Compare commits

..

16 Commits

Author SHA1 Message Date
github-actions[bot] 7a22c9f56d Release 0.3.12 (#416)
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
2024-11-13 13:28:23 +07:00
Huu Le 8431b788ad feat: Add form filling use case for TS and optimize workflows (#417) 2024-11-13 12:45:57 +07:00
Marcus Schiesser 2b712cebec chore: remove dead code 2024-11-07 10:13:47 +08:00
Huu Le 6edea6af5c enhance workflow code for Python (#412)
* enhance workflow shared code

* fix streaming

* refactor code

* add missing helper

* update

* update form filling

* add filters

* simplify the code

* simplify the code

* simplify the code

* update form filling

* update e2e

* update function calling agent

* fix unneeded condition

* Create light-parrots-work.md

* revert change on using functioncallingagent

* update readme

* clean code

* extract call one tool function

* update for blog use case

* fix streaming

* fix e2e

* fix missing await

* improve tools code

* improve assertion code

* skip form filling test for TS framework

* update for tools helper

---------

Co-authored-by: Marcus Schiesser <mail@marcusschiesser.de>
2024-11-06 14:38:12 +07:00
Tom Aarsen d79d1652d1 Add new example HF embedding models (#415)
from https://huggingface.co/models?library=sentence-transformers
2024-11-05 16:12:07 +07:00
github-actions[bot] 8ebd8d7039 Release 0.3.11 (#409)
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
2024-11-04 16:41:34 +07:00
Marcus Schiesser 2b8aaa835d Add support for local models via Hugging Face (#414) 2024-11-04 16:39:27 +07:00
Huu Le 1fe21f85bd chore: Fix highlight.js issue with Next.js static build (#413) 2024-11-04 14:25:26 +07:00
Marcus Schiesser b9570b2eb9 fix: use generic LLMAgent instead of OpenAIAgent (adds support for Gemini and Anthropic for Agentic RAG) (#410) 2024-11-04 11:34:13 +07:00
Thuc Pham 00009ae53e feat: import pdf css (#408) 2024-11-01 17:21:08 +07:00
github-actions[bot] 63558c11fa Release 0.3.10 (#407)
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
2024-11-01 16:07:15 +07:00
Thuc Pham 9172fed2e8 feat: bump LITS 0.8.2 (#406)
---------
Co-authored-by: Marcus Schiesser <mail@marcusschiesser.de>
2024-11-01 15:06:31 +07:00
Thuc Pham 78ccde78fc feat: integrate llamaindex chat-ui (#399)
---------
Co-authored-by: Marcus Schiesser <mail@marcusschiesser.de>
2024-11-01 12:19:29 +07:00
github-actions[bot] 02510703d8 Release 0.3.9 (#405)
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
2024-10-31 16:05:33 +07:00
Huu Le ed59927bd0 feat: Add form filling use case for Python (#403)
---------
Co-authored-by: Marcus Schiesser <mail@marcusschiesser.de>
2024-10-31 16:01:53 +07:00
Thuc Pham 9f866aa981 fix: use uploaded filename to build file url (#404) 2024-10-30 14:47:11 +07:00
106 changed files with 3858 additions and 3263 deletions
+30
View File
@@ -1,5 +1,35 @@
# create-llama
## 0.3.12
### Patch Changes
- 6edea6a: Optimize generated workflow code for Python
- 8431b78: Optimize Typescript multi-agent code
- 8431b78: Add form filling use case (Typescript)
## 0.3.11
### Patch Changes
- 2b8aaa8: Add support for local models via Hugging Face
- b9570b2: Fix: use generic LLMAgent instead of OpenAIAgent (adds support for Gemini and Anthropic for Agentic RAG)
- 1fe21f8: Fix the highlight.js issue with the Next.js static build
- 00009ae: feat: import pdf css
## 0.3.10
### Patch Changes
- 9172fed: feat: bump LITS 0.8.2
- 78ccde7: feat: use llamaindex chat-ui for nextjs frontend
## 0.3.9
### Patch Changes
- ed59927: Add form filling use case (Python)
## 0.3.8
### Patch Changes
+5 -1
View File
@@ -18,7 +18,7 @@ const templateUI: TemplateUI = "shadcn";
const templatePostInstallAction: TemplatePostInstallAction = "runApp";
const appType: AppType = templateFramework === "nextjs" ? "" : "--frontend";
const userMessage = "Write a blog post about physical standards for letters";
const templateAgents = ["financial_report", "blog"];
const templateAgents = ["financial_report", "blog", "form_filling"];
for (const agents of templateAgents) {
test.describe(`Test multiagent template ${agents} ${templateFramework} ${dataSource} ${templateUI} ${appType} ${templatePostInstallAction}`, async () => {
@@ -68,6 +68,10 @@ for (const agents of templateAgents) {
test("Frontend should be able to submit a message and receive the start of a streamed response", async ({
page,
}) => {
test.skip(
agents === "financial_report" || agents === "form_filling",
"Skip chat tests for financial report and form filling.",
);
await page.goto(`http://localhost:${port}`);
await page.fill("form textarea", userMessage);
+14
View File
@@ -336,6 +336,20 @@ const getModelEnvs = (modelConfig: ModelConfig): EnvVar[] => {
},
]
: []),
...(modelConfig.provider === "huggingface"
? [
{
name: "EMBEDDING_BACKEND",
description:
"The backend to use for the Sentence Transformers embedding model, either 'torch', 'onnx', or 'openvino'. Defaults to 'onnx'.",
},
{
name: "EMBEDDING_TRUST_REMOTE_CODE",
description:
"Whether to trust remote code for the embedding model, required for some models with custom code.",
},
]
: []),
...(modelConfig.provider === "t-systems"
? [
{
+68
View File
@@ -0,0 +1,68 @@
import prompts from "prompts";
import { ModelConfigParams } from ".";
import { questionHandlers, toChoice } from "../../questions/utils";
const MODELS = ["HuggingFaceH4/zephyr-7b-alpha"];
type ModelData = {
dimensions: number;
};
const EMBEDDING_MODELS: Record<string, ModelData> = {
"BAAI/bge-small-en-v1.5": { dimensions: 384 },
"BAAI/bge-base-en-v1.5": { dimensions: 768 },
"BAAI/bge-large-en-v1.5": { dimensions: 1024 },
"sentence-transformers/all-MiniLM-L6-v2": { dimensions: 384 },
"sentence-transformers/all-mpnet-base-v2": { dimensions: 768 },
"intfloat/multilingual-e5-large": { dimensions: 1024 },
"mixedbread-ai/mxbai-embed-large-v1": { dimensions: 1024 },
"nomic-ai/nomic-embed-text-v1.5": { dimensions: 768 },
};
const DEFAULT_MODEL = MODELS[0];
const DEFAULT_EMBEDDING_MODEL = Object.keys(EMBEDDING_MODELS)[0];
const DEFAULT_DIMENSIONS = Object.values(EMBEDDING_MODELS)[0].dimensions;
type HuggingfaceQuestionsParams = {
askModels: boolean;
};
export async function askHuggingfaceQuestions({
askModels,
}: HuggingfaceQuestionsParams): Promise<ModelConfigParams> {
const config: ModelConfigParams = {
model: DEFAULT_MODEL,
embeddingModel: DEFAULT_EMBEDDING_MODEL,
dimensions: DEFAULT_DIMENSIONS,
isConfigured(): boolean {
return true;
},
};
if (askModels) {
const { model } = await prompts(
{
type: "select",
name: "model",
message: "Which Hugging Face model would you like to use?",
choices: MODELS.map(toChoice),
initial: 0,
},
questionHandlers,
);
config.model = model;
const { embeddingModel } = await prompts(
{
type: "select",
name: "embeddingModel",
message: "Which embedding model would you like to use?",
choices: Object.keys(EMBEDDING_MODELS).map(toChoice),
initial: 0,
},
questionHandlers,
);
config.embeddingModel = embeddingModel;
config.dimensions = EMBEDDING_MODELS[embeddingModel].dimensions;
}
return config;
}
+5
View File
@@ -5,6 +5,7 @@ import { askAnthropicQuestions } from "./anthropic";
import { askAzureQuestions } from "./azure";
import { askGeminiQuestions } from "./gemini";
import { askGroqQuestions } from "./groq";
import { askHuggingfaceQuestions } from "./huggingface";
import { askLLMHubQuestions } from "./llmhub";
import { askMistralQuestions } from "./mistral";
import { askOllamaQuestions } from "./ollama";
@@ -39,6 +40,7 @@ export async function askModelConfig({
if (framework === "fastapi") {
choices.push({ title: "T-Systems", value: "t-systems" });
choices.push({ title: "Huggingface", value: "huggingface" });
}
const { provider } = await prompts(
{
@@ -76,6 +78,9 @@ export async function askModelConfig({
case "t-systems":
modelConfig = await askLLMHubQuestions({ askModels });
break;
case "huggingface":
modelConfig = await askHuggingfaceQuestions({ askModels });
break;
default:
modelConfig = await askOpenAIQuestions({
openAiKey,
+15
View File
@@ -234,6 +234,21 @@ const getAdditionalDependencies = (
version: "0.2.4",
});
break;
case "huggingface":
dependencies.push({
name: "llama-index-llms-huggingface",
version: "^0.3.5",
});
dependencies.push({
name: "llama-index-embeddings-huggingface",
version: "^0.3.1",
});
dependencies.push({
name: "optimum",
version: "^1.23.3",
extras: ["onnxruntime"],
});
break;
case "t-systems":
dependencies.push({
name: "llama-index-agent-openai",
+16
View File
@@ -267,6 +267,22 @@ For better results, you can specify the region parameter to get results from a s
},
],
},
{
display: "Form Filling",
name: "form_filling",
supportedFrameworks: ["fastapi"],
type: ToolType.LOCAL,
dependencies: [
{
name: "pandas",
version: "^2.2.3",
},
{
name: "tabulate",
version: "^0.9.0",
},
],
},
];
export const getTool = (toolName: string): Tool | undefined => {
+2 -1
View File
@@ -9,6 +9,7 @@ export type ModelProvider =
| "gemini"
| "mistral"
| "azure-openai"
| "huggingface"
| "t-systems";
export type ModelConfig = {
provider: ModelProvider;
@@ -48,7 +49,7 @@ export type TemplateDataSource = {
};
export type TemplateDataSourceType = "file" | "web" | "db";
export type TemplateObservability = "none" | "traceloop" | "llamatrace";
export type TemplateAgents = "financial_report" | "blog";
export type TemplateAgents = "financial_report" | "blog" | "form_filling";
// Config for both file and folder
export type FileSourceConfig =
| {
+10 -7
View File
@@ -136,19 +136,22 @@ export const installTSTemplate = async ({
// Copy agents use case code for multiagent template
if (agents) {
console.log("\nCopying agent:", agents, "\n");
const useCasePath = path.join(compPath, "agents", "typescript", agents);
const agentsCodePath = path.join(useCasePath, "workflow");
const agentsCodePath = path.join(
compPath,
"agents",
"typescript",
agents,
);
// Copy agent codes
await copy("**", path.join(root, relativeEngineDestPath, "workflow"), {
parents: true,
cwd: agentsCodePath,
rename: assetRelocator,
});
// Copy use case files to project root
await copy("*.*", path.join(root), {
parents: true,
cwd: useCasePath,
rename: assetRelocator,
});
} else {
console.log(
red(
+1 -1
View File
@@ -1,6 +1,6 @@
{
"name": "create-llama",
"version": "0.3.8",
"version": "0.3.12",
"description": "Create LlamaIndex-powered apps with one command",
"keywords": [
"rag",
+28
View File
@@ -177,6 +177,34 @@ export const askProQuestions = async (program: QuestionArgs) => {
program.observability = observability;
}
// Ask agents
if (program.template === "multiagent" && !program.agents) {
const { agents } = await prompts(
{
type: "select",
name: "agents",
message: "Which agents would you like to use?",
choices: [
{
title: "Financial report (generate a financial report)",
value: "financial_report",
},
{
title: "Form filling (fill missing value in a CSV file)",
value: "form_filling",
},
{
title: "Blog writer (Write a blog post)",
value: "blog_writer",
},
],
initial: 0,
},
questionHandlers,
);
program.agents = agents;
}
if (!program.modelConfig) {
const modelConfig = await askModelConfig({
openAiKey: program.openAiKey,
+14 -1
View File
@@ -10,6 +10,7 @@ type AppType =
| "rag"
| "code_artifact"
| "financial_report_agent"
| "form_filling"
| "extractor"
| "data_scientist";
@@ -35,8 +36,12 @@ export const askSimpleQuestions = async (
title: "Financial Report Generator (using Workflows)",
value: "financial_report_agent",
},
{
title: "Form Filler (using Workflows)",
value: "form_filling",
},
{ title: "Code Artifact Agent", value: "code_artifact" },
{ title: "Structured extraction", value: "extractor" },
{ title: "Information Extractor", value: "extractor" },
],
},
questionHandlers,
@@ -152,6 +157,14 @@ const convertAnswers = async (
frontend: true,
modelConfig: MODEL_GPT4o,
},
form_filling: {
template: "multiagent",
agents: "form_filling",
tools: getTools(["form_filling"]),
dataSources: EXAMPLE_10K_SEC_FILES,
frontend: true,
modelConfig: MODEL_GPT4o,
},
extractor: {
template: "extractor",
tools: [],
@@ -8,9 +8,9 @@ This example is using three agents to generate a blog post:
There are three different methods how the agents can interact to reach their goal:
1. [Choreography](./app/examples/choreography.py) - the agents decide themselves to delegate a task to another agent
1. [Orchestrator](./app/examples/orchestrator.py) - a central orchestrator decides which agent should execute a task
1. [Explicit Workflow](./app/examples/workflow.py) - a pre-defined workflow specific for the task is used to execute the tasks
1. [Choreography](./app/agents/choreography.py) - the agents decide themselves to delegate a task to another agent
1. [Orchestrator](./app/agents/orchestrator.py) - a central orchestrator decides which agent should execute a task
1. [Explicit Workflow](./app/agents/workflow.py) - a pre-defined workflow specific for the task is used to execute the tasks
## Getting Started
@@ -11,11 +11,11 @@ def get_publisher_tools() -> Tuple[List[FunctionTool], str, str]:
tools = []
# Get configured tools from the tools.yaml file
configured_tools = ToolFactory.from_env(map_result=True)
if "document_generator" in configured_tools.keys():
tools.extend(configured_tools["document_generator"])
if "generate_document" in configured_tools.keys():
tools.append(configured_tools["generate_document"])
prompt_instructions = dedent("""
Normally, reply the blog post content to the user directly.
But if user requested to generate a file, use the document_generator tool to generate the file and reply the link to the file.
But if user requested to generate a file, use the generate_document tool to generate the file and reply the link to the file.
""")
description = "Expert in publishing the blog post, able to publish the blog post in PDF or HTML format."
else:
@@ -42,11 +42,15 @@ def _get_research_tools(**kwargs) -> QueryEngineTool:
query_engine_tool = _create_query_engine_tool(**kwargs)
if query_engine_tool is not None:
tools.append(query_engine_tool)
researcher_tool_names = ["duckduckgo", "wikipedia.WikipediaToolSpec"]
researcher_tool_names = [
"duckduckgo_search",
"duckduckgo_image_search",
"wikipedia.WikipediaToolSpec",
]
configured_tools = ToolFactory.from_env(map_result=True)
for tool_name, tool in configured_tools.items():
if tool_name in researcher_tool_names:
tools.extend(tool)
tools.append(tool)
return tools
@@ -0,0 +1,3 @@
from .blog import create_workflow
__all__ = ["create_workflow"]
@@ -4,17 +4,18 @@ from typing import List, Optional
from app.agents.choreography import create_choreography
from app.agents.orchestrator import create_orchestrator
from app.agents.workflow import create_workflow
from app.agents.workflow import create_workflow as create_blog_workflow
from llama_index.core.chat_engine.types import ChatMessage
from llama_index.core.workflow import Workflow
logger = logging.getLogger("uvicorn")
def get_chat_engine(
def create_workflow(
chat_history: Optional[List[ChatMessage]] = None, **kwargs
) -> Workflow:
# TODO: the EXAMPLE_TYPE could be passed as a chat config parameter?
# Chat filters are not supported yet
kwargs.pop("filters", None)
agent_type = os.getenv("EXAMPLE_TYPE", "").lower()
match agent_type:
case "choreography":
@@ -22,7 +23,7 @@ def get_chat_engine(
case "orchestrator":
agent = create_orchestrator(chat_history, **kwargs)
case _:
agent = create_workflow(chat_history, **kwargs)
agent = create_blog_workflow(chat_history, **kwargs)
logger.info(f"Using agent pattern: {agent_type}")
@@ -1,4 +1,5 @@
from abc import abstractmethod
from enum import Enum
from typing import Any, AsyncGenerator, List, Optional
from llama_index.core.llms import ChatMessage, ChatResponse
@@ -15,7 +16,7 @@ from llama_index.core.workflow import (
Workflow,
step,
)
from pydantic import BaseModel
from pydantic import BaseModel, Field
class InputEvent(Event):
@@ -26,17 +27,27 @@ class ToolCallEvent(Event):
tool_calls: list[ToolSelection]
class AgentRunEventType(Enum):
TEXT = "text"
PROGRESS = "progress"
class AgentRunEvent(Event):
name: str
_msg: str
msg: str
event_type: AgentRunEventType = Field(default=AgentRunEventType.TEXT)
data: Optional[dict] = None
@property
def msg(self):
return self._msg
@msg.setter
def msg(self, value):
self._msg = value
def to_response(self) -> dict:
return {
"type": "agent",
"data": {
"agent": self.name,
"type": self.event_type.value,
"text": self.msg,
"data": self.data,
},
}
class AgentRunResult(BaseModel):
@@ -33,7 +33,7 @@ curl --location 'localhost:8000/api/chat' \
--data '{ "messages": [{ "role": "user", "content": "Create a report comparing the finances of Apple and Tesla" }] }'
```
You can start editing the API by modifying `app/api/routers/chat.py` or `app/financial_report/workflow.py`. The API auto-updates as you save the files.
You can start editing the API by modifying `app/api/routers/chat.py` or `app/workflows/financial_report.py`. The API auto-updates as you save the files.
Open [http://localhost:8000/docs](http://localhost:8000/docs) with your browser to see the Swagger UI of the API.
@@ -1,47 +0,0 @@
from textwrap import dedent
from typing import List, Tuple
from app.engine.tools import ToolFactory
from app.workflows.single import FunctionCallingAgent
from llama_index.core.chat_engine.types import ChatMessage
from llama_index.core.tools import FunctionTool
def _get_analyst_params() -> Tuple[List[type[FunctionTool]], str, str]:
tools = []
prompt_instructions = dedent(
"""
You are an expert in analyzing financial data.
You are given a task and a set of financial data to analyze. Your task is to analyze the financial data and return a report.
Your response should include a detailed analysis of the financial data, including any trends, patterns, or insights that you find.
Construct the analysis in a textual format like tables would be great!
Don't need to synthesize the data, just analyze and provide your findings.
Always use the provided information, don't make up any information yourself.
"""
)
description = "Expert in analyzing financial data"
configured_tools = ToolFactory.from_env(map_result=True)
# Check if the interpreter tool is configured
if "interpreter" in configured_tools.keys():
tools.extend(configured_tools["interpreter"])
prompt_instructions += dedent("""
You are able to visualize the financial data using code interpreter tool.
It's very useful to create and include visualizations to the report (make sure you include the right code and data for the visualization).
Never include any code into the report, just the visualization.
""")
description += (
", able to visualize the financial data using code interpreter tool."
)
return tools, prompt_instructions, description
def create_analyst(chat_history: List[ChatMessage]):
tools, prompt_instructions, description = _get_analyst_params()
return FunctionCallingAgent(
name="analyst",
tools=tools,
description=description,
system_prompt=dedent(prompt_instructions),
chat_history=chat_history,
)
@@ -1,44 +0,0 @@
from textwrap import dedent
from typing import List, Tuple
from app.engine.tools import ToolFactory
from app.workflows.single import FunctionCallingAgent
from llama_index.core.chat_engine.types import ChatMessage
from llama_index.core.tools import BaseTool
def _get_reporter_params(
chat_history: List[ChatMessage],
) -> Tuple[List[type[BaseTool]], str, str]:
tools: List[type[BaseTool]] = []
description = "Expert in representing a financial report"
prompt_instructions = dedent(
"""
You are a report generation assistant tasked with producing a well-formatted report given parsed context.
Given a comprehensive analysis of the user request, your task is to synthesize the information and return a well-formatted report.
## Instructions
You are responsible for representing the analysis in a well-formatted report. If tables or visualizations provided, add them to the right sections that are most relevant.
Use only the provided information to create the report. Do not make up any information yourself.
Finally, the report should be presented in markdown format.
"""
)
configured_tools = ToolFactory.from_env(map_result=True)
if "document_generator" in configured_tools: # type: ignore
tools.extend(configured_tools["document_generator"]) # type: ignore
prompt_instructions += (
"\nYou are also able to generate a file document (PDF/HTML) of the report."
)
description += " and generate a file document (PDF/HTML) of the report."
return tools, description, prompt_instructions
def create_reporter(chat_history: List[ChatMessage]):
tools, description, prompt_instructions = _get_reporter_params(chat_history)
return FunctionCallingAgent(
name="reporter",
tools=tools,
description=description,
system_prompt=prompt_instructions,
chat_history=chat_history,
)
@@ -1,105 +0,0 @@
import os
from textwrap import dedent
from typing import List, Optional
from app.engine.index import IndexConfig, get_index
from app.workflows.single import FunctionCallingAgent
from llama_index.core.chat_engine.types import ChatMessage
from llama_index.core.tools import BaseTool, QueryEngineTool, ToolMetadata
from llama_index.indices.managed.llama_cloud import LlamaCloudIndex
def _create_query_engine_tools(params=None) -> Optional[list[type[BaseTool]]]:
"""
Provide an agent worker that can be used to query the index.
"""
# Add query tool if index exists
index_config = IndexConfig(**(params or {}))
index = get_index(index_config)
if index is None:
return None
top_k = int(os.getenv("TOP_K", 5))
# Construct query engine tools
tools = []
# If index is LlamaCloudIndex, we need to add chunk and doc retriever tools
if isinstance(index, LlamaCloudIndex):
# Document retriever
doc_retriever = index.as_query_engine(
retriever_mode="files_via_content",
similarity_top_k=top_k,
)
chunk_retriever = index.as_query_engine(
retriever_mode="chunks",
similarity_top_k=top_k,
)
tools.append(
QueryEngineTool(
query_engine=doc_retriever,
metadata=ToolMetadata(
name="document_retriever",
description=dedent(
"""
Document retriever that retrieves entire documents from the corpus.
ONLY use for research questions that may require searching over entire research reports.
Will be slower and more expensive than chunk-level retrieval but may be necessary.
"""
),
),
)
)
tools.append(
QueryEngineTool(
query_engine=chunk_retriever,
metadata=ToolMetadata(
name="chunk_retriever",
description=dedent(
"""
Retrieves a small set of relevant document chunks from the corpus.
Use for research questions that want to look up specific facts from the knowledge corpus,
and need entire documents.
"""
),
),
)
)
else:
query_engine = index.as_query_engine(
**({"similarity_top_k": top_k} if top_k != 0 else {})
)
tools.append(
QueryEngineTool(
query_engine=query_engine,
metadata=ToolMetadata(
name="retrieve_information",
description="Use this tool to retrieve information about the text corpus from the index.",
),
)
)
return tools
def create_researcher(chat_history: List[ChatMessage], **kwargs):
"""
Researcher is an agent that take responsibility for using tools to complete a given task.
"""
tools = _create_query_engine_tools(**kwargs)
if tools is None:
raise ValueError("No tools found for researcher agent")
return FunctionCallingAgent(
name="researcher",
tools=tools,
description="expert in retrieving any unknown content from the corpus",
system_prompt=dedent(
"""
You are a researcher agent. You are responsible for retrieving information from the corpus.
## Instructions
+ Don't synthesize the information, just return the whole retrieved information.
+ Don't need to retrieve the information that is already provided in the chat history and response with: "There is no new information, please reuse the information from the conversation."
"""
),
chat_history=chat_history,
)
@@ -1,177 +0,0 @@
from textwrap import dedent
from typing import AsyncGenerator, List, Optional
from app.agents.analyst import create_analyst
from app.agents.reporter import create_reporter
