 'Output matches expected' : 'Output does not match expected',
+    details: {
+      expected: config.expectedOutput,
+      actual: config.actualOutput,
+      threshold
+    },
+    duration: Date.now() - start
+  };
+}
+
+/**
+ * Performance Evaluation - Measure execution time
+ */
+export async function performanceEval(config: PerformanceEvalConfig): Promise<PerformanceResult> {
+  const iterations = config.iterations ?? 10;
+  const warmupRuns = config.warmupRuns ?? 2;
+  const times: number[] = [];
+  
+  // Warmup runs
+  for (let i = 0; i < warmupRuns; i++) {
+    await config.func();
+  }
+  
+  // Actual runs
+  const start = Date.now();
+  for (let i = 0; i < iterations; i++) {
+    const runStart = Date.now();
+    await config.func();
+    times.push(Date.now() - runStart);
+  }
+  
+  const sortedTimes = [...times].sort((a, b) => a - b);
+  const avgTime = times.reduce((a, b) => a + b, 0) / times.length;
+  const minTime = sortedTimes[0];
+  const maxTime = sortedTimes[sortedTimes.length - 1];
+  const p95Index = Math.floor(sortedTimes.length * 0.95);
+  const p95Time = sortedTimes[p95Index] || maxTime;
+  
+  return {
+    passed: true,
+    score: 1,
+    avgTime,
+    minTime,
+    maxTime,
+    p95Time,
+    times,
+    duration: Date.now() - start,
+    details: {
+      iterations,
+      warmupRuns
+    }
+  };
+}
+
+/**
+ * Reliability Evaluation - Check tool call accuracy
+ */
+export async function reliabilityEval(config: ReliabilityEvalConfig): Promise<EvalResult> {
+  const start = Date.now();
+  
+  const expected = new Set(config.expectedToolCalls);
+  const actual = new Set(config.actualToolCalls);
+  
+  const matched = config.expectedToolCalls.filter(t => actual.has(t));
+  const missing = config.expectedToolCalls.filter(t => !actual.has(t));
+  const extra = config.actualToolCalls.filter(t => !expected.has(t));
+  
+  const score = expected.size > 0 ? matched.length / expected.size : 1;
+  const passed = missing.length === 0;
+  
+  return {
+    passed,
+    score,
+    message: passed ? 'All expected tool calls made' : `Missing tool calls: ${missing.join(', ')}`,
+    details: {
+      matched,
+      missing,
+      extra,
+      expected: config.expectedToolCalls,
+      actual: config.actualToolCalls
+    },
+    duration: Date.now() - start
+  };
+}
+
+/**
+ * Calculate text similarity (simple Jaccard similarity)
+ */
+function calculateSimilarity(a: string, b: string): number {
+  const wordsA = new Set(a.toLowerCase().split(/\s+/));
+  const wordsB = new Set(b.toLowerCase().split(/\s+/));
+  
+  const intersection = new Set([...wordsA].filter(x => wordsB.has(x)));
+  const union = new Set([...wordsA, ...wordsB]);
+  
+  return union.size > 0 ? intersection.size / union.size : 0;
+}
+
+/**
+ * Eval Suite - Run multiple evaluations
+ */
+export class EvalSuite {
+  private results: Map<string, EvalResult> = new Map();
+
+  async runAccuracy(name: string, config: AccuracyEvalConfig): Promise<EvalResult> {
+    const result = await accuracyEval(config);
+    this.results.set(name, result);
+    return result;
+  }
+
+  async runPerformance(name: string, config: PerformanceEvalConfig): Promise<PerformanceResult> {
+    const result = await performanceEval(config);
+    this.results.set(name, result);
+    return result;
+  }
+
+  async runReliability(name: string, config: ReliabilityEvalConfig): Promise<EvalResult> {
+    const result = await reliabilityEval(config);
+    this.results.set(name, result);
+    return result;
+  }
+
+  getResults(): Map<string, EvalResult> {
+    return new Map(this.results);
+  }
+
+  getSummary(): { total: number; passed: number; failed: number; avgScore: number } {
+    const results = Array.from(this.results.values());