from app.agents.researcher import create_researcher
from app.workflows.single import AgentRunEvent, AgentRunResult, FunctionCallingAgent
from llama_index.core.chat_engine.types import ChatMessage
from llama_index.core.prompts import PromptTemplate
from llama_index.core.settings import Settings
from llama_index.core.workflow import (
Context,
Event,
StartEvent,
StopEvent,
Workflow,
step,
)
def create_workflow(chat_history: Optional[List[ChatMessage]] = None, **kwargs):
researcher = create_researcher(
chat_history=chat_history,
**kwargs,
)
analyst = create_analyst(chat_history=chat_history)
reporter = create_reporter(chat_history=chat_history)
workflow = FinancialReportWorkflow(timeout=360, chat_history=chat_history)
workflow.add_workflows(
researcher=researcher,
analyst=analyst,
reporter=reporter,
)
return workflow
class ResearchEvent(Event):
input: str
class AnalyzeEvent(Event):
input: str
class ReportEvent(Event):
input: str
class FinancialReportWorkflow(Workflow):
def __init__(
self, timeout: int = 360, chat_history: Optional[List[ChatMessage]] = None
):
super().__init__(timeout=timeout)
self.chat_history = chat_history or []
@step()
async def start(self, ctx: Context, ev: StartEvent) -> ResearchEvent | ReportEvent:
# set streaming
ctx.data["streaming"] = getattr(ev, "streaming", False)
# start the workflow with researching about a topic
ctx.data["task"] = ev.input
ctx.data["user_input"] = ev.input
# Decision-making process
decision = await self._decide_workflow(ev.input, self.chat_history)
if decision != "publish":
return ResearchEvent(input=f"Research for this task: {ev.input}")
else:
chat_history_str = "\n".join(
[f"{msg.role}: {msg.content}" for msg in self.chat_history]
)
return ReportEvent(
input=f"Create a report based on the chat history\n{chat_history_str}\n\n and task: {ev.input}"
)
async def _decide_workflow(
self, input: str, chat_history: List[ChatMessage]
) -> str:
# TODO: Refactor this by using prompt generation
prompt_template = PromptTemplate(
dedent(
"""
You are an expert in decision-making, helping people create financial reports for the provided data.
If the user doesn't need to add or update anything, respond with 'publish'.
Otherwise, respond with 'research'.
Here is the chat history:
{chat_history}
The current user request is:
{input}
Given the chat history and the new user request, decide whether to create a report based on existing information.
Decision (respond with either 'not_publish' or 'publish'):
"""
)
)
chat_history_str = "\n".join(
[f"{msg.role}: {msg.content}" for msg in chat_history]
)
prompt = prompt_template.format(chat_history=chat_history_str, input=input)
output = await Settings.llm.acomplete(prompt)
decision = output.text.strip().lower()
return "publish" if decision == "publish" else "research"
@step()
async def research(
self, ctx: Context, ev: ResearchEvent, researcher: FunctionCallingAgent
) -> AnalyzeEvent:
result: AgentRunResult = await self.run_agent(ctx, researcher, ev.input)
content = result.response.message.content
return AnalyzeEvent(
input=dedent(
f"""
Given the following research content:
{content}
Provide a comprehensive analysis of the data for the user's request: {ctx.data["task"]}
"""
)
)
@step()
async def analyze(
self, ctx: Context, ev: AnalyzeEvent, analyst: FunctionCallingAgent
) -> ReportEvent | StopEvent:
result: AgentRunResult = await self.run_agent(ctx, analyst, ev.input)
content = result.response.message.content
return ReportEvent(
input=dedent(
f"""
Given the following analysis:
{content}
Create a report for the user's request: {ctx.data["task"]}
"""
)
)
@step()
async def report(
self, ctx: Context, ev: ReportEvent, reporter: FunctionCallingAgent
) -> StopEvent:
try:
result: AgentRunResult = await self.run_agent(
ctx, reporter, ev.input, streaming=ctx.data["streaming"]
)
return StopEvent(result=result)
except Exception as e:
ctx.write_event_to_stream(
AgentRunEvent(
name=reporter.name,
msg=f"Error creating a report: {e}",
)
)
return StopEvent(result=None)
async def run_agent(
self,
ctx: Context,
agent: FunctionCallingAgent,
input: str,
streaming: bool = False,
) -> AgentRunResult | AsyncGenerator:
handler = agent.run(input=input, streaming=streaming)
# bubble all events while running the executor to the planner
async for event in handler.stream_events():
# Don't write the StopEvent from sub task to the stream
if type(event) is not StopEvent:
ctx.write_event_to_stream(event)
return await handler
@@ -1,12 +0,0 @@
from typing import List, Optional
from app.agents.workflow import create_workflow
from llama_index.core.chat_engine.types import ChatMessage
from llama_index.core.workflow import Workflow
def get_chat_engine(
chat_history: Optional[List[ChatMessage]] = None, **kwargs
) -> Workflow:
agent_workflow = create_workflow(chat_history, **kwargs)
return agent_workflow
@@ -0,0 +1,3 @@
from .financial_report import create_workflow
__all__ = ["create_workflow"]
@@ -0,0 +1,298 @@
import os
from typing import Any, Dict, List, Optional
from app.engine.index import IndexConfig, get_index
from app.engine.tools import ToolFactory
from app.workflows.events import AgentRunEvent
from app.workflows.tools import (
call_tools,
chat_with_tools,
)
from llama_index.core import Settings
from llama_index.core.base.llms.types import ChatMessage, MessageRole
from llama_index.core.indices.vector_store import VectorStoreIndex
from llama_index.core.llms.function_calling import FunctionCallingLLM
from llama_index.core.memory import ChatMemoryBuffer
from llama_index.core.tools import FunctionTool, QueryEngineTool, ToolSelection
from llama_index.core.workflow import (
Context,
Event,
StartEvent,
StopEvent,
Workflow,
step,
)
def create_workflow(
chat_history: Optional[List[ChatMessage]] = None,
params: Optional[Dict[str, Any]] = None,
filters: Optional[List[Any]] = None,
) -> Workflow:
index_config = IndexConfig(**params)
index: VectorStoreIndex = get_index(config=index_config)
if index is None:
query_engine_tool = None
else:
top_k = int(os.getenv("TOP_K", 10))
query_engine = index.as_query_engine(similarity_top_k=top_k, filters=filters)
query_engine_tool = QueryEngineTool.from_defaults(query_engine=query_engine)
configured_tools: Dict[str, FunctionTool] = ToolFactory.from_env(map_result=True) # type: ignore
code_interpreter_tool = configured_tools.get("interpret")
document_generator_tool = configured_tools.get("generate_document")
return FinancialReportWorkflow(
query_engine_tool=query_engine_tool,
code_interpreter_tool=code_interpreter_tool,
document_generator_tool=document_generator_tool,
chat_history=chat_history,
)
class InputEvent(Event):
input: List[ChatMessage]
response: bool = False
class ResearchEvent(Event):
input: list[ToolSelection]
class AnalyzeEvent(Event):
input: list[ToolSelection] | ChatMessage
class ReportEvent(Event):
input: list[ToolSelection]
class FinancialReportWorkflow(Workflow):
"""
A workflow to generate a financial report using indexed documents.
Requirements:
- Indexed documents containing financial data and a query engine tool to search them
- A code interpreter tool to analyze data and generate reports
- A document generator tool to create report files
Steps:
1. LLM Input: The LLM determines the next step based on function calling.
For example, if the model requests the query engine tool, it returns a ResearchEvent;
if it requests document generation, it returns a ReportEvent.
2. Research: Uses the query engine to find relevant chunks from indexed documents.
After gathering information, it requests analysis (step 3).
3. Analyze: Uses a custom prompt to analyze research results and can call the code
interpreter tool for visualization or calculation. Returns results to the LLM.
4. Report: Uses the document generator tool to create a report. Returns results to the LLM.
"""
_default_system_prompt = """
You are a financial analyst who are given a set of tools to help you.
It's good to using appropriate tools for the user request and always use the information from the tools, don't make up anything yourself.
For the query engine tool, you should break down the user request into a list of queries and call the tool with the queries.
"""
def __init__(
self,
query_engine_tool: QueryEngineTool,
code_interpreter_tool: FunctionTool,
document_generator_tool: FunctionTool,
llm: Optional[FunctionCallingLLM] = None,
timeout: int = 360,
chat_history: Optional[List[ChatMessage]] = None,
system_prompt: Optional[str] = None,
):
super().__init__(timeout=timeout)
self.system_prompt = system_prompt or self._default_system_prompt
self.chat_history = chat_history or []
self.query_engine_tool = query_engine_tool
self.code_interpreter_tool = code_interpreter_tool
self.document_generator_tool = document_generator_tool
assert (
query_engine_tool is not None
), "Query engine tool is not found. Try run generation script or upload a document file first."
assert code_interpreter_tool is not None, "Code interpreter tool is required"
assert (
document_generator_tool is not None
), "Document generator tool is required"
self.tools = [
self.query_engine_tool,
self.code_interpreter_tool,
self.document_generator_tool,
]
self.llm: FunctionCallingLLM = llm or Settings.llm
assert isinstance(self.llm, FunctionCallingLLM)
self.memory = ChatMemoryBuffer.from_defaults(
llm=self.llm, chat_history=self.chat_history
)
@step()
async def prepare_chat_history(self, ctx: Context, ev: StartEvent) -> InputEvent:
ctx.data["input"] = ev.input
if self.system_prompt:
system_msg = ChatMessage(
role=MessageRole.SYSTEM, content=self.system_prompt
)
self.memory.put(system_msg)
# Add user input to memory
self.memory.put(ChatMessage(role=MessageRole.USER, content=ev.input))
return InputEvent(input=self.memory.get())
@step()
async def handle_llm_input( # type: ignore
self,
ctx: Context,
ev: InputEvent,
) -> ResearchEvent | AnalyzeEvent | ReportEvent | StopEvent:
"""
Handle an LLM input and decide the next step.
"""
# Always use the latest chat history from the input
chat_history: list[ChatMessage] = ev.input
# Get tool calls
response = await chat_with_tools(
self.llm,
self.tools, # type: ignore
chat_history,
)
if not response.has_tool_calls():
# If no tool call, return the response generator
return StopEvent(result=response.generator)
# calling different tools at the same time is not supported at the moment
# add an error message to tell the AI to process step by step
if response.is_calling_different_tools():
self.memory.put(
ChatMessage(
role=MessageRole.ASSISTANT,
content="Cannot call different tools at the same time. Try calling one tool at a time.",
)
)
return InputEvent(input=self.memory.get())
self.memory.put(response.tool_call_message)
match response.tool_name():
case self.code_interpreter_tool.metadata.name:
return AnalyzeEvent(input=response.tool_calls)
case self.document_generator_tool.metadata.name:
return ReportEvent(input=response.tool_calls)
case self.query_engine_tool.metadata.name:
return ResearchEvent(input=response.tool_calls)
case _:
raise ValueError(f"Unknown tool: {response.tool_name()}")
@step()
async def research(self, ctx: Context, ev: ResearchEvent) -> AnalyzeEvent:
"""
Do a research to gather information for the user's request.
A researcher should have these tools: query engine, search engine, etc.
"""
ctx.write_event_to_stream(
AgentRunEvent(
name="Researcher",
msg="Starting research",
)
)
tool_calls = ev.input
tool_messages = await call_tools(
ctx=ctx,
agent_name="Researcher",
tools=[self.query_engine_tool],
tool_calls=tool_calls,
)
self.memory.put_messages(tool_messages)
return AnalyzeEvent(
input=ChatMessage(
role=MessageRole.ASSISTANT,
content="I've finished the research. Please analyze the result.",
),
)
@step()
async def analyze(self, ctx: Context, ev: AnalyzeEvent) -> InputEvent:
"""
Analyze the research result.
"""
ctx.write_event_to_stream(
AgentRunEvent(
name="Analyst",
msg="Starting analysis",
)
)
event_requested_by_workflow_llm = isinstance(ev.input, list)
# Requested by the workflow LLM Input step, it's a tool call
if event_requested_by_workflow_llm:
# Set the tool calls
tool_calls = ev.input
else:
# Otherwise, it's triggered by the research step
# Use a custom prompt and independent memory for the analyst agent
analysis_prompt = """
You are a financial analyst, you are given a research result and a set of tools to help you.
Always use the given information, don't make up anything yourself. If there is not enough information, you can asking for more information.
If you have enough numerical information, it's good to include some charts/visualizations to the report so you can use the code interpreter tool to generate a report.
"""
# This is handled by analyst agent
# Clone the shared memory to avoid conflicting with the workflow.
chat_history = self.memory.get()
chat_history.append(
ChatMessage(role=MessageRole.SYSTEM, content=analysis_prompt)
)
chat_history.append(ev.input) # type: ignore
# Check if the analyst agent needs to call tools
response = await chat_with_tools(
self.llm,
[self.code_interpreter_tool],
chat_history,
)
if not response.has_tool_calls():
# If no tool call, fallback analyst message to the workflow
analyst_msg = ChatMessage(
role=MessageRole.ASSISTANT,
content=await response.full_response(),
)
self.memory.put(analyst_msg)
return InputEvent(input=self.memory.get())
else:
# Set the tool calls and the tool call message to the memory
tool_calls = response.tool_calls
self.memory.put(response.tool_call_message)
# Call tools
tool_messages = await call_tools(
ctx=ctx,
agent_name="Analyst",
tools=[self.code_interpreter_tool],
tool_calls=tool_calls, # type: ignore
)
self.memory.put_messages(tool_messages)
# Fallback to the input with the latest chat history
return InputEvent(input=self.memory.get())
@step()
async def report(self, ctx: Context, ev: ReportEvent) -> InputEvent:
"""
Generate a report based on the analysis result.
"""
ctx.write_event_to_stream(
AgentRunEvent(
name="Reporter",
msg="Starting report generation",
)
)
tool_calls = ev.input
tool_messages = await call_tools(
ctx=ctx,
agent_name="Reporter",
tools=[self.document_generator_tool],
tool_calls=tool_calls,
)
self.memory.put_messages(tool_messages)
# After the tool calls, fallback to the input with the latest chat history
return InputEvent(input=self.memory.get())
@@ -0,0 +1,59 @@
This is a [LlamaIndex](https://www.llamaindex.ai/) multi-agents project using [Workflows](https://docs.llamaindex.ai/en/stable/understanding/workflows/).
## Getting Started
First, setup the environment with poetry:
> **_Note:_** This step is not needed if you are using the dev-container.
```shell
poetry install
```
Then check the parameters that have been pre-configured in the `.env` file in this directory.
Make sure you have the `OPENAI_API_KEY` set.
Second, run the development server:
```shell
poetry run python main.py
```
## Use Case: Filling Financial CSV Template
To reproduce the use case, start the [frontend](../frontend/README.md) and follow these steps in the frontend:
1. Upload the Apple and Tesla financial reports from the [data](./data) directory. Just send an empty message.
2. Upload the CSV file [sec_10k_template.csv](./sec_10k_template.csv) and send the message "Fill the missing cells in the CSV file".
The agent will fill the missing cells by retrieving the information from the uploaded financial reports and return a new CSV file with the filled cells.
### API endpoints
The example provides one streaming API endpoint `/api/chat`.
You can test the endpoint with the following curl request:
```
curl --location 'localhost:8000/api/chat' \
--header 'Content-Type: application/json' \
--data '{ "messages": [{ "role": "user", "content": "What can you do?" }] }'
```
You can start editing the API by modifying `app/api/routers/chat.py` or `app/workflows/form_filling.py`. The API auto-updates as you save the files.
Open [http://localhost:8000/docs](http://localhost:8000/docs) with your browser to see the Swagger UI of the API.
The API allows CORS for all origins to simplify development. You can change this behavior by setting the `ENVIRONMENT` environment variable to `prod`:
```
ENVIRONMENT=prod poetry run python main.py
```
## Learn More
To learn more about LlamaIndex, take a look at the following resources:
- [LlamaIndex Documentation](https://docs.llamaindex.ai) - learn about LlamaIndex.
- [Workflows Introduction](https://docs.llamaindex.ai/en/stable/understanding/workflows/) - learn about LlamaIndex workflows.
You can check out [the LlamaIndex GitHub repository](https://github.com/run-llama/llama_index) - your feedback and contributions are welcome!
@@ -0,0 +1,3 @@
from .form_filling import create_workflow
__all__ = ["create_workflow"]
@@ -0,0 +1,241 @@
import os
from typing import Any, Dict, List, Optional
from app.engine.index import IndexConfig, get_index
from app.engine.tools import ToolFactory
from app.workflows.events import AgentRunEvent
from app.workflows.tools import (
call_tools,
chat_with_tools,
)
from llama_index.core import Settings
from llama_index.core.base.llms.types import ChatMessage, MessageRole
from llama_index.core.indices.vector_store import VectorStoreIndex
from llama_index.core.llms.function_calling import FunctionCallingLLM
from llama_index.core.memory import ChatMemoryBuffer
from llama_index.core.tools import FunctionTool, QueryEngineTool, ToolSelection
from llama_index.core.workflow import (
Context,
Event,
StartEvent,
StopEvent,
Workflow,
step,
)
def create_workflow(
chat_history: Optional[List[ChatMessage]] = None,
params: Optional[Dict[str, Any]] = None,
filters: Optional[List[Any]] = None,
) -> Workflow:
if params is None:
params = {}
if filters is None:
filters = []
index_config = IndexConfig(**params)
index: VectorStoreIndex = get_index(config=index_config)
if index is None:
query_engine_tool = None
else:
top_k = int(os.getenv("TOP_K", 10))
query_engine = index.as_query_engine(similarity_top_k=top_k, filters=filters)
query_engine_tool = QueryEngineTool.from_defaults(query_engine=query_engine)
configured_tools = ToolFactory.from_env(map_result=True)
extractor_tool = configured_tools.get("extract_questions") # type: ignore
filling_tool = configured_tools.get("fill_form") # type: ignore
workflow = FormFillingWorkflow(
query_engine_tool=query_engine_tool,
extractor_tool=extractor_tool, # type: ignore
filling_tool=filling_tool, # type: ignore
chat_history=chat_history,
)
return workflow
class InputEvent(Event):
input: List[ChatMessage]
response: bool = False
class ExtractMissingCellsEvent(Event):
tool_calls: list[ToolSelection]
class FindAnswersEvent(Event):
tool_calls: list[ToolSelection]
class FillEvent(Event):
tool_calls: list[ToolSelection]
class FormFillingWorkflow(Workflow):
"""
A predefined workflow for filling missing cells in a CSV file.
Required tools:
- query_engine: A query engine to query for the answers to the questions.
- extract_question: Extract missing cells in a CSV file and generate questions to fill them.
- answer_question: Query for the answers to the questions.
Flow:
1. Extract missing cells in a CSV file and generate questions to fill them.
2. Query for the answers to the questions.
3. Fill the missing cells with the answers.
"""
_default_system_prompt = """
You are a helpful assistant who helps fill missing cells in a CSV file.
Only extract missing cells from CSV files.
Only use provided data - never make up any information yourself. Fill N/A if an answer is not found.
If there is no query engine tool or the gathered information has many N/A values indicating the questions don't match the data, respond with a warning and ask the user to upload a different file or connect to a knowledge base.
"""
def __init__(
self,
query_engine_tool: Optional[QueryEngineTool],
extractor_tool: FunctionTool,
filling_tool: FunctionTool,
llm: Optional[FunctionCallingLLM] = None,
timeout: int = 360,
chat_history: Optional[List[ChatMessage]] = None,
system_prompt: Optional[str] = None,
):
super().__init__(timeout=timeout)
self.system_prompt = system_prompt or self._default_system_prompt
self.chat_history = chat_history or []
self.query_engine_tool = query_engine_tool
self.extractor_tool = extractor_tool
self.filling_tool = filling_tool
if self.extractor_tool is None or self.filling_tool is None:
raise ValueError("Extractor and filling tools are required.")
self.tools = [self.extractor_tool, self.filling_tool]
if self.query_engine_tool is not None:
self.tools.append(self.query_engine_tool) # type: ignore
self.llm: FunctionCallingLLM = llm or Settings.llm
if not isinstance(self.llm, FunctionCallingLLM):
raise ValueError("FormFillingWorkflow only supports FunctionCallingLLM.")
self.memory = ChatMemoryBuffer.from_defaults(
llm=self.llm, chat_history=self.chat_history
)
@step()
async def start(self, ctx: Context, ev: StartEvent) -> InputEvent:
ctx.data["input"] = ev.input
if self.system_prompt:
system_msg = ChatMessage(
role=MessageRole.SYSTEM, content=self.system_prompt
)
self.memory.put(system_msg)
user_input = ev.input
user_msg = ChatMessage(role=MessageRole.USER, content=user_input)
self.memory.put(user_msg)
chat_history = self.memory.get()
return InputEvent(input=chat_history)
@step()
async def handle_llm_input( # type: ignore
self,
ctx: Context,
ev: InputEvent,
) -> ExtractMissingCellsEvent | FillEvent | StopEvent:
"""
Handle an LLM input and decide the next step.
"""
chat_history: list[ChatMessage] = ev.input
response = await chat_with_tools(
self.llm,
self.tools,
chat_history,
)
if not response.has_tool_calls():
return StopEvent(result=response.generator)
# calling different tools at the same time is not supported at the moment
# add an error message to tell the AI to process step by step
if response.is_calling_different_tools():
self.memory.put(
ChatMessage(
role=MessageRole.ASSISTANT,
content="Cannot call different tools at the same time. Try calling one tool at a time.",
)
)
return InputEvent(input=self.memory.get())
self.memory.put(response.tool_call_message)
match response.tool_name():
case self.extractor_tool.metadata.name:
return ExtractMissingCellsEvent(tool_calls=response.tool_calls)
case self.query_engine_tool.metadata.name:
return FindAnswersEvent(tool_calls=response.tool_calls)
case self.filling_tool.metadata.name:
return FillEvent(tool_calls=response.tool_calls)
case _:
raise ValueError(f"Unknown tool: {response.tool_name()}")
@step()
async def extract_missing_cells(
self, ctx: Context, ev: ExtractMissingCellsEvent
) -> InputEvent | FindAnswersEvent:
"""
Extract missing cells in a CSV file and generate questions to fill them.
"""
ctx.write_event_to_stream(
AgentRunEvent(
name="Extractor",
msg="Extracting missing cells",
)
)
# Call the extract questions tool
tool_messages = await call_tools(
agent_name="Extractor",
tools=[self.extractor_tool],
ctx=ctx,
tool_calls=ev.tool_calls,
)
self.memory.put_messages(tool_messages)
return InputEvent(input=self.memory.get())
@step()
async def find_answers(self, ctx: Context, ev: FindAnswersEvent) -> InputEvent:
"""
Call answer questions tool to query for the answers to the questions.
"""
ctx.write_event_to_stream(
AgentRunEvent(
name="Researcher",
msg="Finding answers for missing cells",
)
)
tool_messages = await call_tools(
ctx=ctx,
agent_name="Researcher",
tools=[self.query_engine_tool],
tool_calls=ev.tool_calls,
)
self.memory.put_messages(tool_messages)
return InputEvent(input=self.memory.get())
@step()
async def fill_cells(self, ctx: Context, ev: FillEvent) -> InputEvent:
"""
Call fill cells tool to fill the missing cells with the answers.
"""
ctx.write_event_to_stream(
AgentRunEvent(
name="Processor",
msg="Filling missing cells",
)
)
tool_messages = await call_tools(
agent_name="Processor",
tools=[self.filling_tool],
ctx=ctx,
tool_calls=ev.tool_calls,
)
self.memory.put_messages(tool_messages)
return InputEvent(input=self.memory.get())
@@ -0,0 +1,17 @@
Parameter,2023 Apple (AAPL),2023 Tesla (TSLA)
Revenue,,
Net Income,,
Earnings Per Share (EPS),,
Debt-to-Equity Ratio,,
Current Ratio,,
Gross Margin,,
Operating Margin,,
Net Profit Margin,,
Inventory Turnover,,
Accounts Receivable Turnover,,
Capital Expenditure,,
Research and Development Expense,,
Market Cap,,
Price to Earnings Ratio,,
Dividend Yield,,
Year-over-Year Growth Rate,,
1 Parameter 2023 Apple (AAPL) 2023 Tesla (TSLA)
2 Revenue
3 Net Income
4 Earnings Per Share (EPS)
5 Debt-to-Equity Ratio
6 Current Ratio
7 Gross Margin
8 Operating Margin
9 Net Profit Margin
10 Inventory Turnover
11 Accounts Receivable Turnover
12 Capital Expenditure
13 Research and Development Expense
14 Market Cap
15 Price to Earnings Ratio
16 Dividend Yield
17 Year-over-Year Growth Rate
@@ -1,230 +0,0 @@
import {
Context,
StartEvent,
StopEvent,
Workflow,
WorkflowEvent,
} from "@llamaindex/core/workflow";
import { Message } from "ai";
import { ChatMessage, ChatResponseChunk, Settings } from "llamaindex";
import { getAnnotations } from "../llamaindex/streaming/annotations";
import {
createPublisher,
createResearcher,
createReviewer,
createWriter,
} from "./agents";
import { AgentInput, AgentRunEvent } from "./type";
const TIMEOUT = 360 * 1000;
const MAX_ATTEMPTS = 2;
class ResearchEvent extends WorkflowEvent<{ input: string }> {}
class WriteEvent extends WorkflowEvent<{
input: string;
isGood: boolean;
}> {}
class ReviewEvent extends WorkflowEvent<{ input: string }> {}
class PublishEvent extends WorkflowEvent<{ input: string }> {}
const prepareChatHistory = (chatHistory: Message[]): ChatMessage[] => {
// By default, the chat history only contains the assistant and user messages
// all the agents messages are stored in annotation data which is not visible to the LLM
const MAX_AGENT_MESSAGES = 10;
const agentAnnotations = getAnnotations<{ agent: string; text: string }>(
chatHistory,
{ role: "assistant", type: "agent" },
).slice(-MAX_AGENT_MESSAGES);
const agentMessages = agentAnnotations
.map(
(annotation) =>
`\n<${annotation.data.agent}>\n${annotation.data.text}\n</${annotation.data.agent}>`,
)
.join("\n");
const agentContent = agentMessages
? "Here is the previous conversation of agents:\n" + agentMessages
: "";
if (agentContent) {
const agentMessage: ChatMessage = {
role: "assistant",
content: agentContent,
};
return [
...chatHistory.slice(0, -1),
agentMessage,
chatHistory.slice(-1)[0],
] as ChatMessage[];
}
return chatHistory as ChatMessage[];
};
export const createWorkflow = (messages: Message[], params?: any) => {
const chatHistoryWithAgentMessages = prepareChatHistory(messages);
const runAgent = async (
context: Context,
agent: Workflow,
input: AgentInput,
) => {
const run = agent.run(new StartEvent({ input }));
for await (const event of agent.streamEvents()) {
if (event.data instanceof AgentRunEvent) {
context.writeEventToStream(event.data);
}
}
return await run;
};
const start = async (context: Context, ev: StartEvent) => {
context.set("task", ev.data.input);
const chatHistoryStr = chatHistoryWithAgentMessages
.map((msg) => `${msg.role}: ${msg.content}`)
.join("\n");
// Decision-making process
const decision = await decideWorkflow(ev.data.input, chatHistoryStr);
if (decision !== "publish") {
return new ResearchEvent({
input: `Research for this task: ${ev.data.input}`,
});
} else {
return new PublishEvent({
input: `Publish content based on the chat history\n${chatHistoryStr}\n\n and task: ${ev.data.input}`,
});
}
};
const decideWorkflow = async (task: string, chatHistoryStr: string) => {
const llm = Settings.llm;
const prompt = `You are an expert in decision-making, helping people write and publish blog posts.
If the user is asking for a file or to publish content, respond with 'publish'.
If the user requests to write or update a blog post, respond with 'not_publish'.
Here is the chat history:
${chatHistoryStr}
The current user request is:
${task}
Given the chat history and the new user request, decide whether to publish based on existing information.
Decision (respond with either 'not_publish' or 'publish'):`;
const output = await llm.complete({ prompt: prompt });
const decision = output.text.trim().toLowerCase();
return decision === "publish" ? "publish" : "research";
};
const research = async (context: Context, ev: ResearchEvent) => {
const researcher = await createResearcher(
chatHistoryWithAgentMessages,
params,
);
const researchRes = await runAgent(context, researcher, {
message: ev.data.input,
});
const researchResult = researchRes.data.result;
return new WriteEvent({
input: `Write a blog post given this task: ${context.get("task")} using this research content: ${researchResult}`,
isGood: false,
});
};
const write = async (context: Context, ev: WriteEvent) => {
const writer = createWriter(chatHistoryWithAgentMessages);
context.set("attempts", context.get("attempts", 0) + 1);
const tooManyAttempts = context.get("attempts") > MAX_ATTEMPTS;
if (tooManyAttempts) {
context.writeEventToStream(
new AgentRunEvent({
name: "writer",
msg: `Too many attempts (${MAX_ATTEMPTS}) to write the blog post. Proceeding with the current version.`,
}),
);
}
if (ev.data.isGood || tooManyAttempts) {
// the blog post is good or too many attempts
// stream the final content
const result = await runAgent(context, writer, {
message: `Based on the reviewer's feedback, refine the post and return only the final version of the post. Here's the current version: ${ev.data.input}`,
streaming: true,
});
return result as unknown as StopEvent<AsyncGenerator<ChatResponseChunk>>;
}
const writeRes = await runAgent(context, writer, {
message: ev.data.input,
});
const writeResult = writeRes.data.result;
context.set("result", writeResult); // store the last result
return new ReviewEvent({ input: writeResult });
};
const review = async (context: Context, ev: ReviewEvent) => {
const reviewer = createReviewer(chatHistoryWithAgentMessages);
const reviewRes = await reviewer.run(
new StartEvent<AgentInput>({ input: { message: ev.data.input } }),
);
const reviewResult = reviewRes.data.result;
const oldContent = context.get("result");
const postIsGood = reviewResult.toLowerCase().includes("post is good");
context.writeEventToStream(
new AgentRunEvent({
name: "reviewer",
msg: `The post is ${postIsGood ? "" : "not "}good enough for publishing. Sending back to the writer${
postIsGood ? " for publication." : "."
}`,
}),
);
if (postIsGood) {
return new WriteEvent({
input: "",
isGood: true,
});
}
return new WriteEvent({
input: `Improve the writing of a given blog post by using a given review.
Blog post:
\`\`\`
${oldContent}
\`\`\`
Review:
\`\`\`
${reviewResult}
\`\`\``,
isGood: false,
});
};
const publish = async (context: Context, ev: PublishEvent) => {
const publisher = await createPublisher(chatHistoryWithAgentMessages);
const publishResult = await runAgent(context, publisher, {
message: `${ev.data.input}`,
streaming: true,
});
return publishResult as unknown as StopEvent<
AsyncGenerator<ChatResponseChunk>
>;
};
const workflow = new Workflow({ timeout: TIMEOUT, validate: true });
workflow.addStep(StartEvent, start, {
outputs: [ResearchEvent, PublishEvent],
});
workflow.addStep(ResearchEvent, research, { outputs: WriteEvent });
workflow.addStep(WriteEvent, write, { outputs: [ReviewEvent, StopEvent] });
workflow.addStep(ReviewEvent, review, { outputs: WriteEvent });
workflow.addStep(PublishEvent, publish, { outputs: StopEvent });
return workflow;
};
@@ -1,54 +0,0 @@
import fs from "fs/promises";
import { BaseToolWithCall, QueryEngineTool } from "llamaindex";
import path from "path";
import { getDataSource } from "../engine";
import { createTools } from "../engine/tools/index";
export const getQueryEngineTool = async (
params?: any,
): Promise<QueryEngineTool | null> => {
const index = await getDataSource(params);
if (!index) {
return null;
}
const topK = process.env.TOP_K ? parseInt(process.env.TOP_K) : undefined;
return new QueryEngineTool({
queryEngine: index.asQueryEngine({
similarityTopK: topK,
}),
metadata: {
name: "query_index",
description: `Use this tool to retrieve information about the text corpus from the index.`,
},
});
};
export const getAvailableTools = async () => {
const configFile = path.join("config", "tools.json");
let toolConfig: any;
const tools: BaseToolWithCall[] = [];
try {
toolConfig = JSON.parse(await fs.readFile(configFile, "utf8"));
} catch (e) {
console.info(`Could not read ${configFile} file. Using no tools.`);
}
if (toolConfig) {
tools.push(...(await createTools(toolConfig)));
}
const queryEngineTool = await getQueryEngineTool();
if (queryEngineTool) {
tools.push(queryEngineTool);
}
return tools;
};
export const lookupTools = async (
toolNames: string[],
): Promise<BaseToolWithCall[]> => {
const availableTools = await getAvailableTools();
return availableTools.filter((tool) =>
toolNames.includes(tool.metadata.name),
);
};
@@ -1,19 +1,16 @@
import { ChatMessage } from "llamaindex";
import { getTool } from "../engine/tools";
import { FunctionCallingAgent } from "./single-agent";
import { getQueryEngineTool, lookupTools } from "./tools";
import { getQueryEngineTools } from "./tools";
export const createResearcher = async (
chatHistory: ChatMessage[],
params?: any,
) => {
const queryEngineTool = await getQueryEngineTool(params);
const tools = (
await lookupTools([
"wikipedia_tool",
"duckduckgo_search",
"image_generator",
])
).concat(queryEngineTool ? [queryEngineTool] : []);
export const createResearcher = async (chatHistory: ChatMessage[]) => {
const queryEngineTools = await getQueryEngineTools();
const tools = [
await getTool("wikipedia_tool"),
await getTool("duckduckgo_search"),
await getTool("image_generator"),
...(queryEngineTools ? queryEngineTools : []),
].filter((tool) => tool !== undefined);
return new FunctionCallingAgent({
name: "researcher",
@@ -81,17 +78,17 @@ Example:
};
export const createPublisher = async (chatHistory: ChatMessage[]) => {
const tools = await lookupTools(["document_generator"]);
const tool = await getTool("document_generator");
let systemPrompt = `You are an expert in publishing blog posts. You are given a task to publish a blog post.
If the writer says that there was an error, you should reply with the error and not publish the post.`;
if (tools.length > 0) {
if (tool) {
systemPrompt = `${systemPrompt}.
If the user requests to generate a file, use the document_generator tool to generate the file and reply with the link to the file.
Otherwise, simply return the content of the post.`;
}
return new FunctionCallingAgent({
name: "publisher",
tools: tools,
tools: tool ? [tool] : [],
systemPrompt: systemPrompt,
chatHistory,
});
@@ -0,0 +1,291 @@
import {
HandlerContext,
StartEvent,
StopEvent,
Workflow,
WorkflowContext,
WorkflowEvent,
} from "@llamaindex/workflow";
import { ChatMessage, ChatResponseChunk, Settings } from "llamaindex";
import {
createPublisher,
createResearcher,
createReviewer,
createWriter,
} from "./agents";
import {
FunctionCallingAgent,
FunctionCallingAgentInput,
} from "./single-agent";
import { AgentInput, AgentRunEvent } from "./type";
const TIMEOUT = 360 * 1000;
const MAX_ATTEMPTS = 2;
class ResearchEvent extends WorkflowEvent<{ input: string }> {}
class WriteEvent extends WorkflowEvent<{
input: string;
isGood: boolean;
}> {}
class ReviewEvent extends WorkflowEvent<{ input: string }> {}
class PublishEvent extends WorkflowEvent<{ input: string }> {}
type BlogContext = {
task: string;
attempts: number;
result: string;
};
export const createWorkflow = ({
chatHistory,
params,
}: {
chatHistory: ChatMessage[];
params?: any;
}) => {
const runAgent = async (
context: HandlerContext<BlogContext>,
agent: FunctionCallingAgent,
input: FunctionCallingAgentInput,
) => {
const agentContext = agent.run(input, {
streaming: input.streaming ?? false,
});
for await (const event of agentContext) {
if (event instanceof AgentRunEvent) {
context.sendEvent(event);
}
if (event instanceof StopEvent) {
return event;
}
}
return null;
};
const start = async (
context: HandlerContext<BlogContext>,
ev: StartEvent<AgentInput>,
) => {
const chatHistoryStr = chatHistory
.map((msg) => `${msg.role}: ${msg.content}`)
.join("\n");
// Decision-making process
const decision = await decideWorkflow(
ev.data.message.toString(),
chatHistoryStr,
);
if (decision !== "publish") {
return new ResearchEvent({
input: `Research for this task: ${JSON.stringify(context.data.task)}`,
});
} else {
return new PublishEvent({
input: `Publish content based on the chat history\n${chatHistoryStr}\n\n and task: ${context.data.task}`,
});
}
};
const decideWorkflow = async (task: string, chatHistoryStr: string) => {
const llm = Settings.llm;
const prompt = `You are an expert in decision-making, helping people write and publish blog posts.
If the user is asking for a file or to publish content, respond with 'publish'.
If the user requests to write or update a blog post, respond with 'not_publish'.
Here is the chat history:
${chatHistoryStr}
The current user request is:
${task}
Given the chat history and the new user request, decide whether to publish based on existing information.
Decision (respond with either 'not_publish' or 'publish'):`;
const output = await llm.complete({ prompt: prompt });
const decision = output.text.trim().toLowerCase();
return decision === "publish" ? "publish" : "research";
};
const research = async (
context: HandlerContext<BlogContext>,
ev: ResearchEvent,
) => {
const researcher = await createResearcher(chatHistory);
const researchRes = await runAgent(context, researcher, {
displayName: "Researcher",
message: ev.data.input,
});
const researchResult = researchRes?.data;
return new WriteEvent({
input: `Write a blog post given this task: ${JSON.stringify(
context.data.task,
)} using this research content: ${researchResult}`,
isGood: false,
});
};
const write = async (
context: HandlerContext<BlogContext>,
ev: WriteEvent,
) => {
const writer = createWriter(chatHistory);
context.data.attempts = context.data.attempts + 1;
const tooManyAttempts = context.data.attempts > MAX_ATTEMPTS;
if (tooManyAttempts) {
context.sendEvent(
new AgentRunEvent({
agent: "writer",
text: `Too many attempts (${MAX_ATTEMPTS}) to write the blog post. Proceeding with the current version.`,
type: "text",
}),
);
}
if (ev.data.isGood || tooManyAttempts) {
// the blog post is good or too many attempts
// stream the final content
const result = await runAgent(context, writer, {
message: `Based on the reviewer's feedback, refine the post and return only the final version of the post. Here's the current version: ${ev.data.input}`,
displayName: "Writer",
streaming: true,
});
return result as unknown as StopEvent<AsyncGenerator<ChatResponseChunk>>;
}
const writeRes = await runAgent(context, writer, {
message: ev.data.input,
displayName: "Writer",
streaming: false,
});
const writeResult = writeRes?.data;
context.data.result = writeResult; // store the last result
return new ReviewEvent({ input: writeResult });
};
const review = async (
context: HandlerContext<BlogContext>,
ev: ReviewEvent,
) => {
const reviewer = createReviewer(chatHistory);
const reviewResult = (await runAgent(context, reviewer, {
message: ev.data.input,
displayName: "Reviewer",
streaming: false,
})) as unknown as StopEvent<string>;
const reviewResultStr = reviewResult.data;
const oldContent = context.data.result;
const postIsGood = reviewResultStr.toLowerCase().includes("post is good");
context.sendEvent(
new AgentRunEvent({
agent: "reviewer",
text: `The post is ${postIsGood ? "" : "not "}good enough for publishing. Sending back to the writer${
postIsGood ? " for publication." : "."
}`,
type: "text",
}),
);
if (postIsGood) {
return new WriteEvent({
input: "",
isGood: true,
});
}
return new WriteEvent({
input: `Improve the writing of a given blog post by using a given review.
Blog post:
\`\`\`
${oldContent}
\`\`\`
Review:
\`\`\`
${reviewResult}
\`\`\``,
isGood: false,
});
};
const publish = async (
context: HandlerContext<BlogContext>,
ev: PublishEvent,
) => {
const publisher = await createPublisher(chatHistory);
const publishResult = await runAgent(context, publisher, {
message: `${ev.data.input}`,
displayName: "Publisher",
streaming: true,
});
return publishResult as unknown as StopEvent<
AsyncGenerator<ChatResponseChunk>
>;
};
const workflow: Workflow<
BlogContext,
AgentInput,
string | AsyncGenerator<boolean | ChatResponseChunk>
> = new Workflow();
workflow.addStep(
{
inputs: [StartEvent<AgentInput>],
outputs: [ResearchEvent, PublishEvent],
},
start,
);
workflow.addStep(
{
inputs: [ResearchEvent],
outputs: [WriteEvent],
},
research,
);
workflow.addStep(
{
inputs: [WriteEvent],
outputs: [ReviewEvent, StopEvent<AsyncGenerator<ChatResponseChunk>>],
},
write,
);
workflow.addStep(
{
inputs: [ReviewEvent],
outputs: [WriteEvent],
},
review,
);
workflow.addStep(
{
inputs: [PublishEvent],
outputs: [StopEvent],
},
publish,
);
// Overload run method to initialize the context
workflow.run = function (
input: AgentInput,
): WorkflowContext<
AgentInput,
string | AsyncGenerator<boolean | ChatResponseChunk>,
BlogContext
> {
return Workflow.prototype.run.call(workflow, new StartEvent(input), {
task: input.message.toString(),
attempts: 0,
result: "",
});
};
return workflow;
};
@@ -0,0 +1,47 @@
This is a [LlamaIndex](https://www.llamaindex.ai/) project using [Next.js](https://nextjs.org/) bootstrapped with [`create-llama`](https://github.com/run-llama/LlamaIndexTS/tree/main/packages/create-llama).
## Getting Started
First, install the dependencies:
```
npm install
```
Then check the parameters that have been pre-configured in the `.env` file in this directory.
Make sure you have the `OPENAI_API_KEY` set.
Second, generate the embeddings of the documents in the `./data` directory:
```
npm run generate
```
Third, run the development server:
```
npm run dev
```
Open [http://localhost:3000](http://localhost:3000) with your browser to see the chat UI.
## Use Case: Filling Financial CSV Template
You can start by sending an request on the chat UI to create a report comparing the finances of Apple and Tesla.
Or you can test the `/api/chat` endpoint with the following curl request:
```
curl --location 'localhost:3000/api/chat' \
--header 'Content-Type: application/json' \
--data '{ "messages": [{ "role": "user", "content": "Create a report comparing the finances of Apple and Tesla" }] }'
```
## Learn More
To learn more about LlamaIndex, take a look at the following resources:
- [LlamaIndex Documentation](https://docs.llamaindex.ai) - learn about LlamaIndex (Python features).
- [LlamaIndexTS Documentation](https://ts.llamaindex.ai/docs/llamaindex) - learn about LlamaIndex (Typescript features).
- [Workflows Introduction](https://ts.llamaindex.ai/docs/llamaindex/guide/workflow) - learn about LlamaIndexTS workflows.
You can check out [the LlamaIndexTS GitHub repository](https://github.com/run-llama/LlamaIndexTS) - your feedback and contributions are welcome!
@@ -1,65 +0,0 @@
import { ChatMessage } from "llamaindex";
import { FunctionCallingAgent } from "./single-agent";
import { getQueryEngineTools, lookupTools } from "./tools";
export const createResearcher = async (
chatHistory: ChatMessage[],
params?: any,
) => {
const queryEngineTools = await getQueryEngineTools(params);
if (!queryEngineTools) {
throw new Error("Query engine tool not found");
}
return new FunctionCallingAgent({
name: "researcher",
tools: queryEngineTools,
systemPrompt: `You are a researcher agent. You are responsible for retrieving information from the corpus.
## Instructions:
+ Don't synthesize the information, just return the whole retrieved information.
+ Don't need to retrieve the information that is already provided in the chat history and respond with: "There is no new information, please reuse the information from the conversation."
`,
chatHistory,
});
};
export const createAnalyst = async (chatHistory: ChatMessage[]) => {
let systemPrompt = `You are an expert in analyzing financial data.
You are given a task and a set of financial data to analyze. Your task is to analyze the financial data and return a report.
Your response should include a detailed analysis of the financial data, including any trends, patterns, or insights that you find.
Construct the analysis in textual format; including tables would be great!
Don't need to synthesize the data, just analyze and provide your findings.
Always use the provided information, don't make up any information yourself.`;
const tools = await lookupTools(["interpreter"]);
if (tools.length > 0) {
systemPrompt = `${systemPrompt}
You are able to visualize the financial data using code interpreter tool.
It's very useful to create and include visualizations in the report. Never include any code in the report, just the visualization.`;
}
return new FunctionCallingAgent({
name: "analyst",
tools: tools,
chatHistory,
});
};
export const createReporter = async (chatHistory: ChatMessage[]) => {
const tools = await lookupTools(["document_generator"]);
let systemPrompt = `You are a report generation assistant tasked with producing a well-formatted report given parsed context.
Given a comprehensive analysis of the user request, your task is to synthesize the information and return a well-formatted report.
## Instructions
You are responsible for representing the analysis in a well-formatted report. If tables or visualizations are provided, add them to the most relevant sections.
Finally, the report should be presented in markdown format.`;
if (tools.length > 0) {
systemPrompt = `${systemPrompt}.
You are also able to generate an HTML file of the report.`;
}
return new FunctionCallingAgent({
name: "reporter",
tools: tools,
systemPrompt: systemPrompt,
chatHistory,
});
};
@@ -1,159 +0,0 @@
import {
Context,
StartEvent,
StopEvent,
Workflow,
WorkflowEvent,
} from "@llamaindex/core/workflow";
import { Message } from "ai";
import { ChatMessage, ChatResponseChunk, Settings } from "llamaindex";
import { getAnnotations } from "../llamaindex/streaming/annotations";
import { createAnalyst, createReporter, createResearcher } from "./agents";
import { AgentInput, AgentRunEvent } from "./type";
const TIMEOUT = 360 * 1000;
const MAX_ATTEMPTS = 2;
class ResearchEvent extends WorkflowEvent<{ input: string }> {}
class AnalyzeEvent extends WorkflowEvent<{ input: string }> {}
class ReportEvent extends WorkflowEvent<{ input: string }> {}
const prepareChatHistory = (chatHistory: Message[]): ChatMessage[] => {
// By default, the chat history only contains the assistant and user messages
// all the agents messages are stored in annotation data which is not visible to the LLM
const MAX_AGENT_MESSAGES = 10;
const agentAnnotations = getAnnotations<{ agent: string; text: string }>(
chatHistory,
{ role: "assistant", type: "agent" },
).slice(-MAX_AGENT_MESSAGES);
const agentMessages = agentAnnotations
.map(
(annotation) =>
`\n<${annotation.data.agent}>\n${annotation.data.text}\n</${annotation.data.agent}>`,
)
.join("\n");
const agentContent = agentMessages
? "Here is the previous conversation of agents:\n" + agentMessages
: "";
if (agentContent) {
const agentMessage: ChatMessage = {
role: "assistant",
content: agentContent,
};
return [
...chatHistory.slice(0, -1),
agentMessage,
chatHistory.slice(-1)[0],
] as ChatMessage[];
}
return chatHistory as ChatMessage[];
};
export const createWorkflow = (messages: Message[], params?: any) => {
const chatHistoryWithAgentMessages = prepareChatHistory(messages);
const runAgent = async (
context: Context,
agent: Workflow,
input: AgentInput,
) => {
const run = agent.run(new StartEvent({ input }));
for await (const event of agent.streamEvents()) {
if (event.data instanceof AgentRunEvent) {
context.writeEventToStream(event.data);
}
}
return await run;
};
const start = async (context: Context, ev: StartEvent) => {
context.set("task", ev.data.input);