+    const passed = results.filter(r => r.passed).length;
+    const avgScore = results.length > 0 
+      ? results.reduce((a, b) => a + b.score, 0) / results.length 
+      : 0;
+    
+    return {
+      total: results.length,
+      passed,
+      failed: results.length - passed,
+      avgScore
+    };
+  }
+
+  printSummary(): void {
+    const summary = this.getSummary();
+    console.log('\n=== Evaluation Summary ===');
+    console.log(`Total: ${summary.total}`);
+    console.log(`Passed: ${summary.passed}`);
+    console.log(`Failed: ${summary.failed}`);
+    console.log(`Avg Score: ${(summary.avgScore * 100).toFixed(1)}%`);
+    
+    console.log('\nResults:');
+    for (const [name, result] of this.results) {
+      const status = result.passed ? '✅' : '❌';
+      console.log(`  ${status} ${name}: ${(result.score * 100).toFixed(1)}%`);
+    }
+  }
+}

src/praisonai-ts/src/guardrails/index.ts

@@ -0,0 +1,244 @@
+/**
+ * Guardrails - Input/output validation and safety checks
+ */
+
+export type GuardrailStatus = 'passed' | 'failed' | 'warning';
+
+export interface GuardrailResult {
+  status: GuardrailStatus;
+  message?: string;
+  details?: Record<string, any>;
+  modifiedContent?: string;
+}
+
+export type GuardrailFunction = (
+  content: string,
+  context?: GuardrailContext
+) => Promise<GuardrailResult> | GuardrailResult;
+
+export interface GuardrailContext {
+  role: 'input' | 'output';
+  agentName?: string;
+  sessionId?: string;
+  metadata?: Record<string, any>;
+}
+
+export interface GuardrailConfig {
+  name: string;
+  description?: string;
+  check: GuardrailFunction;
+  onFail?: 'block' | 'warn' | 'modify';
+}
+
+/**
+ * Guardrail class
+ */
+export class Guardrail {
+  readonly name: string;
+  readonly description: string;
+  readonly check: GuardrailFunction;
+  readonly onFail: 'block' | 'warn' | 'modify';
+
+  constructor(config: GuardrailConfig) {
+    this.name = config.name;
+    this.description = config.description || `Guardrail: ${config.name}`;
+    this.check = config.check;
+    this.onFail = config.onFail || 'block';
+  }
+
+  async run(content: string, context?: GuardrailContext): Promise<GuardrailResult> {
+    try {
+      return await this.check(content, context);
+    } catch (error: any) {
+      return {
+        status: 'failed',
+        message: `Guardrail error: ${error.message}`,
+        details: { error: error.message },
+      };
+    }
+  }
+}
+
+/**
+ * Create a guardrail
+ */
+export function guardrail(config: GuardrailConfig): Guardrail {
+  return new Guardrail(config);
+}
+
+/**
+ * Guardrail Manager - Run multiple guardrails
+ */
+export class GuardrailManager {
+  private guardrails: Guardrail[] = [];
+
+  add(g: Guardrail): this {
+    this.guardrails.push(g);
+    return this;
+  }
+
+  async runAll(
+    content: string,
+    context?: GuardrailContext
+  ): Promise<{ passed: boolean; results: Array<{ name: string; result: GuardrailResult }> }> {
+    const results: Array<{ name: string; result: GuardrailResult }> = [];
+    let passed = true;
+
+    for (const g of this.guardrails) {
+      const result = await g.run(content, context);
+      results.push({ name: g.name, result });
+
+      if (result.status === 'failed') {
+        passed = false;
+        if (g.onFail === 'block') {
+          break;
+        }
+      }
+    }
+
+    return { passed, results };
+  }
+
+  get count(): number {
+    return this.guardrails.length;
+  }
+}
+
+/**
+ * Built-in guardrails
+ */
+export const builtinGuardrails = {
+  /**
+   * Check for maximum length
+   */
+  maxLength: (maxChars: number): Guardrail => {
+    return guardrail({
+      name: 'max_length',
+      description: `Ensure content is under ${maxChars} characters`,