const chatHistoryStr = chatHistoryWithAgentMessages
.map((msg) => `${msg.role}: ${msg.content}`)
.join("\n");
// Decision-making process
const decision = await decideWorkflow(ev.data.input, chatHistoryStr);
if (decision !== "publish") {
return new ResearchEvent({
input: `Research for this task: ${ev.data.input}`,
});
} else {
return new ReportEvent({
input: `Publish content based on the chat history\n${chatHistoryStr}\n\n and task: ${ev.data.input}`,
});
}
};
const decideWorkflow = async (task: string, chatHistoryStr: string) => {
const llm = Settings.llm;
const prompt = `You are an expert in decision-making, helping people write and publish blog posts.
If the user is asking for a file or to publish content, respond with 'publish'.
If the user requests to write or update a blog post, respond with 'not_publish'.
Here is the chat history:
${chatHistoryStr}
The current user request is:
${task}
Given the chat history and the new user request, decide whether to publish based on existing information.
Decision (respond with either 'not_publish' or 'publish'):`;
const output = await llm.complete({ prompt: prompt });
const decision = output.text.trim().toLowerCase();
return decision === "publish" ? "publish" : "research";
};
const research = async (context: Context, ev: ResearchEvent) => {
const researcher = await createResearcher(
chatHistoryWithAgentMessages,
params,
);
const researchRes = await runAgent(context, researcher, {
message: ev.data.input,
});
const researchResult = researchRes.data.result;
return new AnalyzeEvent({
input: `Write a blog post given this task: ${context.get("task")} using this research content: ${researchResult}`,
});
};
const analyze = async (context: Context, ev: AnalyzeEvent) => {
const analyst = await createAnalyst(chatHistoryWithAgentMessages);
const analyzeRes = await runAgent(context, analyst, {
message: ev.data.input,
});
return new ReportEvent({
input: `Publish content based on the chat history\n${analyzeRes.data.result}\n\n and task: ${ev.data.input}`,
});
};
const report = async (context: Context, ev: ReportEvent) => {
const reporter = await createReporter(chatHistoryWithAgentMessages);
const reportResult = await runAgent(context, reporter, {
message: `${ev.data.input}`,
streaming: true,
});
return reportResult as unknown as StopEvent<
AsyncGenerator<ChatResponseChunk>
>;
};
const workflow = new Workflow({ timeout: TIMEOUT, validate: true });
workflow.addStep(StartEvent, start, {
outputs: [ResearchEvent, ReportEvent],
});
workflow.addStep(ResearchEvent, research, { outputs: AnalyzeEvent });
workflow.addStep(AnalyzeEvent, analyze, { outputs: ReportEvent });
workflow.addStep(ReportEvent, report, { outputs: StopEvent });
return workflow;
};
@@ -1,86 +0,0 @@
import fs from "fs/promises";
import { BaseToolWithCall, LlamaCloudIndex, QueryEngineTool } from "llamaindex";
import path from "path";
import { getDataSource } from "../engine";
import { createTools } from "../engine/tools/index";
export const getQueryEngineTools = async (
params?: any,
): Promise<QueryEngineTool[] | null> => {
const topK = process.env.TOP_K ? parseInt(process.env.TOP_K) : undefined;
const index = await getDataSource(params);
if (!index) {
return null;
}
// index is LlamaCloudIndex use two query engine tools
if (index instanceof LlamaCloudIndex) {
return [
new QueryEngineTool({
queryEngine: index.asQueryEngine({
similarityTopK: topK,
retrieval_mode: "files_via_content",
}),
metadata: {
name: "document_retriever",
description: `Document retriever that retrieves entire documents from the corpus.
ONLY use for research questions that may require searching over entire research reports.
Will be slower and more expensive than chunk-level retrieval but may be necessary.`,
},
}),
new QueryEngineTool({
queryEngine: index.asQueryEngine({
similarityTopK: topK,
retrieval_mode: "chunks",
}),
metadata: {
name: "chunk_retriever",
description: `Retrieves a small set of relevant document chunks from the corpus.
Use for research questions that want to look up specific facts from the knowledge corpus,
and need entire documents.`,
},
}),
];
} else {
return [
new QueryEngineTool({
queryEngine: (index as any).asQueryEngine({
similarityTopK: topK,
}),
metadata: {
name: "retriever",
description: `Use this tool to retrieve information about the text corpus from the index.`,
},
}),
];
}
};
export const getAvailableTools = async () => {
const configFile = path.join("config", "tools.json");
let toolConfig: any;
const tools: BaseToolWithCall[] = [];
try {
toolConfig = JSON.parse(await fs.readFile(configFile, "utf8"));
} catch (e) {
console.info(`Could not read ${configFile} file. Using no tools.`);
}
if (toolConfig) {
tools.push(...(await createTools(toolConfig)));
}
const queryEngineTools = await getQueryEngineTools();
if (queryEngineTools) {
tools.push(...queryEngineTools);
}
return tools;
};
export const lookupTools = async (
toolNames: string[],
): Promise<BaseToolWithCall[]> => {
const availableTools = await getAvailableTools();
return availableTools.filter((tool) =>
toolNames.includes(tool.metadata.name),
);
};
@@ -0,0 +1,20 @@
import { ChatMessage, ToolCallLLM } from "llamaindex";
import { getTool } from "../engine/tools";
import { FinancialReportWorkflow } from "./fin-report";
import { getQueryEngineTools } from "./tools";
const TIMEOUT = 360 * 1000;
export async function createWorkflow(options: {
chatHistory: ChatMessage[];
llm?: ToolCallLLM;
}) {
return new FinancialReportWorkflow({
chatHistory: options.chatHistory,
queryEngineTools: (await getQueryEngineTools()) || [],
codeInterpreterTool: (await getTool("interpreter"))!,
documentGeneratorTool: (await getTool("document_generator"))!,
llm: options.llm,
timeout: TIMEOUT,
});
}
@@ -0,0 +1,322 @@
import {
HandlerContext,
StartEvent,
StopEvent,
Workflow,
WorkflowEvent,
} from "@llamaindex/workflow";
import {
BaseToolWithCall,
ChatMemoryBuffer,
ChatMessage,
ChatResponseChunk,
Settings,
ToolCall,
ToolCallLLM,
} from "llamaindex";
import { callTools, chatWithTools } from "./tools";
import { AgentInput, AgentRunEvent } from "./type";
// Create a custom event type
class InputEvent extends WorkflowEvent<{ input: ChatMessage[] }> {}
class ResearchEvent extends WorkflowEvent<{
toolCalls: ToolCall[];
}> {}
class AnalyzeEvent extends WorkflowEvent<{
input: ChatMessage | ToolCall[];
}> {}
class ReportGenerationEvent extends WorkflowEvent<{
toolCalls: ToolCall[];
}> {}
const DEFAULT_SYSTEM_PROMPT = `
You are a financial analyst who are given a set of tools to help you.
It's good to using appropriate tools for the user request and always use the information from the tools, don't make up anything yourself.
For the query engine tool, you should break down the user request into a list of queries and call the tool with the queries.
`;
export class FinancialReportWorkflow extends Workflow<
null,
AgentInput,
ChatResponseChunk
> {
llm: ToolCallLLM;
memory: ChatMemoryBuffer;
queryEngineTools: BaseToolWithCall[];
codeInterpreterTool: BaseToolWithCall;
documentGeneratorTool: BaseToolWithCall;
systemPrompt?: string;
constructor(options: {
llm?: ToolCallLLM;
chatHistory: ChatMessage[];
queryEngineTools: BaseToolWithCall[];
codeInterpreterTool: BaseToolWithCall;
documentGeneratorTool: BaseToolWithCall;
systemPrompt?: string;
verbose?: boolean;
timeout?: number;
}) {
super({
verbose: options?.verbose ?? false,
timeout: options?.timeout ?? 360,
});
this.llm = options.llm ?? (Settings.llm as ToolCallLLM);
if (!(this.llm instanceof ToolCallLLM)) {
throw new Error("LLM is not a ToolCallLLM");
}
this.systemPrompt = options.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;
this.queryEngineTools = options.queryEngineTools;
this.codeInterpreterTool = options.codeInterpreterTool;
this.documentGeneratorTool = options.documentGeneratorTool;
this.memory = new ChatMemoryBuffer({
llm: this.llm,
chatHistory: options.chatHistory,
});
// Add steps
this.addStep(
{
inputs: [StartEvent<AgentInput>],
outputs: [InputEvent],
},
this.prepareChatHistory,
);
this.addStep(
{
inputs: [InputEvent],
outputs: [
InputEvent,
ResearchEvent,
AnalyzeEvent,
ReportGenerationEvent,
StopEvent,
],
},
this.handleLLMInput,
);
this.addStep(
{
inputs: [ResearchEvent],
outputs: [AnalyzeEvent],
},
this.handleResearch,
);
this.addStep(
{
inputs: [AnalyzeEvent],
outputs: [InputEvent],
},
this.handleAnalyze,
);
this.addStep(
{
inputs: [ReportGenerationEvent],
outputs: [InputEvent],
},
this.handleReportGeneration,
);
}
prepareChatHistory = async (
ctx: HandlerContext<null>,
ev: StartEvent<AgentInput>,
): Promise<InputEvent> => {
const { message } = ev.data;
if (this.systemPrompt) {
this.memory.put({ role: "system", content: this.systemPrompt });
}
this.memory.put({ role: "user", content: message });
return new InputEvent({ input: this.memory.getMessages() });
};
handleLLMInput = async (
ctx: HandlerContext<null>,
ev: InputEvent,
): Promise<
| InputEvent
| ResearchEvent
| AnalyzeEvent
| ReportGenerationEvent
| StopEvent
> => {
const chatHistory = ev.data.input;
const tools = [this.codeInterpreterTool, this.documentGeneratorTool];
if (this.queryEngineTools) {
tools.push(...this.queryEngineTools);
}
const toolCallResponse = await chatWithTools(this.llm, tools, chatHistory);
if (!toolCallResponse.hasToolCall()) {
return new StopEvent(toolCallResponse.responseGenerator);
}
if (toolCallResponse.hasMultipleTools()) {
this.memory.put({
role: "assistant",
content:
"Calling different tools is not allowed. Please only use multiple calls of the same tool.",
});
return new InputEvent({ input: this.memory.getMessages() });
}
// Put the LLM tool call message into the memory
// And trigger the next step according to the tool call
if (toolCallResponse.toolCallMessage) {
this.memory.put(toolCallResponse.toolCallMessage);
}
const toolName = toolCallResponse.getToolNames()[0];
switch (toolName) {
case this.codeInterpreterTool.metadata.name:
return new AnalyzeEvent({
input: toolCallResponse.toolCalls,
});
case this.documentGeneratorTool.metadata.name:
return new ReportGenerationEvent({
toolCalls: toolCallResponse.toolCalls,
});
default:
if (
this.queryEngineTools &&
this.queryEngineTools.some((tool) => tool.metadata.name === toolName)
) {
return new ResearchEvent({
toolCalls: toolCallResponse.toolCalls,
});
}
throw new Error(`Unknown tool: ${toolName}`);
}
};
handleResearch = async (
ctx: HandlerContext<null>,
ev: ResearchEvent,
): Promise<AnalyzeEvent> => {
ctx.sendEvent(
new AgentRunEvent({
agent: "Researcher",
text: "Researching data",
type: "text",
}),
);
const { toolCalls } = ev.data;
const toolMsgs = await callTools({
tools: this.queryEngineTools,
toolCalls,
ctx,
agentName: "Researcher",
});
for (const toolMsg of toolMsgs) {
this.memory.put(toolMsg);
}
return new AnalyzeEvent({
input: {
role: "assistant",
content:
"I have finished researching the data, please analyze the data.",
},
});
};
/**
* Analyze a research result or a tool call for code interpreter from the LLM
*/
handleAnalyze = async (
ctx: HandlerContext<null>,
ev: AnalyzeEvent,
): Promise<InputEvent> => {
ctx.sendEvent(
new AgentRunEvent({
agent: "Analyst",
text: `Starting analysis`,
type: "text",
}),
);
// Request by workflow LLM, input is a list of tool calls
let toolCalls: ToolCall[] = [];
if (Array.isArray(ev.data.input)) {
toolCalls = ev.data.input;
} else {
// Requested by Researcher, input is a ChatMessage
// We start new LLM chat specifically for analyzing the data
const analysisPrompt = `
You are an expert in analyzing financial data.
You are given a set of financial data to analyze. Your task is to analyze the financial data and return a report.
Your response should include a detailed analysis of the financial data, including any trends, patterns, or insights that you find.
Construct the analysis in textual format; including tables would be great!
Don't need to synthesize the data, just analyze and provide your findings.
`;
// Clone the current chat history
// Add the analysis system prompt and the message from the researcher
const newChatHistory = [
...this.memory.getMessages(),
{ role: "system", content: analysisPrompt },
ev.data.input,
];
const toolCallResponse = await chatWithTools(
this.llm,
[this.codeInterpreterTool],
newChatHistory as ChatMessage[],
);
if (!toolCallResponse.hasToolCall()) {
this.memory.put(await toolCallResponse.asFullResponse());
return new InputEvent({
input: this.memory.getMessages(),
});
} else {
this.memory.put(toolCallResponse.toolCallMessage);
toolCalls = toolCallResponse.toolCalls;
}
}
// Call the tools
const toolMsgs = await callTools({
tools: [this.codeInterpreterTool],
toolCalls,
ctx,
agentName: "Analyst",
});
for (const toolMsg of toolMsgs) {
this.memory.put(toolMsg);
}
return new InputEvent({
input: this.memory.getMessages(),
});
};
handleReportGeneration = async (
ctx: HandlerContext<null>,
ev: ReportGenerationEvent,
): Promise<InputEvent> => {
const { toolCalls } = ev.data;
const toolMsgs = await callTools({
tools: [this.documentGeneratorTool],
toolCalls,
ctx,
agentName: "Reporter",
});
for (const toolMsg of toolMsgs) {
this.memory.put(toolMsg);
}
return new InputEvent({ input: this.memory.getMessages() });
};
}
@@ -0,0 +1,37 @@
This is a [LlamaIndex](https://www.llamaindex.ai/) project using [Next.js](https://nextjs.org/) bootstrapped with [`create-llama`](https://github.com/run-llama/LlamaIndexTS/tree/main/packages/create-llama).
## Getting Started
First, install the dependencies:
```
npm install
```
Then check the parameters that have been pre-configured in the `.env` file in this directory.
Make sure you have the `OPENAI_API_KEY` set.
Second, run the development server:
```
npm run dev
```
Open [http://localhost:3000](http://localhost:3000) with your browser to see the chat UI.
## Use Case: Filling Financial CSV Template
1. Upload the Apple and Tesla financial reports from the [data](./data) directory. Just send an empty message.
2. Upload the CSV file [sec_10k_template.csv](./sec_10k_template.csv) and send the message "Fill the missing cells in the CSV file".
The agent will fill the missing cells by retrieving the information from the uploaded financial reports and return a new CSV file with the filled cells.
## Learn More
To learn more about LlamaIndex, take a look at the following resources:
- [LlamaIndex Documentation](https://docs.llamaindex.ai) - learn about LlamaIndex (Python features).
- [LlamaIndexTS Documentation](https://ts.llamaindex.ai/docs/llamaindex) - learn about LlamaIndex (Typescript features).
- [Workflows Introduction](https://ts.llamaindex.ai/docs/llamaindex/guide/workflow) - learn about LlamaIndexTS workflows.
You can check out [the LlamaIndexTS GitHub repository](https://github.com/run-llama/LlamaIndexTS) - your feedback and contributions are welcome!
@@ -0,0 +1,17 @@
Parameter,2023 Apple (AAPL),2023 Tesla (TSLA)
Revenue,,
Net Income,,
Earnings Per Share (EPS),,
Debt-to-Equity Ratio,,
Current Ratio,,
Gross Margin,,
Operating Margin,,
Net Profit Margin,,
Inventory Turnover,,
Accounts Receivable Turnover,,
Capital Expenditure,,
Research and Development Expense,,
Market Cap,,
Price to Earnings Ratio,,
Dividend Yield,,
Year-over-Year Growth Rate,,
1 Parameter 2023 Apple (AAPL) 2023 Tesla (TSLA)
2 Revenue
3 Net Income
4 Earnings Per Share (EPS)
5 Debt-to-Equity Ratio
6 Current Ratio
7 Gross Margin
8 Operating Margin
9 Net Profit Margin
10 Inventory Turnover
11 Accounts Receivable Turnover
12 Capital Expenditure
13 Research and Development Expense
14 Market Cap
15 Price to Earnings Ratio
16 Dividend Yield
17 Year-over-Year Growth Rate
@@ -0,0 +1,20 @@
import { ChatMessage, ToolCallLLM } from "llamaindex";
import { getTool } from "../engine/tools";
import { FormFillingWorkflow } from "./form-filling";
import { getQueryEngineTools } from "./tools";
const TIMEOUT = 360 * 1000;
export async function createWorkflow(options: {
chatHistory: ChatMessage[];
llm?: ToolCallLLM;
}) {
return new FormFillingWorkflow({
chatHistory: options.chatHistory,
queryEngineTools: (await getQueryEngineTools()) || [],
extractorTool: (await getTool("extract_missing_cells"))!,
fillMissingCellsTool: (await getTool("fill_missing_cells"))!,
llm: options.llm,
timeout: TIMEOUT,
});
}
@@ -0,0 +1,275 @@
import {
HandlerContext,
StartEvent,
StopEvent,
Workflow,
WorkflowEvent,
} from "@llamaindex/workflow";
import {
BaseToolWithCall,
ChatMemoryBuffer,
ChatMessage,
ChatResponseChunk,
Settings,
ToolCall,
ToolCallLLM,
} from "llamaindex";
import { callTools, chatWithTools } from "./tools";
import { AgentInput, AgentRunEvent } from "./type";
// Create a custom event type
class InputEvent extends WorkflowEvent<{ input: ChatMessage[] }> {}
class ExtractMissingCellsEvent extends WorkflowEvent<{
toolCalls: ToolCall[];
}> {}
class FindAnswersEvent extends WorkflowEvent<{
toolCalls: ToolCall[];
}> {}
class FillMissingCellsEvent extends WorkflowEvent<{
toolCalls: ToolCall[];
}> {}
const DEFAULT_SYSTEM_PROMPT = `
You are a helpful assistant who helps fill missing cells in a CSV file.
Only use the information from the retriever tool - don't make up any information yourself. Fill N/A if an answer is not found.
If there is no retriever tool or the gathered information has many N/A values indicating the questions don't match the data, respond with a warning and ask the user to upload a different file or connect to a knowledge base.
You can make multiple tool calls at once but only call with the same tool.
Only use the local file path for the tools.
`;
export class FormFillingWorkflow extends Workflow<
null,
AgentInput,
ChatResponseChunk
> {
llm: ToolCallLLM;
memory: ChatMemoryBuffer;
extractorTool: BaseToolWithCall;
queryEngineTools?: BaseToolWithCall[];
fillMissingCellsTool: BaseToolWithCall;
systemPrompt?: string;
constructor(options: {
llm?: ToolCallLLM;
chatHistory: ChatMessage[];
extractorTool: BaseToolWithCall;
queryEngineTools?: BaseToolWithCall[];
fillMissingCellsTool: BaseToolWithCall;
systemPrompt?: string;
verbose?: boolean;
timeout?: number;
}) {
super({
verbose: options?.verbose ?? false,
timeout: options?.timeout ?? 360,
});
this.llm = options.llm ?? (Settings.llm as ToolCallLLM);
if (!(this.llm instanceof ToolCallLLM)) {
throw new Error("LLM is not a ToolCallLLM");
}
this.systemPrompt = options.systemPrompt ?? DEFAULT_SYSTEM_PROMPT;
this.extractorTool = options.extractorTool;
this.queryEngineTools = options.queryEngineTools;
this.fillMissingCellsTool = options.fillMissingCellsTool;
this.memory = new ChatMemoryBuffer({
llm: this.llm,
chatHistory: options.chatHistory,
});
// Add steps
this.addStep(
{
inputs: [StartEvent<AgentInput>],
outputs: [InputEvent],
},
this.prepareChatHistory,
);
this.addStep(
{
inputs: [InputEvent],
outputs: [
InputEvent,
ExtractMissingCellsEvent,
FindAnswersEvent,
FillMissingCellsEvent,
StopEvent,
],
},
this.handleLLMInput,
);
this.addStep(
{
inputs: [ExtractMissingCellsEvent],
outputs: [InputEvent],
},
this.handleExtractMissingCells,
);
this.addStep(
{
inputs: [FindAnswersEvent],
outputs: [InputEvent],
},
this.handleFindAnswers,
);
this.addStep(
{
inputs: [FillMissingCellsEvent],
outputs: [InputEvent],
},
this.handleFillMissingCells,
);
}
prepareChatHistory = async (
ctx: HandlerContext<null>,
ev: StartEvent<AgentInput>,
): Promise<InputEvent> => {
const { message } = ev.data;
if (this.systemPrompt) {
this.memory.put({ role: "system", content: this.systemPrompt });
}
this.memory.put({ role: "user", content: message });
return new InputEvent({ input: this.memory.getMessages() });
};
handleLLMInput = async (
ctx: HandlerContext<null>,
ev: InputEvent,
): Promise<
| InputEvent
| ExtractMissingCellsEvent
| FindAnswersEvent
| FillMissingCellsEvent
| StopEvent
> => {
const chatHistory = ev.data.input;
const tools = [this.extractorTool, this.fillMissingCellsTool];
if (this.queryEngineTools) {
tools.push(...this.queryEngineTools);
}
const toolCallResponse = await chatWithTools(this.llm, tools, chatHistory);
if (!toolCallResponse.hasToolCall()) {
return new StopEvent(toolCallResponse.responseGenerator);
}
if (toolCallResponse.hasMultipleTools()) {
this.memory.put({
role: "assistant",
content:
"Calling different tools is not allowed. Please only use multiple calls of the same tool.",
});
return new InputEvent({ input: this.memory.getMessages() });
}
// Put the LLM tool call message into the memory
// And trigger the next step according to the tool call
if (toolCallResponse.toolCallMessage) {
this.memory.put(toolCallResponse.toolCallMessage);
}
const toolName = toolCallResponse.getToolNames()[0];
switch (toolName) {
case this.extractorTool.metadata.name:
return new ExtractMissingCellsEvent({
toolCalls: toolCallResponse.toolCalls,
});
case this.fillMissingCellsTool.metadata.name:
return new FillMissingCellsEvent({
toolCalls: toolCallResponse.toolCalls,
});
default:
if (
this.queryEngineTools &&
this.queryEngineTools.some((tool) => tool.metadata.name === toolName)
) {
return new FindAnswersEvent({
toolCalls: toolCallResponse.toolCalls,
});
}
throw new Error(`Unknown tool: ${toolName}`);
}
};
handleExtractMissingCells = async (
ctx: HandlerContext<null>,
ev: ExtractMissingCellsEvent,
): Promise<InputEvent> => {
ctx.sendEvent(
new AgentRunEvent({
agent: "CSVExtractor",
text: "Extracting missing cells",
type: "text",
}),
);
const { toolCalls } = ev.data;
const toolMsgs = await callTools({
tools: [this.extractorTool],
toolCalls,
ctx,
agentName: "CSVExtractor",
});
for (const toolMsg of toolMsgs) {
this.memory.put(toolMsg);
}
return new InputEvent({ input: this.memory.getMessages() });
};
handleFindAnswers = async (
ctx: HandlerContext<null>,
ev: FindAnswersEvent,
): Promise<InputEvent> => {
const { toolCalls } = ev.data;
if (!this.queryEngineTools) {
throw new Error("Query engine tool is not available");
}
ctx.sendEvent(
new AgentRunEvent({
agent: "Researcher",
text: "Finding answers",
type: "text",
}),
);
const toolMsgs = await callTools({
tools: this.queryEngineTools,
toolCalls,
ctx,
agentName: "Researcher",
});
for (const toolMsg of toolMsgs) {
this.memory.put(toolMsg);
}
return new InputEvent({ input: this.memory.getMessages() });
};
handleFillMissingCells = async (
ctx: HandlerContext<null>,
ev: FillMissingCellsEvent,
): Promise<InputEvent> => {
const { toolCalls } = ev.data;
const toolMsgs = await callTools({
tools: [this.fillMissingCellsTool],
toolCalls,
ctx,
agentName: "Processor",
});
for (const toolMsg of toolMsgs) {
this.memory.put(toolMsg);
}
return new InputEvent({ input: this.memory.getMessages() });
};
}
@@ -56,7 +56,7 @@ class ToolFactory:
A dictionary of tool names to lists of FunctionTools if map_result is True,
otherwise a list of FunctionTools.
"""
tools: Union[Dict[str, List[FunctionTool]], List[FunctionTool]] = (
tools: Union[Dict[str, FunctionTool], List[FunctionTool]] = (
{} if map_result else []
)
@@ -69,7 +69,9 @@ class ToolFactory:
tool_type, tool_name, config
)
if map_result:
tools[tool_name] = loaded_tools # type: ignore
tools.update( # type: ignore
{tool.metadata.name: tool for tool in loaded_tools}
)
else:
tools.extend(loaded_tools) # type: ignore
@@ -0,0 +1,224 @@
import logging
import os
import uuid
from textwrap import dedent
from typing import Optional
import pandas as pd
from app.services.file import FileService
from llama_index.core import Settings
from llama_index.core.prompts import PromptTemplate
from llama_index.core.tools import FunctionTool
from pydantic import BaseModel, Field
logger = logging.getLogger(__name__)
class MissingCell(BaseModel):
"""
A missing cell in a table.
"""
row_index: int = Field(description="The index of the row of the missing cell")
column_index: int = Field(description="The index of the column of the missing cell")
question_to_answer: str = Field(
description="The question to answer to fill the missing cell"
)
class MissingCells(BaseModel):
"""
A list of missing cells.
"""
missing_cells: list[MissingCell] = Field(description="The missing cells")
class CellValue(BaseModel):
row_index: int = Field(description="The row index of the cell")
column_index: int = Field(description="The column index of the cell")
value: str = Field(
description="The value of the cell. Should be a concise value (numerical value or specific value)"
)
class FormFillingTool:
"""
Fill out missing cells in a CSV file using information from the knowledge base.
"""
save_dir: str = os.path.join("output", "tools")
# Default prompt for extracting questions
# Replace the default prompt with a custom prompt by setting the EXTRACT_QUESTIONS_PROMPT environment variable.
_default_extract_questions_prompt = dedent(
"""
You are a data analyst. You are given a table with missing cells.
Your task is to identify the missing cells and the questions needed to fill them.
IMPORTANT: Column indices should be 0-based, where the first data column is index 1
(index 0 is typically the row names/index column).
# Instructions:
- Understand the entire content of the table and the topics of the table.
- Identify the missing cells and the meaning of the data in the cells.
- For each missing cell, provide the row index and the correct column index (remember: first data column is 1).
- For each missing cell, provide the question needed to fill the cell (it's important to provide the question that is relevant to the topic of the table).
- Since the cell's value should be concise, the question should request a numerical answer or a specific value.
# Example:
# | | Name | Age | City |
# |----|------|-----|------|
# | 0 | John | | Paris|
# | 1 | Mary | | |
# | 2 | | 30 | |
#
# Your thoughts:
# - The table is about people's names, ages, and cities.
# - Row: 1, Column: 1 (Age column), Question: "How old is Mary? Please provide only the numerical answer."
# - Row: 1, Column: 2 (City column), Question: "In which city does Mary live? Please provide only the city name."
Please provide your answer in the requested format.
# Here is your task:
- Table content:
{table_content}
- Your answer:
"""
)
def extract_questions(
self,
file_path: Optional[str] = None,
file_content: Optional[str] = None,
) -> dict:
"""
Use this tool to extract missing cells in a CSV file and generate questions to fill them.
Pass either the path to the CSV file or the content of the CSV file.
Args:
file_path (Optional[str]): The local file path to the CSV file to extract missing cells from (Don't pass a sandbox path).
file_content (Optional[str]): The content of the CSV file to extract missing cells from.
Returns:
dict: A dictionary containing the missing cells and their corresponding questions.
"""
extract_questions_prompt = os.getenv(
"EXTRACT_QUESTIONS_PROMPT", self._default_extract_questions_prompt
)
if file_path is None and file_content is None:
raise ValueError("Either `file_path` or `file_content` must be provided")
table_content = None
if file_path:
file_name, file_extension = self._get_file_name_and_extension(
file_path, file_content
)
try:
df = pd.read_csv(file_path)
except FileNotFoundError as e:
return {
"error": str(e),
"message": "Please check and update the file path and ensure it's a local path - not a sandbox path.",
}
table_content = df.to_markdown()
if table_content is None:
raise ValueError("Could not convert the table to markdown")
if file_content:
table_content = file_content
if table_content is None:
raise ValueError("Table content not found")
response: MissingCells = Settings.llm.structured_predict(
output_cls=MissingCells,
prompt=PromptTemplate(extract_questions_prompt),
table_content=table_content,
)
return response.model_dump()
def fill_form(
self,
cell_values: list[CellValue],
file_path: Optional[str] = None,
file_content: Optional[str] = None,
) -> dict:
"""
Use this tool to fill cell values into a CSV file.
Requires cell values to be used for filling out, as well as either the path to the CSV file or the content of the CSV file.
Args:
cell_values (list[CellValue]): The cell values used to fill out the CSV file (call `extract_questions` and query engine to construct the cell values).
file_path (Optional[str]): The local file path to the CSV file that should be filled out (not as sandbox path).
file_content (Optional[str]): The content of the CSV file that should be filled out.
Returns:
dict: A dictionary containing the content and metadata of the filled-out file.
"""
file_name, file_extension = self._get_file_name_and_extension(
file_path, file_content
)
df = pd.read_csv(file_path)
# Fill the dataframe with the cell values
filled_df = df.copy()
for cell_value in cell_values:
if not isinstance(cell_value, CellValue):
cell_value = CellValue(**cell_value)
filled_df.iloc[cell_value.row_index, cell_value.column_index] = (
cell_value.value
)
# Save the filled table to a new CSV file
csv_content: str = filled_df.to_csv(index=False)
file_metadata = FileService.save_file(
content=csv_content,
file_name=f"{file_name}_filled.csv",
save_dir=self.save_dir,
)
new_content: str = filled_df.to_markdown()
result = {
"filled_content": new_content,
"filled_file": file_metadata,
}
return result
def _get_file_name_and_extension(
self, file_path: Optional[str], file_content: Optional[str]
) -> tuple[str, str]:
if file_path is None and file_content is None:
raise ValueError("Either `file_path` or `file_content` must be provided")
if file_path is None:
file_name = str(uuid.uuid4())
file_extension = ".csv"
else:
file_name, file_extension = os.path.splitext(file_path)
if file_extension != ".csv":
raise ValueError("Form filling is only supported for CSV files")
return file_name, file_extension
def _save_output(self, file_name: str, output: str) -> dict:
"""
Save the output to a file.
"""
file_metadata = FileService.save_file(
content=output,
file_name=file_name,
save_dir=self.save_dir,
)
return file_metadata.model_dump()
def get_tools(**kwargs):
tool = FormFillingTool()
return [
FunctionTool.from_defaults(tool.extract_questions),
FunctionTool.from_defaults(tool.fill_form),
]
@@ -1,7 +1,7 @@
import {
BaseChatEngine,
BaseToolWithCall,
OpenAIAgent,
LLMAgent,
QueryEngineTool,
} from "llamaindex";
import fs from "node:fs/promises";
@@ -42,7 +42,7 @@ export async function createChatEngine(documentIds?: string[], params?: any) {
tools.push(...(await createTools(toolConfig)));
}
const agent = new OpenAIAgent({
const agent = new LLMAgent({
tools,
systemPrompt: process.env.SYSTEM_PROMPT,
}) as unknown as BaseChatEngine;
@@ -0,0 +1,296 @@
import { JSONSchemaType } from "ajv";
import fs from "fs";
import { BaseTool, Settings, ToolMetadata } from "llamaindex";
import Papa from "papaparse";
import path from "path";
import { saveDocument } from "../../llamaindex/documents/helper";
type ExtractMissingCellsParameter = {
filePath: string;
};
export type MissingCell = {
rowIndex: number;
columnIndex: number;
question: string;
};
const CSV_EXTRACTION_PROMPT = `You are a data analyst. You are given a table with missing cells.
Your task is to identify the missing cells and the questions needed to fill them.
IMPORTANT: Column indices should be 0-based
# Instructions:
- Understand the entire content of the table and the topics of the table.
- Identify the missing cells and the meaning of the data in the cells.
- For each missing cell, provide the row index and the correct column index (remember: first data column is 1).
- For each missing cell, provide the question needed to fill the cell (it's important to provide the question that is relevant to the topic of the table).
- Since the cell's value should be concise, the question should request a numerical answer or a specific value.
- Finally, only return the answer in JSON format with the following schema:
{
"missing_cells": [
{
"rowIndex": number,
"columnIndex": number,
"question": string
}
]
}
- If there are no missing cells, return an empty array.
- The answer is only the JSON object, nothing else and don't wrap it inside markdown code block.
# Example:
# | | Name | Age | City |
# |----|------|-----|------|
# | 0 | John | | Paris|
# | 1 | Mary | | |
# | 2 | | 30 | |
#
# Your thoughts:
# - The table is about people's names, ages, and cities.
# - Row: 1, Column: 2 (Age column), Question: "How old is Mary? Please provide only the numerical answer."
# - Row: 1, Column: 3 (City column), Question: "In which city does Mary live? Please provide only the city name."
# Your answer:
# {
# "missing_cells": [
# {
# "rowIndex": 1,
# "columnIndex": 2,
# "question": "How old is Mary? Please provide only the numerical answer."
# },
# {
# "rowIndex": 1,
# "columnIndex": 3,
# "question": "In which city does Mary live? Please provide only the city name."
# }
# ]
# }
# Here is your task:
- Table content:
{table_content}
- Your answer:
`;
const DEFAULT_METADATA: ToolMetadata<
JSONSchemaType<ExtractMissingCellsParameter>
> = {
name: "extract_missing_cells",
description: `Use this tool to extract missing cells in a CSV file and generate questions to fill them. This tool only works with local file path.`,
parameters: {
type: "object",
properties: {
filePath: {
type: "string",
description: "The local file path to the CSV file.",
},
},
required: ["filePath"],
},
};
export interface ExtractMissingCellsParams {
metadata?: ToolMetadata<JSONSchemaType<ExtractMissingCellsParameter>>;
}
export class ExtractMissingCellsTool
implements BaseTool<ExtractMissingCellsParameter>
{
metadata: ToolMetadata<JSONSchemaType<ExtractMissingCellsParameter>>;
defaultExtractionPrompt: string;
constructor(params: ExtractMissingCellsParams) {
this.metadata = params.metadata ?? DEFAULT_METADATA;
this.defaultExtractionPrompt = CSV_EXTRACTION_PROMPT;
}
private readCsvFile(filePath: string): Promise<string[][]> {
return new Promise((resolve, reject) => {
fs.readFile(filePath, "utf8", (err, data) => {
if (err) {
reject(err);
return;
}
const parsedData = Papa.parse<string[]>(data, {
skipEmptyLines: false,
});
if (parsedData.errors.length) {
reject(parsedData.errors);
return;
}
// Ensure all rows have the same number of columns as the header
const maxColumns = parsedData.data[0].length;
const paddedRows = parsedData.data.map((row) => {
return [...row, ...Array(maxColumns - row.length).fill("")];
});
resolve(paddedRows);
});
});
}
private formatToMarkdownTable(data: string[][]): string {
if (data.length === 0) return "";
const maxColumns = data[0].length;
const headerRow = `| ${data[0].join(" | ")} |`;
const separatorRow = `| ${Array(maxColumns).fill("---").join(" | ")} |`;
const dataRows = data.slice(1).map((row) => {
return `| ${row.join(" | ")} |`;
});
return [headerRow, separatorRow, ...dataRows].join("\n");
}
async call(input: ExtractMissingCellsParameter): Promise<MissingCell[]> {
const { filePath } = input;
let tableContent: string[][];
try {
tableContent = await this.readCsvFile(filePath);
} catch (error) {
throw new Error(
`Failed to read CSV file. Make sure that you are reading a local file path (not a sandbox path).`,
);
}
const prompt = this.defaultExtractionPrompt.replace(
"{table_content}",
this.formatToMarkdownTable(tableContent),
);
const llm = Settings.llm;
const response = await llm.complete({
prompt,
});
const rawAnswer = response.text;
const parsedResponse = JSON.parse(rawAnswer) as {
missing_cells: MissingCell[];
};
if (!parsedResponse.missing_cells) {
throw new Error(
"The answer is not in the correct format. There should be a missing_cells array.",
);
}
const answer = parsedResponse.missing_cells;
return answer;
}
}
type FillMissingCellsParameter = {
filePath: string;
cells: {
rowIndex: number;
columnIndex: number;
answer: string;
}[];
};
const FILL_CELLS_METADATA: ToolMetadata<
JSONSchemaType<FillMissingCellsParameter>
> = {
name: "fill_missing_cells",
description: `Use this tool to fill missing cells in a CSV file with provided answers. This tool only works with local file path.`,
parameters: {
type: "object",
properties: {
filePath: {
type: "string",
description: "The local file path to the CSV file.",
},
cells: {
type: "array",
items: {
type: "object",
properties: {
rowIndex: { type: "number" },
columnIndex: { type: "number" },
answer: { type: "string" },
},
required: ["rowIndex", "columnIndex", "answer"],
},
description: "Array of cells to fill with their answers",
},
},
required: ["filePath", "cells"],
},
};
export interface FillMissingCellsParams {
metadata?: ToolMetadata<JSONSchemaType<FillMissingCellsParameter>>;
}
export class FillMissingCellsTool
implements BaseTool<FillMissingCellsParameter>
{
metadata: ToolMetadata<JSONSchemaType<FillMissingCellsParameter>>;
constructor(params: FillMissingCellsParams = {}) {
this.metadata = params.metadata ?? FILL_CELLS_METADATA;
}
async call(input: FillMissingCellsParameter): Promise<string> {
const { filePath, cells } = input;
// Read the CSV file
const fileContent = await new Promise<string>((resolve, reject) => {
fs.readFile(filePath, "utf8", (err, data) => {
if (err) {
reject(err);
} else {
resolve(data);
}
});
});
// Parse CSV with PapaParse
const parseResult = Papa.parse<string[]>(fileContent, {
header: false, // Ensure the header is not treated as a separate object
skipEmptyLines: false, // Ensure empty lines are not skipped
});
if (parseResult.errors.length) {
throw new Error(
"Failed to parse CSV file: " + parseResult.errors[0].message,
);
}
const rows = parseResult.data;
// Fill the cells with answers
for (const cell of cells) {
// Adjust rowIndex to start from 1 for data rows
const adjustedRowIndex = cell.rowIndex + 1;
if (
adjustedRowIndex < rows.length &&
cell.columnIndex < rows[adjustedRowIndex].length
) {
rows[adjustedRowIndex][cell.columnIndex] = cell.answer;
}
}
// Convert back to CSV format
const updatedContent = Papa.unparse(rows, {
delimiter: parseResult.meta.delimiter,
});
// Use the helper function to write the file
const parsedPath = path.parse(filePath);
const newFileName = `${parsedPath.name}-filled${parsedPath.ext}`;
const newFilePath = path.join("output/tools", newFileName);
const newFileUrl = await saveDocument(newFilePath, updatedContent);
return (
"Successfully filled missing cells in the CSV file. File URL to show to the user: " +
newFileUrl
);
}
}
@@ -1,11 +1,19 @@
import { BaseToolWithCall } from "llamaindex";
import { ToolsFactory } from "llamaindex/tools/ToolsFactory";
import fs from "node:fs/promises";
import path from "node:path";
import { CodeGeneratorTool, CodeGeneratorToolParams } from "./code-generator";
import {
DocumentGenerator,
DocumentGeneratorParams,
} from "./document-generator";
import { DuckDuckGoSearchTool, DuckDuckGoToolParams } from "./duckduckgo";
import {
ExtractMissingCellsParams,
ExtractMissingCellsTool,
FillMissingCellsParams,
FillMissingCellsTool,
} from "./form-filling";
import { ImgGeneratorTool, ImgGeneratorToolParams } from "./img-gen";
import { InterpreterTool, InterpreterToolParams } from "./interpreter";
import { OpenAPIActionTool } from "./openapi-action";
@@ -54,6 +62,12 @@ const toolFactory: Record<string, ToolCreator> = {
document_generator: async (config: unknown) => {
return [new DocumentGenerator(config as DocumentGeneratorParams)];
},
form_filling: async (config: unknown) => {
return [
new ExtractMissingCellsTool(config as ExtractMissingCellsParams),
new FillMissingCellsTool(config as FillMissingCellsParams),
];
},
};
async function createLocalTools(
@@ -70,3 +84,19 @@ async function createLocalTools(
return tools;
}
export async function getConfiguredTools(
configPath?: string,
): Promise<BaseToolWithCall[]> {
const configFile = path.join(configPath ?? "config", "tools.json");
const toolConfig = JSON.parse(await fs.readFile(configFile, "utf8"));
const tools = await createTools(toolConfig);
return tools;
}
export async function getTool(
toolName: string,
): Promise<BaseToolWithCall | undefined> {
const tools = await getConfiguredTools();
return tools.find((tool) => tool.metadata.name === toolName);
}
@@ -13,7 +13,7 @@ const MIME_TYPE_TO_EXT: Record<string, string> = {
"docx",
};
const UPLOADED_FOLDER = "output/uploaded";
export const UPLOADED_FOLDER = "output/uploaded";
export async function storeAndParseFile(
name: string,
@@ -1,7 +1,5 @@
import { Document, LLamaCloudFileService, VectorStoreIndex } from "llamaindex";
import { LlamaCloudIndex } from "llamaindex/cloud/LlamaCloudIndex";
import fs from "node:fs/promises";
import path from "node:path";
import { DocumentFile } from "../streaming/annotations";
import { parseFile, storeFile } from "./helper";
import { runPipeline } from "./pipeline";
@@ -18,8 +16,8 @@ export async function uploadDocument(
// Store file
const fileMetadata = await storeFile(name, fileBuffer, mimeType);
// If the file is csv and has codeExecutorTool, we don't need to index the file.
if (mimeType === "text/csv" && (await hasCodeExecutorTool())) {
// Do not index csv files
if (mimeType === "text/csv") {
return fileMetadata;
}
let documentIds: string[] = [];
@@ -49,7 +47,11 @@ export async function uploadDocument(
}
} else {
// run the pipeline for other vector store indexes
const documents: Document[] = await parseFile(fileBuffer, name, mimeType);
const documents: Document[] = await parseFile(
fileBuffer,
fileMetadata.name,
mimeType,
);
documentIds = await runPipeline(index, documents);
}
@@ -57,14 +59,3 @@ export async function uploadDocument(
fileMetadata.refs = documentIds;
return fileMetadata;
}
const hasCodeExecutorTool = async () => {
const codeExecutorTools = ["interpreter", "artifact"];
const configFile = path.join("config", "tools.json");
const toolConfig = JSON.parse(await fs.readFile(configFile, "utf8"));
const localTools = toolConfig.local || {};
// Check if local tools contains codeExecutorTools
return codeExecutorTools.some((tool) => localTools[tool] !== undefined);
};
@@ -1,5 +1,11 @@
import { JSONValue, Message } from "ai";
import { MessageContent, MessageContentDetail } from "llamaindex";
import {
ChatMessage,
MessageContent,
MessageContentDetail,
MessageType,
} from "llamaindex";
import { UPLOADED_FOLDER } from "../documents/helper";
export type DocumentFileType = "csv" | "pdf" | "txt" | "docx";
@@ -58,6 +64,45 @@ export function retrieveMessageContent(messages: Message[]): MessageContent {
];
}
export function convertToChatHistory(messages: Message[]): ChatMessage[] {
if (!messages || !Array.isArray(messages)) {
return [];
}
const agentHistory = retrieveAgentHistoryMessage(messages);
if (agentHistory) {
const previousMessages = messages.slice(0, -1);
return [...previousMessages, agentHistory].map((msg) => ({
role: msg.role as MessageType,
content: msg.content,
}));
}
return messages.map((msg) => ({
role: msg.role as MessageType,
content: msg.content,
}));
}
function retrieveAgentHistoryMessage(
messages: Message[],
maxAgentMessages = 10,
): ChatMessage | null {
const agentAnnotations = getAnnotations<{ agent: string; text: string }>(
messages,
{ role: "assistant", type: "agent" },
).slice(-maxAgentMessages);
if (agentAnnotations.length > 0) {
const messageContent =
"Here is the previous conversation of agents:\n" +
agentAnnotations.map((annotation) => annotation.data.text).join("\n");
return {
role: "assistant",
content: messageContent,
};
}
return null;
}
function getFileContent(file: DocumentFile): string {
let defaultContent = `=====File: ${file.name}=====\n`;
// Include file URL if it's available
@@ -84,6 +129,10 @@ function getFileContent(file: DocumentFile): string {
const sandboxFilePath = `/tmp/${file.name}`;
defaultContent += `Sandbox file path (instruction: only use sandbox path for artifact or code interpreter tool): ${sandboxFilePath}\n`;
// Include local file path
const localFilePath = `${UPLOADED_FOLDER}/${file.name}`;
defaultContent += `Local file path (instruction: use for local tool that requires a local path): ${localFilePath}\n`;
return defaultContent;
}
@@ -127,13 +176,10 @@ function retrieveLatestArtifact(messages: Message[]): MessageContentDetail[] {
}
function convertAnnotations(messages: Message[]): MessageContentDetail[] {
// annotations from the last user message that has annotations
const annotations: Annotation[] =
messages
.slice()
.reverse()
.find((message) => message.role === "user" && message.annotations)
?.annotations?.map(getValidAnnotation) || [];
// get all annotations from user messages
const annotations: Annotation[] = messages
.filter((message) => message.role === "user" && message.annotations)
.flatMap((message) => message.annotations?.map(getValidAnnotation) || []);
if (annotations.length === 0) return [];
const content: MessageContentDetail[] = [];
@@ -1,7 +1,7 @@
import {
FILE_EXT_TO_READER,
SimpleDirectoryReader,
} from "llamaindex/readers/SimpleDirectoryReader";
} from "llamaindex/readers/index";
export const DATA_DIR = "./data";
@@ -2,7 +2,7 @@ import { LlamaParseReader } from "llamaindex";
import {
FILE_EXT_TO_READER,
SimpleDirectoryReader,
} from "llamaindex/readers/SimpleDirectoryReader";
} from "llamaindex/readers/index";
export const DATA_DIR = "./data";
@@ -4,7 +4,8 @@ from app.api.routers.models import (
ChatData,
)
from app.api.routers.vercel_response import VercelStreamResponse
from app.engine.engine import get_chat_engine
from app.engine.query_filter import generate_filters
from app.workflows import create_workflow
from fastapi import APIRouter, BackgroundTasks, HTTPException, Request, status
chat_router = r = APIRouter()
@@ -22,19 +23,20 @@ async def chat(
last_message_content = data.get_last_message_content()
messages = data.get_history_messages(include_agent_messages=True)
# The chat API supports passing private document filters and chat params
# but agent workflow does not support them yet
# ignore chat params and use all documents for now
# TODO: generate filters based on doc_ids
doc_ids = data.get_chat_document_ids()
filters = generate_filters(doc_ids)
params = data.data or {}
engine = get_chat_engine(chat_history=messages, params=params)
event_handler = engine.run(input=last_message_content, streaming=True)
workflow = create_workflow(
chat_history=messages, params=params, filters=filters
)
event_handler = workflow.run(input=last_message_content, streaming=True)
return VercelStreamResponse(
request=request,
chat_data=data,
event_handler=event_handler,
events=engine.stream_events(),
events=workflow.stream_events(),
)
except Exception as e:
logger.exception("Error in chat engine", exc_info=True)
@@ -1,12 +1,11 @@
import asyncio
import json
import logging
from typing import AsyncGenerator, List
from typing import AsyncGenerator, Awaitable, List
from aiostream import stream
from app.api.routers.models import ChatData, Message
from app.api.services.suggestion import NextQuestionSuggestion
from app.workflows.single import AgentRunEvent, AgentRunResult
from fastapi import Request
from fastapi.responses import StreamingResponse
@@ -55,8 +54,8 @@ class VercelStreamResponse(StreamingResponse):
self,
request: Request,
chat_data: ChatData,
event_handler: AgentRunResult | AsyncGenerator,
events: AsyncGenerator[AgentRunEvent, None],
event_handler: Awaitable,
events: AsyncGenerator,
verbose: bool = True,
):
# Yield the text response
@@ -64,15 +63,17 @@ class VercelStreamResponse(StreamingResponse):
result = await event_handler
final_response = ""
if isinstance(result, AgentRunResult):
for token in result.response.message.content:
final_response += token
yield self.convert_text(token)
if isinstance(result, AsyncGenerator):
async for token in result:
final_response += token.delta
final_response += str(token.delta)
yield self.convert_text(token.delta)
else:
if hasattr(result, "response"):
content = result.response.message.content
if content:
for token in content:
final_response += str(token)
yield self.convert_text(token)
# Generate next questions if next question prompt is configured
question_data = await self._generate_next_questions(
@@ -86,7 +87,7 @@ class VercelStreamResponse(StreamingResponse):
# Yield the events from the event handler
async def _event_generator():