+      check: (content) => {
+        if (content.length > maxChars) {
+          return {
+            status: 'failed',
+            message: `Content exceeds maximum length of ${maxChars} characters`,
+            details: { length: content.length, max: maxChars },
+          };
+        }
+        return { status: 'passed' };
+      },
+    });
+  },
+
+  /**
+   * Check for minimum length
+   */
+  minLength: (minChars: number): Guardrail => {
+    return guardrail({
+      name: 'min_length',
+      description: `Ensure content is at least ${minChars} characters`,
+      check: (content) => {
+        if (content.length < minChars) {
+          return {
+            status: 'failed',
+            message: `Content is below minimum length of ${minChars} characters`,
+            details: { length: content.length, min: minChars },
+          };
+        }
+        return { status: 'passed' };
+      },
+    });
+  },
+
+  /**
+   * Check for blocked words
+   */
+  blockedWords: (words: string[]): Guardrail => {
+    return guardrail({
+      name: 'blocked_words',
+      description: 'Check for blocked words',
+      check: (content) => {
+        const lowerContent = content.toLowerCase();
+        const found = words.filter(w => lowerContent.includes(w.toLowerCase()));
+        if (found.length > 0) {
+          return {
+            status: 'failed',
+            message: `Content contains blocked words`,
+            details: { blockedWords: found },
+          };
+        }
+        return { status: 'passed' };
+      },
+    });
+  },
+
+  /**
+   * Check for required words
+   */
+  requiredWords: (words: string[]): Guardrail => {
+    return guardrail({
+      name: 'required_words',
+      description: 'Check for required words',
+      check: (content) => {
+        const lowerContent = content.toLowerCase();
+        const missing = words.filter(w => !lowerContent.includes(w.toLowerCase()));
+        if (missing.length > 0) {
+          return {
+            status: 'failed',
+            message: `Content missing required words`,
+            details: { missingWords: missing },
+          };
+        }
+        return { status: 'passed' };
+      },
+    });
+  },
+
+  /**
+   * Regex pattern check
+   */
+  pattern: (regex: RegExp, mustMatch: boolean = true): Guardrail => {
+    return guardrail({
+      name: 'pattern',
+      description: `Check content against pattern: ${regex}`,
+      check: (content) => {
+        const matches = regex.test(content);
+        if (mustMatch && !matches) {
+          return {
+            status: 'failed',
+            message: `Content does not match required pattern`,
+            details: { pattern: regex.toString() },
+          };
+        }
+        if (!mustMatch && matches) {
+          return {
+            status: 'failed',
+            message: `Content matches forbidden pattern`,
+            details: { pattern: regex.toString() },
+          };
+        }
+        return { status: 'passed' };
+      },
+    });
+  },
+
+  /**
+   * JSON validity check
+   */
+  validJson: (): Guardrail => {
+    return guardrail({
+      name: 'valid_json',
+      description: 'Ensure content is valid JSON',
+      check: (content) => {
+        try {
+          JSON.parse(content);
+          return { status: 'passed' };
+        } catch (e: any) {
+          return {
+            status: 'failed',
+            message: 'Content is not valid JSON',
+            details: { error: e.message },
+          };
+        }
+      },
+    });
+  },
+};

src/praisonai-ts/src/index.ts

@@ -4,4 +4,74 @@ export * from './knowledge';
 export * from './llm';
 export * from './memory';
 export * from './process';
-export * from './tools';
+
+// Export tools (excluding conflicting types)