async for event in events:
event_response = self._event_to_response(event)
event_response = event.to_response()
if verbose:
logger.debug(event_response)
if event_response is not None:
@@ -95,13 +96,6 @@ class VercelStreamResponse(StreamingResponse):
combine = stream.merge(_chat_response_generator(), _event_generator())
return combine
@staticmethod
def _event_to_response(event: AgentRunEvent) -> dict:
return {
"type": "agent",
"data": {"agent": event.name, "text": event.msg},
}
@classmethod
def convert_text(cls, token: str):
# Escape newlines and double quotes to avoid breaking the stream
@@ -0,0 +1,27 @@
from enum import Enum
from typing import Optional
from llama_index.core.workflow import Event
class AgentRunEventType(Enum):
TEXT = "text"
PROGRESS = "progress"
class AgentRunEvent(Event):
name: str
msg: str
event_type: AgentRunEventType = AgentRunEventType.TEXT
data: Optional[dict] = None
def to_response(self) -> dict:
return {
"type": "agent",
"data": {
"agent": self.name,
"type": self.event_type.value,
"text": self.msg,
"data": self.data,
},
}
@@ -0,0 +1,121 @@
from typing import Any, List, Optional
from app.workflows.events import AgentRunEvent
from app.workflows.tools import ToolCallResponse, call_tools, chat_with_tools
from llama_index.core.base.llms.types import ChatMessage
from llama_index.core.llms.function_calling import FunctionCallingLLM
from llama_index.core.memory import ChatMemoryBuffer
from llama_index.core.settings import Settings
from llama_index.core.tools.types import BaseTool
from llama_index.core.workflow import (
Context,
Event,
StartEvent,
StopEvent,
Workflow,
step,
)
class InputEvent(Event):
input: list[ChatMessage]
class ToolCallEvent(Event):
input: ToolCallResponse
class FunctionCallingAgent(Workflow):
"""
A simple workflow to request LLM with tools independently.
You can share the previous chat history to provide the context for the LLM.
"""
def __init__(
self,
*args: Any,
llm: FunctionCallingLLM | None = None,
chat_history: Optional[List[ChatMessage]] = None,
tools: List[BaseTool] | None = None,
system_prompt: str | None = None,
verbose: bool = False,
timeout: float = 360.0,
name: str,
write_events: bool = True,
**kwargs: Any,
) -> None:
super().__init__(*args, verbose=verbose, timeout=timeout, **kwargs) # type: ignore
self.tools = tools or []
self.name = name
self.write_events = write_events
if llm is None:
llm = Settings.llm
self.llm = llm
if not self.llm.metadata.is_function_calling_model:
raise ValueError("The provided LLM must support function calling.")
self.system_prompt = system_prompt
self.memory = ChatMemoryBuffer.from_defaults(
llm=self.llm, chat_history=chat_history
)
self.sources = [] # type: ignore
@step()
async def prepare_chat_history(self, ctx: Context, ev: StartEvent) -> InputEvent:
# clear sources
self.sources = []
# set streaming
ctx.data["streaming"] = getattr(ev, "streaming", False)
# set system prompt
if self.system_prompt is not None:
system_msg = ChatMessage(role="system", content=self.system_prompt)
self.memory.put(system_msg)
# get user input
user_input = ev.input
user_msg = ChatMessage(role="user", content=user_input)
self.memory.put(user_msg)
if self.write_events:
ctx.write_event_to_stream(
AgentRunEvent(name=self.name, msg=f"Start to work on: {user_input}")
)
return InputEvent(input=self.memory.get())
@step()
async def handle_llm_input(
self,
ctx: Context,
ev: InputEvent,
) -> ToolCallEvent | StopEvent:
chat_history = ev.input
response = await chat_with_tools(
self.llm,
self.tools,
chat_history,
)
is_tool_call = isinstance(response, ToolCallResponse)
if not is_tool_call:
if ctx.data["streaming"]:
return StopEvent(result=response)
else:
full_response = ""
async for chunk in response.generator:
full_response += chunk.message.content
return StopEvent(result=full_response)
return ToolCallEvent(input=response)
@step()
async def handle_tool_calls(self, ctx: Context, ev: ToolCallEvent) -> InputEvent:
tool_calls = ev.input.tool_calls
tool_call_message = ev.input.tool_call_message
self.memory.put(tool_call_message)
tool_messages = await call_tools(self.name, self.tools, ctx, tool_calls)
self.memory.put_messages(tool_messages)
return InputEvent(input=self.memory.get())
@@ -0,0 +1,227 @@
import logging
import uuid
from abc import ABC, abstractmethod
from typing import Any, AsyncGenerator, Callable, Optional
from app.workflows.events import AgentRunEvent, AgentRunEventType
from llama_index.core.base.llms.types import ChatMessage, ChatResponse, MessageRole
from llama_index.core.llms.function_calling import FunctionCallingLLM
from llama_index.core.tools import (
BaseTool,
FunctionTool,
ToolOutput,
ToolSelection,
)
from llama_index.core.workflow import Context
from pydantic import BaseModel, ConfigDict
logger = logging.getLogger("uvicorn")
class ContextAwareTool(FunctionTool, ABC):
@abstractmethod
async def acall(self, ctx: Context, input: Any) -> ToolOutput: # type: ignore
pass
class ChatWithToolsResponse(BaseModel):
"""
A tool call response from chat_with_tools.
"""
tool_calls: Optional[list[ToolSelection]]
tool_call_message: Optional[ChatMessage]
generator: Optional[AsyncGenerator[ChatResponse | None, None]]
model_config = ConfigDict(arbitrary_types_allowed=True)
def is_calling_different_tools(self) -> bool:
tool_names = {tool_call.tool_name for tool_call in self.tool_calls}
return len(tool_names) > 1
def has_tool_calls(self) -> bool:
return self.tool_calls is not None and len(self.tool_calls) > 0
def tool_name(self) -> str:
assert self.has_tool_calls()
assert not self.is_calling_different_tools()
return self.tool_calls[0].tool_name
async def full_response(self) -> str:
assert self.generator is not None
full_response = ""
async for chunk in self.generator:
full_response += chunk.message.content
return full_response
async def chat_with_tools( # type: ignore
llm: FunctionCallingLLM,
tools: list[BaseTool],
chat_history: list[ChatMessage],
) -> ChatWithToolsResponse:
"""
Request LLM to call tools or not.
This function doesn't change the memory.
"""
generator = _tool_call_generator(llm, tools, chat_history)
is_tool_call = await generator.__anext__()
if is_tool_call:
# Last chunk is the full response
# Wait for the last chunk
full_response = None
async for chunk in generator:
full_response = chunk
assert isinstance(full_response, ChatResponse)
return ChatWithToolsResponse(
tool_calls=llm.get_tool_calls_from_response(full_response),
tool_call_message=full_response.message,
generator=None,
)
else:
return ChatWithToolsResponse(
tool_calls=None,
tool_call_message=None,
generator=generator,
)
async def call_tools(
ctx: Context,
agent_name: str,
tools: list[BaseTool],
tool_calls: list[ToolSelection],
emit_agent_events: bool = True,
) -> list[ChatMessage]:
if len(tool_calls) == 0:
return []
tools_by_name = {tool.metadata.get_name(): tool for tool in tools}
if len(tool_calls) == 1:
return [
await call_tool(
ctx,
tools_by_name[tool_calls[0].tool_name],
tool_calls[0],
lambda msg: ctx.write_event_to_stream(
AgentRunEvent(
name=agent_name,
msg=msg,
)
),
)
]
# Multiple tool calls, show progress
tool_msgs: list[ChatMessage] = []
progress_id = str(uuid.uuid4())
total_steps = len(tool_calls)
if emit_agent_events:
ctx.write_event_to_stream(
AgentRunEvent(
name=agent_name,
msg=f"Making {total_steps} tool calls",
)
)
for i, tool_call in enumerate(tool_calls):
tool = tools_by_name.get(tool_call.tool_name)
if not tool:
tool_msgs.append(
ChatMessage(
role=MessageRole.ASSISTANT,
content=f"Tool {tool_call.tool_name} does not exist",
)
)
continue
tool_msg = await call_tool(
ctx,
tool,
tool_call,
event_emitter=lambda msg: ctx.write_event_to_stream(
AgentRunEvent(
name=agent_name,
msg=msg,
event_type=AgentRunEventType.PROGRESS,
data={
"id": progress_id,
"total": total_steps,
"current": i,
},
)
),
)
tool_msgs.append(tool_msg)
return tool_msgs
async def call_tool(
ctx: Context,
tool: BaseTool,
tool_call: ToolSelection,
event_emitter: Optional[Callable[[str], None]],
) -> ChatMessage:
if event_emitter:
event_emitter(
f"Calling tool {tool_call.tool_name}, {str(tool_call.tool_kwargs)}"
)
try:
if isinstance(tool, ContextAwareTool):
if ctx is None:
raise ValueError("Context is required for context aware tool")
# inject context for calling an context aware tool
response = await tool.acall(ctx=ctx, **tool_call.tool_kwargs)
else:
response = await tool.acall(**tool_call.tool_kwargs) # type: ignore
return ChatMessage(
role=MessageRole.TOOL,
content=str(response.raw_output),
additional_kwargs={
"tool_call_id": tool_call.tool_id,
"name": tool.metadata.get_name(),
},
)
except Exception as e:
logger.error(f"Got error in tool {tool_call.tool_name}: {str(e)}")
if event_emitter:
event_emitter(f"Got error in tool {tool_call.tool_name}: {str(e)}")
return ChatMessage(
role=MessageRole.TOOL,
content=f"Error: {str(e)}",
additional_kwargs={
"tool_call_id": tool_call.tool_id,
"name": tool.metadata.get_name(),
},
)
async def _tool_call_generator(
llm: FunctionCallingLLM,
tools: list[BaseTool],
chat_history: list[ChatMessage],
) -> AsyncGenerator[ChatResponse | bool, None]:
response_stream = await llm.astream_chat_with_tools(
tools,
chat_history=chat_history,
allow_parallel_tool_calls=False,
)
full_response = None
yielded_indicator = False
async for chunk in response_stream:
if "tool_calls" not in chunk.message.additional_kwargs:
# Yield a boolean to indicate whether the response is a tool call
if not yielded_indicator:
yield False
yielded_indicator = True
# if not a tool call, yield the chunks!
yield chunk # type: ignore
elif not yielded_indicator:
# Yield the indicator for a tool call
yield True
yielded_indicator = True
full_response = chunk
if full_response:
yield full_response # type: ignore
@@ -1,36 +1,34 @@
import { StopEvent } from "@llamaindex/core/workflow";
import { Message, streamToResponse } from "ai";
import { Request, Response } from "express";
import { ChatResponseChunk } from "llamaindex";
import {
convertToChatHistory,
retrieveMessageContent,
} from "./llamaindex/streaming/annotations";
import { createWorkflow } from "./workflow/factory";
import { toDataStream, workflowEventsToStreamData } from "./workflow/stream";
import { createStreamFromWorkflowContext } from "./workflow/stream";
export const chat = async (req: Request, res: Response) => {
try {
const { messages, data }: { messages: Message[]; data?: any } = req.body;
const userMessage = messages.pop();
if (!messages || !userMessage || userMessage.role !== "user") {
const { messages }: { messages: Message[] } = req.body;
if (!messages || messages.length === 0) {
return res.status(400).json({
error:
"messages are required in the request body and the last message must be from the user",
error: "messages are required in the request body",
});
}
const chatHistory = convertToChatHistory(messages);
const userMessageContent = retrieveMessageContent(messages);
const agent = createWorkflow(messages, data);
const result = agent.run<AsyncGenerator<ChatResponseChunk>>(
userMessage.content,
) as unknown as Promise<StopEvent<AsyncGenerator<ChatResponseChunk>>>;
const workflow = await createWorkflow({ chatHistory });
// convert the workflow events to a vercel AI stream data object
const agentStreamData = await workflowEventsToStreamData(
agent.streamEvents(),
);
// convert the workflow result to a vercel AI content stream
const stream = toDataStream(result, {
onFinal: () => agentStreamData.close(),
const context = workflow.run({
message: userMessageContent,
streaming: true,
});
return streamToResponse(stream, res, {}, agentStreamData);
const { stream, dataStream } =
await createStreamFromWorkflowContext(context);
return streamToResponse(stream, res, {}, dataStream);
} catch (error) {
console.error("[LlamaIndex]", error);
return res.status(500).json({
@@ -1,11 +1,14 @@
import { initObservability } from "@/app/observability";
import { StopEvent } from "@llamaindex/core/workflow";
import { Message, StreamingTextResponse } from "ai";
import { ChatResponseChunk } from "llamaindex";
import { StreamingTextResponse, type Message } from "ai";
import { NextRequest, NextResponse } from "next/server";
import { initSettings } from "./engine/settings";
import {
convertToChatHistory,
isValidMessages,
retrieveMessageContent,
} from "./llamaindex/streaming/annotations";
import { createWorkflow } from "./workflow/factory";
import { toDataStream, workflowEventsToStreamData } from "./workflow/stream";
import { createStreamFromWorkflowContext } from "./workflow/stream";
initObservability();
initSettings();
@@ -16,9 +19,8 @@ export const dynamic = "force-dynamic";
export async function POST(request: NextRequest) {
try {
const body = await request.json();
const { messages, data }: { messages: Message[]; data?: any } = body;
const userMessage = messages.pop();
if (!messages || !userMessage || userMessage.role !== "user") {
const { messages }: { messages: Message[]; data?: any } = body;
if (!isValidMessages(messages)) {
return NextResponse.json(
{
error:
@@ -28,20 +30,20 @@ export async function POST(request: NextRequest) {
);
}
const agent = createWorkflow(messages, data);
// TODO: fix type in agent.run in LITS
const result = agent.run<AsyncGenerator<ChatResponseChunk>>(
userMessage.content,
) as unknown as Promise<StopEvent<AsyncGenerator<ChatResponseChunk>>>;
// convert the workflow events to a vercel AI stream data object
const agentStreamData = await workflowEventsToStreamData(
agent.streamEvents(),
);
// convert the workflow result to a vercel AI content stream
const stream = toDataStream(result, {
onFinal: () => agentStreamData.close(),
const chatHistory = convertToChatHistory(messages);
const userMessageContent = retrieveMessageContent(messages);
const workflow = await createWorkflow({ chatHistory });
const context = workflow.run({
message: userMessageContent,
streaming: true,
});
return new StreamingTextResponse(stream, {}, agentStreamData);
const { stream, dataStream } =
await createStreamFromWorkflowContext(context);
// Return the two streams in one response
return new StreamingTextResponse(stream, {}, dataStream);
} catch (error) {
console.error("[LlamaIndex]", error);
return NextResponse.json(
@@ -1,22 +1,21 @@
import {
Context,
HandlerContext,
StartEvent,
StopEvent,
Workflow,
WorkflowEvent,
} from "@llamaindex/core/workflow";
} from "@llamaindex/workflow";
import {
BaseToolWithCall,
ChatMemoryBuffer,
ChatMessage,
ChatResponse,
ChatResponseChunk,
QueryEngineTool,
Settings,
ToolCall,
ToolCallLLM,
ToolCallLLMMessageOptions,
callTool,
} from "llamaindex";
import { callTools, chatWithTools } from "./tools";
import { AgentInput, AgentRunEvent } from "./type";
class InputEvent extends WorkflowEvent<{
@@ -27,11 +26,23 @@ class ToolCallEvent extends WorkflowEvent<{
toolCalls: ToolCall[];
}> {}
export class FunctionCallingAgent extends Workflow {
type FunctionCallingAgentContextData = {
streaming: boolean;
};
export type FunctionCallingAgentInput = AgentInput & {
displayName: string;
};
export class FunctionCallingAgent extends Workflow<
FunctionCallingAgentContextData,
FunctionCallingAgentInput,
string | AsyncGenerator<boolean | ChatResponseChunk<object>>
> {
name: string;
llm: ToolCallLLM;
memory: ChatMemoryBuffer;
tools: BaseToolWithCall[];
tools: BaseToolWithCall[] | QueryEngineTool[];
systemPrompt?: string;
writeEvents: boolean;
role?: string;
@@ -53,7 +64,9 @@ export class FunctionCallingAgent extends Workflow {
});
this.name = options?.name;
this.llm = options.llm ?? (Settings.llm as ToolCallLLM);
this.checkToolCallSupport();
if (!(this.llm instanceof ToolCallLLM)) {
throw new Error("LLM is not a ToolCallLLM");
}
this.memory = new ChatMemoryBuffer({
llm: this.llm,
chatHistory: options.chatHistory,
@@ -64,175 +77,103 @@ export class FunctionCallingAgent extends Workflow {
this.role = options?.role;
// add steps
this.addStep(StartEvent<AgentInput>, this.prepareChatHistory, {
outputs: InputEvent,
});
this.addStep(InputEvent, this.handleLLMInput, {
outputs: [ToolCallEvent, StopEvent],
});
this.addStep(ToolCallEvent, this.handleToolCalls, {
outputs: InputEvent,
});
this.addStep(
{
inputs: [StartEvent<AgentInput>],
outputs: [InputEvent],
},
this.prepareChatHistory,
);
this.addStep(
{
inputs: [InputEvent],
outputs: [ToolCallEvent, StopEvent],
},
this.handleLLMInput,
);
this.addStep(
{
inputs: [ToolCallEvent],
outputs: [InputEvent],
},
this.handleToolCalls,
);
}
private get chatHistory() {
return this.memory.getMessages();
}
private async prepareChatHistory(
ctx: Context,
prepareChatHistory = async (
ctx: HandlerContext<FunctionCallingAgentContextData>,
ev: StartEvent<AgentInput>,
): Promise<InputEvent> {
const { message, streaming } = ev.data.input;
ctx.set("streaming", streaming);
): Promise<InputEvent> => {
const { message, streaming } = ev.data;
ctx.data.streaming = streaming ?? false;
this.writeEvent(`Start to work on: ${message}`, ctx);
if (this.systemPrompt) {
this.memory.put({ role: "system", content: this.systemPrompt });
}
this.memory.put({ role: "user", content: message });
return new InputEvent({ input: this.chatHistory });
}
};
private async handleLLMInput(
ctx: Context,
handleLLMInput = async (
ctx: HandlerContext<FunctionCallingAgentContextData>,
ev: InputEvent,
): Promise<StopEvent<string | AsyncGenerator> | ToolCallEvent> {
if (ctx.get("streaming")) {
return await this.handleLLMInputStream(ctx, ev);
): Promise<StopEvent<string | AsyncGenerator> | ToolCallEvent> => {
const toolCallResponse = await chatWithTools(
this.llm,
this.tools,
this.chatHistory,
);
if (toolCallResponse.toolCallMessage) {
this.memory.put(toolCallResponse.toolCallMessage);
}
const result = await this.llm.chat({
messages: this.chatHistory,
tools: this.tools,
});
this.memory.put(result.message);
const toolCalls = this.getToolCallsFromResponse(result);
if (toolCalls.length) {
return new ToolCallEvent({ toolCalls });
if (toolCallResponse.hasToolCall()) {
return new ToolCallEvent({ toolCalls: toolCallResponse.toolCalls });
}
this.writeEvent("Finished task", ctx);
return new StopEvent({ result: result.message.content.toString() });
}
private async handleLLMInputStream(
context: Context,
ev: InputEvent,
): Promise<StopEvent<AsyncGenerator> | ToolCallEvent> {
const { llm, tools, memory } = this;
const llmArgs = { messages: this.chatHistory, tools };
const responseGenerator = async function* () {
const responseStream = await llm.chat({ ...llmArgs, stream: true });
let fullResponse = null;
let yieldedIndicator = false;
for await (const chunk of responseStream) {
const hasToolCalls = chunk.options && "toolCall" in chunk.options;
if (!hasToolCalls) {
if (!yieldedIndicator) {
yield false;
yieldedIndicator = true;
}
yield chunk;
} else if (!yieldedIndicator) {
yield true;
yieldedIndicator = true;
}
fullResponse = chunk;
if (ctx.data.streaming) {
if (!toolCallResponse.responseGenerator) {
throw new Error("No streaming response");
}
if (fullResponse?.options && Object.keys(fullResponse.options).length) {
memory.put({
role: "assistant",
content: "",
options: fullResponse.options,
});
yield fullResponse;
}
};
const generator = responseGenerator();
const isToolCall = await generator.next();
if (isToolCall.value) {
const fullResponse = await generator.next();
const toolCalls = this.getToolCallsFromResponse(
fullResponse.value as ChatResponseChunk<ToolCallLLMMessageOptions>,
);
return new ToolCallEvent({ toolCalls });
return new StopEvent(toolCallResponse.responseGenerator);
}
this.writeEvent("Finished task", context);
return new StopEvent({ result: generator });
}
const fullResponse = await toolCallResponse.asFullResponse();
this.memory.put(fullResponse);
return new StopEvent(fullResponse.content.toString());
};
private async handleToolCalls(
ctx: Context,
handleToolCalls = async (
ctx: HandlerContext<FunctionCallingAgentContextData>,
ev: ToolCallEvent,
): Promise<InputEvent> {
): Promise<InputEvent> => {
const { toolCalls } = ev.data;
const toolMsgs: ChatMessage[] = [];
for (const call of toolCalls) {
const targetTool = this.tools.find(
(tool) => tool.metadata.name === call.name,
);
// TODO: make logger optional in callTool in framework
const toolOutput = await callTool(targetTool, call, {
log: () => {},
error: (...args: unknown[]) => {
console.error(`[Tool ${call.name} Error]:`, ...args);
},
warn: () => {},
});
toolMsgs.push({
content: JSON.stringify(toolOutput.output),
role: "user",
options: {
toolResult: {
result: toolOutput.output,
isError: toolOutput.isError,
id: call.id,
},
},
});
}
const toolMsgs = await callTools({
tools: this.tools,
toolCalls,
ctx,
agentName: this.name,
});
for (const msg of toolMsgs) {
this.memory.put(msg);
}
return new InputEvent({ input: this.memory.getMessages() });
}
};
private writeEvent(msg: string, context: Context) {
writeEvent = (
msg: string,
ctx: HandlerContext<FunctionCallingAgentContextData>,
) => {
if (!this.writeEvents) return;
context.writeEventToStream({
data: new AgentRunEvent({ name: this.name, msg }),
});
}
private checkToolCallSupport() {
const { supportToolCall } = this.llm as ToolCallLLM;
if (!supportToolCall) throw new Error("LLM does not support tool calls");
}
private getToolCallsFromResponse(
response:
| ChatResponse<ToolCallLLMMessageOptions>
| ChatResponseChunk<ToolCallLLMMessageOptions>,
): ToolCall[] {
let options;
if ("message" in response) {
options = response.message.options;
} else {
options = response.options;
}
if (options && "toolCall" in options) {
return options.toolCall as ToolCall[];
}
return [];
}
ctx.sendEvent(
new AgentRunEvent({ agent: this.name, text: msg, type: "text" }),
);
};
}
@@ -1,65 +1,77 @@
import { StopEvent } from "@llamaindex/core/workflow";
import {
createCallbacksTransformer,
createStreamDataTransformer,
StopEvent,
WorkflowContext,
WorkflowEvent,
} from "@llamaindex/workflow";
import {
StreamData,
createStreamDataTransformer,
trimStartOfStreamHelper,
type AIStreamCallbacksAndOptions,
} from "ai";
import { ChatResponseChunk } from "llamaindex";
import { AgentRunEvent } from "./type";
export function toDataStream(
result: Promise<StopEvent<AsyncGenerator<ChatResponseChunk>>>,
callbacks?: AIStreamCallbacksAndOptions,
) {
return toReadableStream(result)
.pipeThrough(createCallbacksTransformer(callbacks))
.pipeThrough(createStreamDataTransformer());
}
function toReadableStream(
result: Promise<StopEvent<AsyncGenerator<ChatResponseChunk>>>,
) {
export async function createStreamFromWorkflowContext<Input, Output, Context>(
context: WorkflowContext<Input, Output, Context>,
): Promise<{ stream: ReadableStream<string>; dataStream: StreamData }> {
const trimStartOfStream = trimStartOfStreamHelper();
return new ReadableStream<string>({
start(controller) {
controller.enqueue(""); // Kickstart the stream
const dataStream = new StreamData();
const encoder = new TextEncoder();
let generator: AsyncGenerator<ChatResponseChunk> | undefined;
const closeStreams = (controller: ReadableStreamDefaultController) => {
controller.close();
dataStream.close();
};
const mainStream = new ReadableStream({
async start(controller) {
// Kickstart the stream by sending an empty string
controller.enqueue(encoder.encode(""));
},
async pull(controller): Promise<void> {
const stopEvent = await result;
const generator = stopEvent.data.result;
const { value, done } = await generator.next();
async pull(controller) {
while (!generator) {
// get next event from workflow context
const { value: event, done } =
await context[Symbol.asyncIterator]().next();
if (done) {
closeStreams(controller);
return;
}
generator = handleEvent(event, dataStream);
}
const { value: chunk, done } = await generator.next();
if (done) {
controller.close();
closeStreams(controller);
return;
}
const text = trimStartOfStream(value.delta ?? "");
if (text) controller.enqueue(text);
const text = trimStartOfStream(chunk.delta ?? "");
if (text) {
controller.enqueue(encoder.encode(text));
}
},
});