+export { 
+  Tool, BaseTool, 
+  FunctionTool, tool, ToolRegistry, getRegistry, registerTool, getTool,
+  type ToolConfig, type ToolContext, type ToolParameters
+} from './tools';
+export * from './tools/arxivTools';
+export * from './tools/mcpSse';
+
+// Export session management
+export * from './session';
+
+// Export database adapters
+export * from './db';
+
+// Export workflows
+export * from './workflows';
+
+// Export guardrails
+export * from './guardrails';
+
+// Export handoff
+export { Handoff, handoff, handoffFilters, type HandoffConfig, type HandoffContext, type HandoffResult } from './agent/handoff';
+
+// Export router agent
+export { RouterAgent, createRouter, routeConditions, type RouterConfig, type RouteConfig, type RouteContext } from './agent/router';
+
+// Export context agent
+export { ContextAgent, createContextAgent, type ContextAgentConfig, type ContextMessage } from './agent/context';
+
+// Export knowledge base (RAG)
+export { KnowledgeBase, createKnowledgeBase, type Document, type SearchResult, type EmbeddingProvider, type KnowledgeBaseConfig } from './knowledge/rag';
+
+// Export evaluation framework
+export { accuracyEval, performanceEval, reliabilityEval, EvalSuite, type EvalResult, type PerformanceResult, type AccuracyEvalConfig, type PerformanceEvalConfig, type ReliabilityEvalConfig } from './eval';
+
+// Export observability
+export { MemoryObservabilityAdapter, ConsoleObservabilityAdapter, setObservabilityAdapter, getObservabilityAdapter, type ObservabilityAdapter, type TraceContext, type SpanContext, type SpanData, type TraceData } from './observability';
+
+// Export skills
+export { SkillManager, createSkillManager, parseSkillFile, type Skill, type SkillMetadata, type SkillDiscoveryOptions } from './skills';
+
+// Export CLI
+export { chat, listProviders, version, help } from './cli';
+
+// Export providers with explicit names to avoid conflicts
+export {
+  createProvider,
+  getDefaultProvider,
+  parseModelString,
+  isProviderAvailable,
+  getAvailableProviders,
+  OpenAIProvider,
+  AnthropicProvider,
+  GoogleProvider,
+  BaseProvider,
+  type LLMProvider,
+  type ProviderConfig,
+  type ProviderFactory,
+  type GenerateTextOptions,
+  type GenerateTextResult,
+  type StreamTextOptions,
+  type StreamChunk,
+  type GenerateObjectOptions,
+  type GenerateObjectResult,
+  type TokenUsage,
+  type Message as ProviderMessage,
+  type ToolCall,
+  type ToolDefinition as ProviderToolDefinition,
+} from './llm/providers';

src/praisonai-ts/src/knowledge/rag.ts

@@ -0,0 +1,193 @@
+/**
+ * Knowledge Base (RAG) - Retrieval Augmented Generation
+ */
+
+export interface Document {
+  id: string;
+  content: string;
+  metadata?: Record<string, any>;
+  embedding?: number[];
+}
+
+export interface SearchResult {
+  document: Document;
+  score: number;
+}
+
+export interface EmbeddingProvider {
+  embed(text: string): Promise<number[]>;
+  embedBatch(texts: string[]): Promise<number[][]>;
+}
+
+export interface KnowledgeBaseConfig {
+  embeddingProvider?: EmbeddingProvider;
+  similarityThreshold?: number;
+  maxResults?: number;
+}
+
+/**
+ * Simple in-memory vector store for RAG
+ */
+export class KnowledgeBase {
+  private documents: Map<string, Document> = new Map();
+  private embeddingProvider?: EmbeddingProvider;
+  private similarityThreshold: number;
+  private maxResults: number;
+
+  constructor(config: KnowledgeBaseConfig = {}) {
+    this.embeddingProvider = config.embeddingProvider;
+    this.similarityThreshold = config.similarityThreshold ?? 0.7;
+    this.maxResults = config.maxResults ?? 5;
+  }
+
+  /**
+   * Add a document to the knowledge base
+   */
+  async add(doc: Omit<Document, 'embedding'>): Promise<Document> {