return {
stream: mainStream.pipeThrough(createStreamDataTransformer()),
dataStream,
};
}
export async function workflowEventsToStreamData(
events: AsyncIterable<AgentRunEvent>,
): Promise<StreamData> {
const streamData = new StreamData();
(async () => {
for await (const event of events) {
if (event instanceof AgentRunEvent) {
const { name, msg } = event.data;
if ((streamData as any).isClosed) {
break;
}
streamData.appendMessageAnnotation({
type: "agent",
data: { agent: name, text: msg },
});
}
}
})();
return streamData;
function handleEvent(
event: WorkflowEvent<any>,
dataStream: StreamData,
): AsyncGenerator<ChatResponseChunk> | undefined {
// Handle for StopEvent
if (event instanceof StopEvent) {
return event.data as AsyncGenerator<ChatResponseChunk>;
}
// Handle for AgentRunEvent
if (event instanceof AgentRunEvent) {
dataStream.appendMessageAnnotation({
type: "agent",
data: event.data,
});
}
}
@@ -0,0 +1,342 @@
import { HandlerContext } from "@llamaindex/workflow";
import {
BaseToolWithCall,
callTool,
ChatMessage,
ChatResponse,
ChatResponseChunk,
LlamaCloudIndex,
PartialToolCall,
QueryEngineTool,
ToolCall,
ToolCallLLM,
ToolCallLLMMessageOptions,
} from "llamaindex";
import crypto from "node:crypto";
import { getDataSource } from "../engine";
import { AgentRunEvent } from "./type";
export const getQueryEngineTools = async (): Promise<
QueryEngineTool[] | null
> => {
const topK = process.env.TOP_K ? parseInt(process.env.TOP_K) : undefined;
const index = await getDataSource();
if (!index) {
return null;
}
// index is LlamaCloudIndex use two query engine tools
if (index instanceof LlamaCloudIndex) {
return [
new QueryEngineTool({
queryEngine: index.asQueryEngine({
similarityTopK: topK,
retrieval_mode: "files_via_content",
}),
metadata: {
name: "document_retriever",
description: `Document retriever that retrieves entire documents from the corpus.
ONLY use for research questions that may require searching over entire research reports.
Will be slower and more expensive than chunk-level retrieval but may be necessary.`,
},
}),
new QueryEngineTool({
queryEngine: index.asQueryEngine({
similarityTopK: topK,
retrieval_mode: "chunks",
}),
metadata: {
name: "chunk_retriever",
description: `Retrieves a small set of relevant document chunks from the corpus.
Use for research questions that want to look up specific facts from the knowledge corpus,
and need entire documents.`,
},
}),
];
} else {
return [
new QueryEngineTool({
queryEngine: index.asQueryEngine({
similarityTopK: topK,
}),
metadata: {
name: "retriever",
description: `Use this tool to retrieve information about the text corpus from the index.`,
},
}),
];
}
};
/**
* Call multiple tools and return the tool messages
*/
export const callTools = async <T>({
tools,
toolCalls,
ctx,
agentName,
writeEvent = true,
}: {
toolCalls: ToolCall[];
tools: BaseToolWithCall[];
ctx: HandlerContext<T>;
agentName: string;
writeEvent?: boolean;
}): Promise<ChatMessage[]> => {
const toolMsgs: ChatMessage[] = [];
if (toolCalls.length === 0) {
return toolMsgs;
}
if (toolCalls.length === 1) {
const tool = tools.find((tool) => tool.metadata.name === toolCalls[0].name);
if (!tool) {
throw new Error(`Tool ${toolCalls[0].name} not found`);
}
return [
await callSingleTool(
tool,
toolCalls[0],
writeEvent
? (msg: string) => {
ctx.sendEvent(
new AgentRunEvent({
agent: agentName,
text: msg,
type: "text",
}),
);
}
: undefined,
),
];
}
// Multiple tool calls, show events in progress
const progressId = crypto.randomUUID();
const totalSteps = toolCalls.length;
let currentStep = 0;
for (const toolCall of toolCalls) {
const tool = tools.find((tool) => tool.metadata.name === toolCall.name);
if (!tool) {
throw new Error(`Tool ${toolCall.name} not found`);
}
const toolMsg = await callSingleTool(tool, toolCall, (msg: string) => {
ctx.sendEvent(
new AgentRunEvent({
agent: agentName,
text: msg,
type: "progress",
data: {
id: progressId,
total: totalSteps,
current: currentStep,
},
}),
);
currentStep++;
});
toolMsgs.push(toolMsg);
}
return toolMsgs;
};
export const callSingleTool = async (
tool: BaseToolWithCall,
toolCall: ToolCall,
eventEmitter?: (msg: string) => void,
): Promise<ChatMessage> => {
if (eventEmitter) {
eventEmitter(
`Calling tool ${toolCall.name} with input: ${JSON.stringify(toolCall.input)}`,
);
}
const toolOutput = await callTool(tool, toolCall, {
log: () => {},
error: (...args: unknown[]) => {
console.error(`Tool ${toolCall.name} got error:`, ...args);
if (eventEmitter) {
eventEmitter(`Tool ${toolCall.name} got error: ${args.join(" ")}`);
}
return {
content: JSON.stringify({
error: args.join(" "),
}),
role: "user",
options: {
toolResult: {
id: toolCall.id,
result: JSON.stringify({
error: args.join(" "),
}),
isError: true,
},
},
};
},
warn: () => {},
});
return {
content: JSON.stringify(toolOutput.output),
role: "user",
options: {
toolResult: {
result: toolOutput.output,
isError: toolOutput.isError,
id: toolCall.id,
},
},
};
};
class ChatWithToolsResponse {
toolCalls: ToolCall[];
toolCallMessage?: ChatMessage;
responseGenerator?: AsyncGenerator<ChatResponseChunk>;
constructor(options: {
toolCalls: ToolCall[];
toolCallMessage?: ChatMessage;
responseGenerator?: AsyncGenerator<ChatResponseChunk>;
}) {
this.toolCalls = options.toolCalls;
this.toolCallMessage = options.toolCallMessage;
this.responseGenerator = options.responseGenerator;
}
hasMultipleTools() {
const uniqueToolNames = new Set(this.getToolNames());
return uniqueToolNames.size > 1;
}
hasToolCall() {
return this.toolCalls.length > 0;
}
getToolNames() {
return this.toolCalls.map((toolCall) => toolCall.name);
}
async asFullResponse(): Promise<ChatMessage> {
if (!this.responseGenerator) {
throw new Error("No response generator");
}
let fullResponse = "";
for await (const chunk of this.responseGenerator) {
fullResponse += chunk.delta;
}
return {
role: "assistant",
content: fullResponse,
};
}
}
export const chatWithTools = async (
llm: ToolCallLLM,
tools: BaseToolWithCall[],
messages: ChatMessage[],
): Promise<ChatWithToolsResponse> => {
const responseGenerator = async function* (): AsyncGenerator<
boolean | ChatResponseChunk,
void,
unknown
> {
const responseStream = await llm.chat({ messages, tools, stream: true });
let fullResponse = null;
let yieldedIndicator = false;
const toolCallMap = new Map();
for await (const chunk of responseStream) {
const hasToolCalls = chunk.options && "toolCall" in chunk.options;
if (!hasToolCalls) {
if (!yieldedIndicator) {
yield false;
yieldedIndicator = true;
}
yield chunk;
} else if (!yieldedIndicator) {
yield true;
yieldedIndicator = true;
}
if (chunk.options && "toolCall" in chunk.options) {
for (const toolCall of chunk.options.toolCall as PartialToolCall[]) {
if (toolCall.id) {
toolCallMap.set(toolCall.id, toolCall);
}
}
}
if (
hasToolCalls &&
(chunk.raw as any)?.choices?.[0]?.finish_reason !== null
) {
// Update the fullResponse with the tool calls
const toolCalls = Array.from(toolCallMap.values());
fullResponse = {
...chunk,
options: {
...chunk.options,
toolCall: toolCalls,
},
};
}
}
if (fullResponse) {
yield fullResponse;
}
};
const generator = responseGenerator();
const isToolCall = await generator.next();
if (isToolCall.value) {
// If it's a tool call, we need to wait for the full response
let fullResponse = null;
for await (const chunk of generator) {
fullResponse = chunk;
}
if (fullResponse) {
const responseChunk = fullResponse as ChatResponseChunk;
const toolCalls = getToolCallsFromResponse(responseChunk);
return new ChatWithToolsResponse({
toolCalls,
toolCallMessage: {
options: responseChunk.options,
role: "assistant",
content: "",
},
});
} else {
throw new Error("Cannot get tool calls from response");
}
}
return new ChatWithToolsResponse({
toolCalls: [],
responseGenerator: generator as AsyncGenerator<ChatResponseChunk>,
});
};
export const getToolCallsFromResponse = (
response:
| ChatResponse<ToolCallLLMMessageOptions>
| ChatResponseChunk<ToolCallLLMMessageOptions>,
): ToolCall[] => {
let options;
if ("message" in response) {
options = response.message.options;
} else {
options = response.options;
}
if (options && "toolCall" in options) {
return options.toolCall as ToolCall[];
}
return [];
};
@@ -1,11 +1,24 @@
import { WorkflowEvent } from "@llamaindex/core/workflow";
import { WorkflowEvent } from "@llamaindex/workflow";
import { MessageContent } from "llamaindex";
export type AgentInput = {
message: string;
message: MessageContent;
streaming?: boolean;
};
export type AgentRunEventType = "text" | "progress";
export type ProgressEventData = {
id: string;
total: number;
current: number;
};
export type AgentRunEventData = ProgressEventData;
export class AgentRunEvent extends WorkflowEvent<{
name: string;
msg: string;
agent: string;
text: string;
type: AgentRunEventType;
data?: AgentRunEventData;
}> {}
@@ -21,6 +21,8 @@ def init_settings():
init_mistral()
case "azure-openai":
init_azure_openai()
case "huggingface":
init_huggingface()
case "t-systems":
from .llmhub import init_llmhub
@@ -138,6 +140,42 @@ def init_fastembed():
)
def init_huggingface_embedding():
try:
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
except ImportError:
raise ImportError(
"Hugging Face support is not installed. Please install it with `poetry add llama-index-embeddings-huggingface`"
)
embedding_model = os.getenv("EMBEDDING_MODEL", "all-MiniLM-L6-v2")
backend = os.getenv("EMBEDDING_BACKEND", "onnx") # "torch", "onnx", or "openvino"
trust_remote_code = (
os.getenv("EMBEDDING_TRUST_REMOTE_CODE", "false").lower() == "true"
)
Settings.embed_model = HuggingFaceEmbedding(
model_name=embedding_model,
trust_remote_code=trust_remote_code,
backend=backend,
)
def init_huggingface():
try:
from llama_index.llms.huggingface import HuggingFaceLLM
except ImportError:
raise ImportError(
"Hugging Face support is not installed. Please install it with `poetry add llama-index-llms-huggingface` and `poetry add llama-index-embeddings-huggingface`"
)
Settings.llm = HuggingFaceLLM(
model_name=os.getenv("MODEL"),
tokenizer_name=os.getenv("MODEL"),
)
init_huggingface_embedding()
def init_groq():
try:
from llama_index.llms.groq import Groq
@@ -20,19 +20,22 @@
"dotenv": "^16.3.1",
"duck-duck-scrape": "^2.2.5",
"express": "^4.18.2",
"llamaindex": "0.7.10",
"llamaindex": "0.8.2",
"pdf2json": "3.0.5",
"ajv": "^8.12.0",
"@e2b/code-interpreter": "0.0.9-beta.3",
"got": "^14.4.1",
"@apidevtools/swagger-parser": "^10.1.0",
"formdata-node": "^6.0.3",
"marked": "^14.1.2"
"marked": "^14.1.2",
"papaparse": "^5.4.1"
},
"devDependencies": {
"@types/cors": "^2.8.16",
"@types/express": "^4.17.21",
"@types/node": "^20.9.5",
"@llamaindex/workflow": "^0.0.3",
"@types/papaparse": "^5.3.15",
"concurrently": "^8.2.2",
"eslint": "^8.54.0",
"eslint-config-prettier": "^8.10.0",
@@ -60,9 +60,11 @@ class AnnotationFileData(BaseModel):
# Include document IDs if it's available
if file.refs is not None:
default_content += f"Document IDs: {file.refs}\n"
# Include sandbox file path
# file path
sandbox_file_path = f"/tmp/{file.name}"
local_file_path = f"output/uploaded/{file.name}"
default_content += f"Sandbox file path (instruction: only use sandbox path for artifact or code interpreter tool): {sandbox_file_path}\n"
default_content += f"Local file path (instruction: Use for local tools: form filling, extractor): {local_file_path}\n"
return default_content
def to_llm_content(self) -> Optional[str]:
@@ -128,24 +130,29 @@ class ChatData(BaseModel):
def get_last_message_content(self) -> str:
"""
Get the content of the last message along with the data content if available.
Fallback to use data content from previous messages
Get the content of the last message along with the data content from all user messages
"""
if len(self.messages) == 0:
raise ValueError("There is not any message in the chat")
last_message = self.messages[-1]
message_content = last_message.content
for message in reversed(self.messages):
# Collect annotation contents from all user messages
all_annotation_contents: List[str] = []
for message in self.messages:
if message.role == MessageRole.USER and message.annotations is not None:
annotation_contents = filter(
None,
[annotation.to_content() for annotation in message.annotations],
)
if not annotation_contents:
continue
annotation_text = "\n".join(annotation_contents)
message_content = f"{message_content}\n{annotation_text}"
break
all_annotation_contents.extend(annotation_contents)
# Add all annotation contents if any exist
if len(all_annotation_contents) > 0:
annotation_text = "\n".join(all_annotation_contents)
message_content = f"{message_content}\n{annotation_text}"
return message_content
def _get_agent_messages(self, max_messages: int = 10) -> List[str]:
@@ -6,7 +6,7 @@ import re
import uuid
from io import BytesIO
from pathlib import Path
from typing import Dict, List, Optional, Tuple
from typing import List, Optional, Tuple
from llama_index.core import VectorStoreIndex
from llama_index.core.ingestion import IngestionPipeline
@@ -14,7 +14,6 @@ from llama_index.core.readers.file.base import (
_try_loading_included_file_formats as get_file_loaders_map,
)
from llama_index.core.schema import Document
from llama_index.core.tools.function_tool import FunctionTool
from llama_index.indices.managed.llama_cloud.base import LlamaCloudIndex
from llama_index.readers.file import FlatReader
from pydantic import BaseModel, Field
@@ -78,10 +77,8 @@ class FileService:
save_dir=PRIVATE_STORE_PATH,
)
tools = _get_available_tools()
code_executor_tools = ["interpreter", "artifact"]
# If the file is CSV and there is a code executor tool, we don't need to index.
if extension == "csv" and any(tool in tools for tool in code_executor_tools):
# Don't index csv files (they are handled by tools)
if extension == "csv":
return document_file
else:
# Insert the file into the index and update document ids to the file metadata
@@ -283,18 +280,3 @@ def _default_file_loaders_map():
default_loaders[".txt"] = FlatReader
default_loaders[".csv"] = FlatReader
return default_loaders
def _get_available_tools() -> Dict[str, List[FunctionTool]]:
try:
from app.engine.tools import ToolFactory # type: ignore
except ImportError:
logger.warning("ToolFactory not found, no tools will be available")
return {}
try:
tools = ToolFactory.from_env(map_result=True)
return tools # type: ignore
except Exception as e:
logger.error(f"Error loading tools from environment: {str(e)}")
raise ValueError(f"Failed to get available tools: {str(e)}") from e
@@ -1,57 +1,27 @@
"use client";
import { ChatSection as ChatSectionUI } from "@llamaindex/chat-ui";
import "@llamaindex/chat-ui/styles/code.css";
import "@llamaindex/chat-ui/styles/katex.css";
import "@llamaindex/chat-ui/styles/pdf.css";
import { useChat } from "ai/react";
import { useState } from "react";
import { ChatInput, ChatMessages } from "./ui/chat";
import CustomChatInput from "./ui/chat/chat-input";
import CustomChatMessages from "./ui/chat/chat-messages";
import { useClientConfig } from "./ui/chat/hooks/use-config";
export default function ChatSection() {
const { backend } = useClientConfig();
const [requestData, setRequestData] = useState<any>();
const {
messages,
input,
isLoading,
handleSubmit,
handleInputChange,
reload,
stop,
append,
setInput,
} = useChat({
body: { data: requestData },
const handler = useChat({
api: `${backend}/api/chat`,
headers: {
"Content-Type": "application/json", // using JSON because of vercel/ai 2.2.26
},
onError: (error: unknown) => {
if (!(error instanceof Error)) throw error;
const message = JSON.parse(error.message);
alert(message.detail);
alert(JSON.parse(error.message).detail);
},
sendExtraMessageFields: true,
});
return (
<div className="space-y-4 w-full h-full flex flex-col">
<ChatMessages
messages={messages}
isLoading={isLoading}
reload={reload}
stop={stop}
append={append}
/>
<ChatInput
input={input}
handleSubmit={handleSubmit}
handleInputChange={handleInputChange}
isLoading={isLoading}
messages={messages}
append={append}
setInput={setInput}
requestParams={{ params: requestData }}
setRequestData={setRequestData}
/>
</div>
<ChatSectionUI handler={handler} className="w-full h-full">
<CustomChatMessages />
<CustomChatInput />
</ChatSectionUI>
);
}
@@ -1 +0,0 @@
Using the chat component from https://github.com/marcusschiesser/ui (based on https://ui.shadcn.com/)
@@ -1,28 +0,0 @@
import { PauseCircle, RefreshCw } from "lucide-react";
import { Button } from "../button";
import { ChatHandler } from "./chat.interface";
export default function ChatActions(
props: Pick<ChatHandler, "stop" | "reload"> & {
showReload?: boolean;
showStop?: boolean;
},
) {
return (
<div className="space-x-4">
{props.showStop && (
<Button variant="outline" size="sm" onClick={props.stop}>
<PauseCircle className="mr-2 h-4 w-4" />
Stop generating
</Button>
)}
{props.showReload && (
<Button variant="outline" size="sm" onClick={props.reload}>
<RefreshCw className="mr-2 h-4 w-4" />
Regenerate
</Button>
)}
</div>
);
}
@@ -1,8 +1,10 @@
import { useChatMessage } from "@llamaindex/chat-ui";
import { User2 } from "lucide-react";
import Image from "next/image";
export default function ChatAvatar({ role }: { role: string }) {
if (role === "user") {
export function ChatMessageAvatar() {
const { message } = useChatMessage();
if (message.role === "user") {
return (
<div className="flex h-8 w-8 shrink-0 select-none items-center justify-center rounded-md border bg-background shadow">
<User2 className="h-4 w-4" />
@@ -1,34 +1,13 @@
import { JSONValue } from "ai";
import React from "react";
import { DocumentFile } from ".";
import { Button } from "../button";
import { DocumentPreview } from "../document-preview";
import FileUploader from "../file-uploader";
import { Textarea } from "../textarea";
import UploadImagePreview from "../upload-image-preview";
import { ChatHandler } from "./chat.interface";
import { useFile } from "./hooks/use-file";
import { LlamaCloudSelector } from "./widgets/LlamaCloudSelector";
"use client";
const ALLOWED_EXTENSIONS = ["png", "jpg", "jpeg", "csv", "pdf", "txt", "docx"];
import { ChatInput, useChatUI, useFile } from "@llamaindex/chat-ui";
import { DocumentPreview, ImagePreview } from "@llamaindex/chat-ui/widgets";
import { LlamaCloudSelector } from "./custom/llama-cloud-selector";
import { useClientConfig } from "./hooks/use-config";
export default function ChatInput(
props: Pick<
ChatHandler,
| "isLoading"
| "input"
| "onFileUpload"
| "onFileError"
| "handleSubmit"
| "handleInputChange"
| "messages"
| "setInput"
| "append"
> & {
requestParams?: any;
setRequestData?: React.Dispatch<any>;
},
) {
export default function CustomChatInput() {
const { requestData, isLoading, input } = useChatUI();
const { backend } = useClientConfig();
const {
imageUrl,
setImageUrl,
@@ -37,101 +16,65 @@ export default function ChatInput(
removeDoc,
reset,
getAnnotations,
} = useFile();
// default submit function does not handle including annotations in the message
// so we need to use append function to submit new message with annotations
const handleSubmitWithAnnotations = (
e: React.FormEvent<HTMLFormElement>,
annotations: JSONValue[] | undefined,
) => {
e.preventDefault();
props.append!({
content: props.input,
role: "user",
createdAt: new Date(),
annotations,
});
props.setInput!("");
};
const onSubmit = (e: React.FormEvent<HTMLFormElement>) => {
e.preventDefault();
const annotations = getAnnotations();
if (annotations.length) {
handleSubmitWithAnnotations(e, annotations);
return reset();
}
props.handleSubmit(e);
};
} = useFile({ uploadAPI: `${backend}/api/chat/upload` });
/**
* Handles file uploads. Overwrite to hook into the file upload behavior.
* @param file The file to upload
*/
const handleUploadFile = async (file: File) => {
// There's already an image uploaded, only allow one image at a time
if (imageUrl) {
alert("You can only upload one image at a time.");
return;
}
try {
await uploadFile(file, props.requestParams);
props.onFileUpload?.(file);
// Upload the file and send with it the current request data
await uploadFile(file, requestData);
} catch (error: any) {
const onFileUploadError = props.onFileError || window.alert;
onFileUploadError(error.message);
// Show error message if upload fails
alert(error.message);
}
};
const handleKeyDown = (e: React.KeyboardEvent<HTMLTextAreaElement>) => {
if (e.key === "Enter" && !e.shiftKey) {
e.preventDefault();
onSubmit(e as unknown as React.FormEvent<HTMLFormElement>);
}
};
// Get references to the upload files in message annotations format, see https://github.com/run-llama/chat-ui/blob/main/packages/chat-ui/src/hook/use-file.tsx#L56
const annotations = getAnnotations();
return (
<form
onSubmit={onSubmit}
className="rounded-xl bg-white p-4 shadow-xl space-y-4 shrink-0"
<ChatInput
className="shadow-xl rounded-xl"
resetUploadedFiles={reset}
annotations={annotations}
>
{imageUrl && (
<UploadImagePreview url={imageUrl} onRemove={() => setImageUrl(null)} />
)}
{files.length > 0 && (
<div className="flex gap-4 w-full overflow-auto py-2">
{files.map((file: DocumentFile) => (
<DocumentPreview
key={file.id}
file={file}
onRemove={() => removeDoc(file)}
/>
))}
</div>
)}
<div className="flex w-full items-start justify-between gap-4 ">
<Textarea
id="chat-input"
autoFocus
name="message"
placeholder="Type a message"
className="flex-1 min-h-0 h-[40px]"
value={props.input}
onChange={props.handleInputChange}
onKeyDown={handleKeyDown}
/>
<FileUploader
onFileUpload={handleUploadFile}
onFileError={props.onFileError}
config={{
allowedExtensions: ALLOWED_EXTENSIONS,
disabled: props.isLoading,
}}
/>
{process.env.NEXT_PUBLIC_USE_LLAMACLOUD === "true" &&
props.setRequestData && (
<LlamaCloudSelector setRequestData={props.setRequestData} />
)}
<Button type="submit" disabled={props.isLoading || !props.input.trim()}>
Send message
</Button>
<div>
{/* Image preview section */}
{imageUrl && (
<ImagePreview url={imageUrl} onRemove={() => setImageUrl(null)} />
)}
{/* Document previews section */}
{files.length > 0 && (
<div className="flex gap-4 w-full overflow-auto py-2">
{files.map((file) => (
<DocumentPreview
key={file.id}
file={file}
onRemove={() => removeDoc(file)}
/>
))}
</div>
)}
</div>
</form>
<ChatInput.Form>
<ChatInput.Field />
<ChatInput.Upload onUpload={handleUploadFile} />
<LlamaCloudSelector />
<ChatInput.Submit
disabled={
isLoading || (!input.trim() && files.length === 0 && !imageUrl)
}
/>
</ChatInput.Form>
</ChatInput>
);
}
@@ -0,0 +1,30 @@
import {
ChatMessage,
ContentPosition,
getSourceAnnotationData,
useChatMessage,
} from "@llamaindex/chat-ui";
import { Markdown } from "./custom/markdown";
import { ToolAnnotations } from "./tools/chat-tools";
export function ChatMessageContent() {
const { message } = useChatMessage();
const customContent = [
{
// override the default markdown component
position: ContentPosition.MARKDOWN,
component: (
<Markdown
content={message.content}
sources={getSourceAnnotationData(message.annotations)?.[0]}
/>
),
},
{
// add the tool annotations after events
position: ContentPosition.AFTER_EVENTS,
component: <ToolAnnotations message={message} />,
},
];
return <ChatMessage.Content content={customContent} />;
}
@@ -1,153 +0,0 @@
import { icons, LucideIcon } from "lucide-react";
import { useMemo } from "react";
import { Button } from "../../button";
import {
Drawer,
DrawerClose,
DrawerContent,
DrawerHeader,
DrawerTitle,
DrawerTrigger,
} from "../../drawer";
import { AgentEventData } from "../index";
import Markdown from "./markdown";
const AgentIcons: Record<string, LucideIcon> = {
bot: icons.Bot,
researcher: icons.ScanSearch,
writer: icons.PenLine,
reviewer: icons.MessageCircle,
publisher: icons.BookCheck,
};
type MergedEvent = {
agent: string;
texts: string[];
icon: LucideIcon;
};
export function ChatAgentEvents({
data,
isFinished,
}: {
data: AgentEventData[];
isFinished: boolean;
}) {
const events = useMemo(() => mergeAdjacentEvents(data), [data]);