+    const document: Document = { ...doc };
+    
+    if (this.embeddingProvider) {
+      document.embedding = await this.embeddingProvider.embed(doc.content);
+    }
+    
+    this.documents.set(doc.id, document);
+    return document;
+  }
+
+  /**
+   * Add multiple documents
+   */
+  async addBatch(docs: Array<Omit<Document, 'embedding'>>): Promise<Document[]> {
+    if (this.embeddingProvider && docs.length > 0) {
+      const embeddings = await this.embeddingProvider.embedBatch(docs.map(d => d.content));
+      return Promise.all(docs.map(async (doc, i) => {
+        const document: Document = { ...doc, embedding: embeddings[i] };
+        this.documents.set(doc.id, document);
+        return document;
+      }));
+    }
+    
+    return Promise.all(docs.map(doc => this.add(doc)));
+  }
+
+  /**
+   * Get a document by ID
+   */
+  get(id: string): Document | undefined {
+    return this.documents.get(id);
+  }
+
+  /**
+   * Delete a document
+   */
+  delete(id: string): boolean {
+    return this.documents.delete(id);
+  }
+
+  /**
+   * Search for similar documents
+   */
+  async search(query: string, limit?: number): Promise<SearchResult[]> {
+    const maxResults = limit ?? this.maxResults;
+    
+    if (!this.embeddingProvider) {
+      // Fallback to simple text matching
+      return this.textSearch(query, maxResults);
+    }
+
+    const queryEmbedding = await this.embeddingProvider.embed(query);
+    const results: SearchResult[] = [];
+
+    for (const doc of this.documents.values()) {
+      if (doc.embedding) {
+        const score = this.cosineSimilarity(queryEmbedding, doc.embedding);
+        if (score >= this.similarityThreshold) {
+          results.push({ document: doc, score });
+        }
+      }
+    }
+
+    return results
+      .sort((a, b) => b.score - a.score)
+      .slice(0, maxResults);
+  }
+
+  /**
+   * Simple text-based search fallback
+   */
+  private textSearch(query: string, limit: number): SearchResult[] {
+    const queryLower = query.toLowerCase();
+    const results: SearchResult[] = [];
+
+    for (const doc of this.documents.values()) {
+      const contentLower = doc.content.toLowerCase();
+      if (contentLower.includes(queryLower)) {
+        const score = queryLower.length / contentLower.length;
+        results.push({ document: doc, score: Math.min(score * 10, 1) });
+      }
+    }
+
+    return results
+      .sort((a, b) => b.score - a.score)
+      .slice(0, limit);
+  }
+
+  /**
+   * Calculate cosine similarity between two vectors
+   */
+  private cosineSimilarity(a: number[], b: number[]): number {
+    if (a.length !== b.length) return 0;
+    
+    let dotProduct = 0;
+    let normA = 0;
+    let normB = 0;
+    
+    for (let i = 0; i < a.length; i++) {
+      dotProduct += a[i] * b[i];
+      normA += a[i] * a[i];
+      normB += b[i] * b[i];
+    }
+    
+    const denominator = Math.sqrt(normA) * Math.sqrt(normB);
+    return denominator === 0 ? 0 : dotProduct / denominator;
+  }
+
+  /**
+   * Get all documents
+   */
+  list(): Document[] {
+    return Array.from(this.documents.values());
+  }
+
+  /**
+   * Clear all documents
+   */
+  clear(): void {
+    this.documents.clear();
+  }
+
+  /**
+   * Get document count
+   */
+  get size(): number {
+    return this.documents.size;
+  }
+
+  /**
+   * Build context from search results for RAG
+   */
+  buildContext(results: SearchResult[]): string {
+    if (results.length === 0) return '';
+    
+    return results
+      .map((r, i) => `[${i + 1}] ${r.document.content}`)
+      .join('\n\n');
+  }
+}
+
+/**
+ * Create a knowledge base
+ */
+export function createKnowledgeBase(config?: KnowledgeBaseConfig): KnowledgeBase {
+  return new KnowledgeBase(config);
+}