return (
<div className="pl-2">
<div className="text-sm space-y-4">
{events.map((eventItem, index) => (
<AgentEventContent
key={index}
event={eventItem}
isLast={index === events.length - 1}
isFinished={isFinished}
/>
))}
</div>
</div>
);
}
const MAX_TEXT_LENGTH = 150;
function AgentEventContent({
event,
isLast,
isFinished,
}: {
event: MergedEvent;
isLast: boolean;
isFinished: boolean;
}) {
const { agent, texts } = event;
const AgentIcon = event.icon;
return (
<div className="flex gap-4 border-b pb-4 items-center fadein-agent">
<div className="w-[100px] flex flex-col items-center gap-2">
<div className="relative">
{isLast && !isFinished && (
<div className="absolute -top-0 -right-4">
<span className="relative flex h-3 w-3">
<span className="animate-ping absolute inline-flex h-full w-full rounded-full bg-sky-400 opacity-75"></span>
<span className="relative inline-flex rounded-full h-3 w-3 bg-sky-500"></span>
</span>
</div>
)}
<AgentIcon />
</div>
<span className="font-bold">{agent}</span>
</div>
<ul className="flex-1 list-decimal space-y-2">
{texts.map((text, index) => (
<li className="whitespace-break-spaces" key={index}>
{text.length <= MAX_TEXT_LENGTH && <span>{text}</span>}
{text.length > MAX_TEXT_LENGTH && (
<div>
<span>{text.slice(0, MAX_TEXT_LENGTH)}...</span>
<AgentEventDialog
content={text}
title={`Agent "${agent}" - Step: ${index + 1}`}
>
<span className="font-semibold underline cursor-pointer ml-2">
Show more
</span>
</AgentEventDialog>
</div>
)}
</li>
))}
</ul>
</div>
);
}
type AgentEventDialogProps = {
title: string;
content: string;
children: React.ReactNode;
};
function AgentEventDialog(props: AgentEventDialogProps) {
return (
<Drawer direction="left">
<DrawerTrigger asChild>{props.children}</DrawerTrigger>
<DrawerContent className="w-3/5 mt-24 h-full max-h-[96%] ">
<DrawerHeader className="flex justify-between">
<div className="space-y-2">
<DrawerTitle>{props.title}</DrawerTitle>
</div>
<DrawerClose asChild>
<Button variant="outline">Close</Button>
</DrawerClose>
</DrawerHeader>
<div className="m-4 overflow-auto">
<Markdown content={props.content} />
</div>
</DrawerContent>
</Drawer>
);
}
function mergeAdjacentEvents(events: AgentEventData[]): MergedEvent[] {
const mergedEvents: MergedEvent[] = [];
for (const event of events) {
const lastMergedEvent = mergedEvents[mergedEvents.length - 1];
if (lastMergedEvent && lastMergedEvent.agent === event.agent) {
// If the last event in mergedEvents has the same non-null agent, add the title to it
lastMergedEvent.texts.push(event.text);
} else {
// Otherwise, create a new merged event
mergedEvents.push({
agent: event.agent,
texts: [event.text],
icon: AgentIcons[event.agent] ?? icons.Bot,
});
}
}
return mergedEvents;
}
@@ -1,50 +0,0 @@
import { ChevronDown, ChevronRight, Loader2 } from "lucide-react";
import { useState } from "react";
import { Button } from "../../button";
import {
Collapsible,
CollapsibleContent,
CollapsibleTrigger,
} from "../../collapsible";
import { EventData } from "../index";
export function ChatEvents({
data,
isLoading,
}: {
data: EventData[];
isLoading: boolean;
}) {
const [isOpen, setIsOpen] = useState(false);
const buttonLabel = isOpen ? "Hide events" : "Show events";
const EventIcon = isOpen ? (
<ChevronDown className="h-4 w-4" />
) : (
<ChevronRight className="h-4 w-4" />
);
return (
<div className="border-l-2 border-indigo-400 pl-2">
<Collapsible open={isOpen} onOpenChange={setIsOpen}>
<CollapsibleTrigger asChild>
<Button variant="secondary" className="space-x-2">
{isLoading ? <Loader2 className="h-4 w-4 animate-spin" /> : null}
<span>{buttonLabel}</span>
{EventIcon}
</Button>
</CollapsibleTrigger>
<CollapsibleContent asChild>
<div className="mt-4 text-sm space-y-2">
{data.map((eventItem, index) => (
<div className="whitespace-break-spaces" key={index}>
{eventItem.title}
</div>
))}
</div>
</CollapsibleContent>
</Collapsible>
</div>
);
}
@@ -1,13 +0,0 @@
import { DocumentPreview } from "../../document-preview";
import { DocumentFileData } from "../index";
export function ChatFiles({ data }: { data: DocumentFileData }) {
if (!data.files.length) return null;
return (
<div className="flex gap-2 items-center">
{data.files.map((file, index) => (
<DocumentPreview key={file.id} file={file} />
))}
</div>
);
}
@@ -1,17 +0,0 @@
import Image from "next/image";
import { type ImageData } from "../index";
export function ChatImage({ data }: { data: ImageData }) {
return (
<div className="rounded-md max-w-[200px] shadow-md">
<Image
src={data.url}
width={0}
height={0}
sizes="100vw"
style={{ width: "100%", height: "auto" }}
alt=""
/>
</div>
);
}
@@ -1,173 +0,0 @@
import { Check, Copy } from "lucide-react";
import { useMemo } from "react";
import { Button } from "../../button";
import { PreviewCard } from "../../document-preview";
import {
HoverCard,
HoverCardContent,
HoverCardTrigger,
} from "../../hover-card";
import { cn } from "../../lib/utils";
import { useCopyToClipboard } from "../hooks/use-copy-to-clipboard";
import { DocumentFileType, SourceData, SourceNode } from "../index";
import PdfDialog from "../widgets/PdfDialog";
type Document = {
url: string;
sources: SourceNode[];
};
export function ChatSources({ data }: { data: SourceData }) {
const documents: Document[] = useMemo(() => {
// group nodes by document (a document must have a URL)
const nodesByUrl: Record<string, SourceNode[]> = {};
data.nodes.forEach((node) => {
const key = node.url;
nodesByUrl[key] ??= [];
nodesByUrl[key].push(node);
});
// convert to array of documents
return Object.entries(nodesByUrl).map(([url, sources]) => ({
url,
sources,
}));
}, [data.nodes]);
if (documents.length === 0) return null;
return (
<div className="space-y-2 text-sm">
<div className="font-semibold text-lg">Sources:</div>
<div className="flex gap-3 flex-wrap">
{documents.map((document) => {
return <DocumentInfo key={document.url} document={document} />;
})}
</div>
</div>
);
}
function SourceInfo({ node, index }: { node?: SourceNode; index: number }) {
if (!node) return <SourceNumberButton index={index} />;
return (
<HoverCard>
<HoverCardTrigger
className="cursor-default"
onClick={(e) => {
e.preventDefault();
e.stopPropagation();
}}
>
<SourceNumberButton
index={index}
className="hover:text-white hover:bg-primary"
/>
</HoverCardTrigger>
<HoverCardContent className="w-[400px]">
<NodeInfo nodeInfo={node} />
</HoverCardContent>
</HoverCard>
);
}
export function SourceNumberButton({
index,
className,
}: {
index: number;
className?: string;
}) {
return (
<span
className={cn(
"text-xs w-5 h-5 rounded-full bg-gray-100 inline-flex items-center justify-center",
className,
)}
>
{index + 1}
</span>
);
}
export function DocumentInfo({
document,
className,
}: {
document: Document;
className?: string;
}) {
const { url, sources } = document;
// Extract filename from URL
const urlParts = url.split("/");
const fileName = urlParts.length > 0 ? urlParts[urlParts.length - 1] : url;
const fileExt = fileName?.split(".").pop() as DocumentFileType | undefined;
const previewFile = {
name: fileName,
type: fileExt as DocumentFileType,
};
const DocumentDetail = (
<div className={`relative ${className}`}>
<PreviewCard className={"cursor-pointer"} file={previewFile} />
<div className="absolute bottom-2 right-2 space-x-2 flex">
{sources.map((node: SourceNode, index: number) => (
<div key={node.id}>
<SourceInfo node={node} index={index} />
</div>
))}
</div>
</div>
);
if (url.endsWith(".pdf")) {
// open internal pdf dialog for pdf files when click document card
return <PdfDialog documentId={url} url={url} trigger={DocumentDetail} />;
}
// open external link when click document card for other file types
return <div onClick={() => window.open(url, "_blank")}>{DocumentDetail}</div>;
}
function NodeInfo({ nodeInfo }: { nodeInfo: SourceNode }) {
const { isCopied, copyToClipboard } = useCopyToClipboard({ timeout: 1000 });
const pageNumber =
// XXX: page_label is used in Python, but page_number is used by Typescript
(nodeInfo.metadata?.page_number as number) ??
(nodeInfo.metadata?.page_label as number) ??
null;
return (
<div className="space-y-4">
<div className="flex justify-between items-center">
<span className="font-semibold">
{pageNumber ? `On page ${pageNumber}:` : "Node content:"}
</span>
{nodeInfo.text && (
<Button
onClick={(e) => {
e.stopPropagation();
copyToClipboard(nodeInfo.text);
}}
size="icon"
variant="ghost"
className="h-12 w-12 shrink-0"
>
{isCopied ? (
<Check className="h-4 w-4" />
) : (
<Copy className="h-4 w-4" />
)}
</Button>
)}
</div>
{nodeInfo.text && (
<pre className="max-h-[200px] overflow-auto whitespace-pre-line">
&ldquo;{nodeInfo.text}&rdquo;
</pre>
)}
</div>
);
}
@@ -1,31 +0,0 @@
import { ChatHandler, SuggestedQuestionsData } from "..";
export function SuggestedQuestions({
questions,
append,
isLastMessage,
}: {
questions: SuggestedQuestionsData;
append: Pick<ChatHandler, "append">["append"];
isLastMessage: boolean;
}) {
const showQuestions = isLastMessage && questions.length > 0;
return (
showQuestions &&
append !== undefined && (
<div className="flex flex-col space-y-2">
{questions.map((question, index) => (
<a
key={index}
onClick={() => {
append({ role: "user", content: question });
}}
className="text-sm italic hover:underline cursor-pointer"
>
{"->"} {question}
</a>
))}
</div>
)
);
}
@@ -1,40 +0,0 @@
import { ToolData } from "../index";
import { Artifact, CodeArtifact } from "../widgets/Artifact";
import { WeatherCard, WeatherData } from "../widgets/WeatherCard";
// TODO: If needed, add displaying more tool outputs here
export default function ChatTools({
data,
artifactVersion,
}: {
data: ToolData;
artifactVersion?: number;
}) {
if (!data) return null;
const { toolCall, toolOutput } = data;
if (toolOutput.isError) {
return (
<div className="border-l-2 border-red-400 pl-2">
There was an error when calling the tool {toolCall.name} with input:{" "}
<br />
{JSON.stringify(toolCall.input)}
</div>
);
}
switch (toolCall.name) {
case "get_weather_information":
const weatherData = toolOutput.output as unknown as WeatherData;
return <WeatherCard data={weatherData} />;
case "artifact":
return (
<Artifact
artifact={toolOutput.output as CodeArtifact}
version={artifactVersion}
/>
);
default:
return null;
}
}
@@ -1,131 +0,0 @@
"use client";
import hljs from "highlight.js";
// instead of atom-one-dark theme, there are a lot of others: https://highlightjs.org/demo
import "highlight.js/styles/atom-one-dark-reasonable.css";
import { Check, Copy, Download } from "lucide-react";
import { FC, memo, useEffect, useRef } from "react";
import { Button } from "../../button";
import { useCopyToClipboard } from "../hooks/use-copy-to-clipboard";
interface Props {
language: string;
value: string;
className?: string;
}
interface languageMap {
[key: string]: string | undefined;
}
export const programmingLanguages: languageMap = {
javascript: ".js",
python: ".py",
java: ".java",
c: ".c",
cpp: ".cpp",
"c++": ".cpp",
"c#": ".cs",
ruby: ".rb",
php: ".php",
swift: ".swift",
"objective-c": ".m",
kotlin: ".kt",
typescript: ".ts",
go: ".go",
perl: ".pl",
rust: ".rs",
scala: ".scala",
haskell: ".hs",
lua: ".lua",
shell: ".sh",
sql: ".sql",
html: ".html",
css: ".css",
// add more file extensions here, make sure the key is same as language prop in CodeBlock.tsx component
};
export const generateRandomString = (length: number, lowercase = false) => {
const chars = "ABCDEFGHJKLMNPQRSTUVWXY3456789"; // excluding similar looking characters like Z, 2, I, 1, O, 0
let result = "";
for (let i = 0; i < length; i++) {
result += chars.charAt(Math.floor(Math.random() * chars.length));
}
return lowercase ? result.toLowerCase() : result;
};
const CodeBlock: FC<Props> = memo(({ language, value, className }) => {
const { isCopied, copyToClipboard } = useCopyToClipboard({ timeout: 2000 });
const codeRef = useRef<HTMLElement>(null);
useEffect(() => {
if (codeRef.current && codeRef.current.dataset.highlighted !== "yes") {
hljs.highlightElement(codeRef.current);
}
}, [language, value]);
const downloadAsFile = () => {
if (typeof window === "undefined") {
return;
}
const fileExtension = programmingLanguages[language] || ".file";
const suggestedFileName = `file-${generateRandomString(
3,
true,
)}${fileExtension}`;
const fileName = window.prompt("Enter file name", suggestedFileName);
if (!fileName) {
// User pressed cancel on prompt.
return;
}
const blob = new Blob([value], { type: "text/plain" });
const url = URL.createObjectURL(blob);
const link = document.createElement("a");
link.download = fileName;
link.href = url;
link.style.display = "none";
document.body.appendChild(link);
link.click();
document.body.removeChild(link);
URL.revokeObjectURL(url);
};
const onCopy = () => {
if (isCopied) return;
copyToClipboard(value);
};
return (
<div
className={`codeblock relative w-full bg-zinc-950 font-sans ${className}`}
>
<div className="flex w-full items-center justify-between bg-zinc-800 px-6 py-2 pr-4 text-zinc-100">
<span className="text-xs lowercase">{language}</span>
<div className="flex items-center space-x-1">
<Button variant="ghost" onClick={downloadAsFile} size="icon">
<Download />
<span className="sr-only">Download</span>
</Button>
<Button variant="ghost" size="icon" onClick={onCopy}>
{isCopied ? (
<Check className="h-4 w-4" />
) : (
<Copy className="h-4 w-4" />
)}
<span className="sr-only">Copy code</span>
</Button>
</div>
</div>
<pre className="border border-zinc-700">
<code ref={codeRef} className={`language-${language} font-mono`}>
{value}
</code>
</pre>
</div>
);
});
CodeBlock.displayName = "CodeBlock";
export { CodeBlock };
@@ -1,184 +0,0 @@
import { Check, Copy } from "lucide-react";
import { Message } from "ai";
import { Fragment } from "react";
import { Button } from "../../button";
import { useCopyToClipboard } from "../hooks/use-copy-to-clipboard";
import {
AgentEventData,
ChatHandler,
DocumentFileData,
EventData,
ImageData,
MessageAnnotation,
MessageAnnotationType,
SuggestedQuestionsData,
ToolData,
getAnnotationData,
getSourceAnnotationData,
} from "../index";
import { ChatAgentEvents } from "./chat-agent-events";
import ChatAvatar from "./chat-avatar";
import { ChatEvents } from "./chat-events";
import { ChatFiles } from "./chat-files";
import { ChatImage } from "./chat-image";
import { ChatSources } from "./chat-sources";
import { SuggestedQuestions } from "./chat-suggestedQuestions";
import ChatTools from "./chat-tools";
import Markdown from "./markdown";
type ContentDisplayConfig = {
order: number;
component: JSX.Element | null;
};
function ChatMessageContent({
message,
isLoading,
append,
isLastMessage,
artifactVersion,
}: {
message: Message;
isLoading: boolean;
append: Pick<ChatHandler, "append">["append"];
isLastMessage: boolean;
artifactVersion: number | undefined;
}) {
const annotations = message.annotations as MessageAnnotation[] | undefined;
if (!annotations?.length) return <Markdown content={message.content} />;
const imageData = getAnnotationData<ImageData>(
annotations,
MessageAnnotationType.IMAGE,
);
const contentFileData = getAnnotationData<DocumentFileData>(
annotations,
MessageAnnotationType.DOCUMENT_FILE,
);
const eventData = getAnnotationData<EventData>(
annotations,
MessageAnnotationType.EVENTS,
);
const agentEventData = getAnnotationData<AgentEventData>(
annotations,
MessageAnnotationType.AGENT_EVENTS,
);
const sourceData = getSourceAnnotationData(annotations);
const toolData = getAnnotationData<ToolData>(
annotations,
MessageAnnotationType.TOOLS,
);
const suggestedQuestionsData = getAnnotationData<SuggestedQuestionsData>(
annotations,
MessageAnnotationType.SUGGESTED_QUESTIONS,
);
const contents: ContentDisplayConfig[] = [
{
order: 1,
component: imageData[0] ? <ChatImage data={imageData[0]} /> : null,
},
{
order: -3,
component:
eventData.length > 0 ? (
<ChatEvents isLoading={isLoading} data={eventData} />
) : null,
},
{
order: -2,
component:
agentEventData.length > 0 ? (
<ChatAgentEvents
data={agentEventData}
isFinished={!!message.content}
/>
) : null,
},
{
order: 2,
component: contentFileData[0] ? (
<ChatFiles data={contentFileData[0]} />
) : null,
},
{
order: -1,
component: toolData[0] ? (
<ChatTools data={toolData[0]} artifactVersion={artifactVersion} />
) : null,
},
{
order: 0,
component: <Markdown content={message.content} sources={sourceData[0]} />,
},
{
order: 3,
component: sourceData[0] ? <ChatSources data={sourceData[0]} /> : null,
},
{
order: 4,
component: suggestedQuestionsData[0] ? (
<SuggestedQuestions
questions={suggestedQuestionsData[0]}
append={append}
isLastMessage={isLastMessage}
/>
) : null,
},
];
return (
<div className="flex-1 gap-4 flex flex-col">
{contents
.sort((a, b) => a.order - b.order)
.map((content, index) => (
<Fragment key={index}>{content.component}</Fragment>
))}
</div>
);
}
export default function ChatMessage({
chatMessage,
isLoading,
append,
isLastMessage,
artifactVersion,
}: {
chatMessage: Message;
isLoading: boolean;
append: Pick<ChatHandler, "append">["append"];
isLastMessage: boolean;
artifactVersion: number | undefined;
}) {
const { isCopied, copyToClipboard } = useCopyToClipboard({ timeout: 2000 });
return (
<div className="flex items-start gap-4 pr-5 pt-5">
<ChatAvatar role={chatMessage.role} />
<div className="group flex flex-1 justify-between gap-2">
<ChatMessageContent
message={chatMessage}
isLoading={isLoading}
append={append}
isLastMessage={isLastMessage}
artifactVersion={artifactVersion}
/>
<Button
onClick={() => copyToClipboard(chatMessage.content)}
size="icon"
variant="ghost"
className="h-8 w-8 opacity-0 group-hover:opacity-100"
>
{isCopied ? (
<Check className="h-4 w-4" />
) : (
<Copy className="h-4 w-4" />
)}
</Button>
</div>
</div>
);
}
@@ -1,172 +0,0 @@
import "katex/dist/katex.min.css";
import { FC, memo } from "react";
import ReactMarkdown, { Options } from "react-markdown";
import rehypeKatex from "rehype-katex";
import remarkGfm from "remark-gfm";
import remarkMath from "remark-math";
import { DOCUMENT_FILE_TYPES, DocumentFileType, SourceData } from "..";
import { useClientConfig } from "../hooks/use-config";
import { DocumentInfo, SourceNumberButton } from "./chat-sources";
import { CodeBlock } from "./codeblock";
const MemoizedReactMarkdown: FC<Options> = memo(
ReactMarkdown,
(prevProps, nextProps) =>
prevProps.children === nextProps.children &&
prevProps.className === nextProps.className,
);
const preprocessLaTeX = (content: string) => {
// Replace block-level LaTeX delimiters \[ \] with $$ $$
const blockProcessedContent = content.replace(
/\\\[([\s\S]*?)\\\]/g,
(_, equation) => `$$${equation}$$`,
);
// Replace inline LaTeX delimiters \( \) with $ $
const inlineProcessedContent = blockProcessedContent.replace(
/\\\[([\s\S]*?)\\\]/g,
(_, equation) => `$${equation}$`,
);
return inlineProcessedContent;
};
const preprocessMedia = (content: string) => {
// Remove `sandbox:` from the beginning of the URL
// to fix OpenAI's models issue appending `sandbox:` to the relative URL
return content.replace(/(sandbox|attachment|snt):/g, "");
};
/**
* Update the citation flag [citation:id]() to the new format [citation:index](url)
*/
const preprocessCitations = (content: string, sources?: SourceData) => {
if (sources) {
const citationRegex = /\[citation:(.+?)\]\(\)/g;
let match;
// Find all the citation references in the content
while ((match = citationRegex.exec(content)) !== null) {
const citationId = match[1];
// Find the source node with the id equal to the citation-id, also get the index of the source node
const sourceNode = sources.nodes.find((node) => node.id === citationId);
// If the source node is found, replace the citation reference with the new format
if (sourceNode !== undefined) {
content = content.replace(
match[0],
`[citation:${sources.nodes.indexOf(sourceNode)}]()`,
);
} else {
// If the source node is not found, remove the citation reference
content = content.replace(match[0], "");
}
}
}
return content;
};
const preprocessContent = (content: string, sources?: SourceData) => {
return preprocessCitations(
preprocessMedia(preprocessLaTeX(content)),
sources,
);
};
export default function Markdown({
content,
sources,
}: {
content: string;
sources?: SourceData;
}) {
const processedContent = preprocessContent(content, sources);
const { backend } = useClientConfig();
return (
<MemoizedReactMarkdown
className="prose dark:prose-invert prose-p:leading-relaxed prose-pre:p-0 break-words custom-markdown"
remarkPlugins={[remarkGfm, remarkMath]}
rehypePlugins={[rehypeKatex as any]}
components={{
p({ children }) {
return <div className="mb-2 last:mb-0">{children}</div>;
},
code({ node, inline, className, children, ...props }) {
if (children.length) {
if (children[0] == "▍") {
return (
<span className="mt-1 animate-pulse cursor-default"></span>
);
}
children[0] = (children[0] as string).replace("`▍`", "▍");
}
const match = /language-(\w+)/.exec(className || "");
if (inline) {
return (
<code className={className} {...props}>
{children}
</code>
);
}
return (
<CodeBlock
key={Math.random()}
language={(match && match[1]) || ""}
value={String(children).replace(/\n$/, "")}
className="mb-2"
{...props}
/>
);
},
a({ href, children }) {
// If href starts with `{backend}/api/files`, then it's a local document and we use DocumenInfo for rendering
if (href?.startsWith(backend + "/api/files")) {
// Check if the file is document file type
const fileExtension = href.split(".").pop()?.toLowerCase();
if (
fileExtension &&
DOCUMENT_FILE_TYPES.includes(fileExtension as DocumentFileType)
) {
return (
<DocumentInfo
document={{
url: backend
? new URL(decodeURIComponent(href)).href
: href,
sources: [],
}}
className="mb-2 mt-2"
/>
);
}
}
// If a text link starts with 'citation:', then render it as a citation reference
if (
Array.isArray(children) &&
typeof children[0] === "string" &&
children[0].startsWith("citation:")
) {
const index = Number(children[0].replace("citation:", ""));
if (!isNaN(index)) {
return <SourceNumberButton index={index} />;
} else {
// citation is not looked up yet, don't render anything
return <></>;
}
}
return (
<a href={href} target="_blank">
{children}
</a>
);
},
}}
>
{processedContent}
</MemoizedReactMarkdown>
);
}
@@ -1,136 +1,30 @@
import { Loader2 } from "lucide-react";
import { useEffect, useMemo, useRef, useState } from "react";
"use client";
import { ToolData } from ".";
import { Button } from "../button";
import ChatActions from "./chat-actions";
import ChatMessage from "./chat-message";
import { ChatHandler } from "./chat.interface";
import { useClientConfig } from "./hooks/use-config";
export default function ChatMessages(
props: Pick<
ChatHandler,
"messages" | "isLoading" | "reload" | "stop" | "append"
>,
) {
const { backend } = useClientConfig();
const [starterQuestions, setStarterQuestions] = useState<string[]>();
const scrollableChatContainerRef = useRef<HTMLDivElement>(null);
const messageLength = props.messages.length;
const lastMessage = props.messages[messageLength - 1];
const scrollToBottom = () => {
if (scrollableChatContainerRef.current) {
scrollableChatContainerRef.current.scrollTop =
scrollableChatContainerRef.current.scrollHeight;
}
};
const isLastMessageFromAssistant =
messageLength > 0 && lastMessage?.role !== "user";
const showReload =
props.reload && !props.isLoading && isLastMessageFromAssistant;
const showStop = props.stop && props.isLoading;
// `isPending` indicate
// that stream response is not yet received from the server,
// so we show a loading indicator to give a better UX.
const isPending = props.isLoading && !isLastMessageFromAssistant;
useEffect(() => {
scrollToBottom();
}, [messageLength, lastMessage]);
useEffect(() => {
if (!starterQuestions) {
fetch(`${backend}/api/chat/config`)
.then((response) => response.json())
.then((data) => {
if (data?.starterQuestions) {
setStarterQuestions(data.starterQuestions);
}
})
.catch((error) => console.error("Error fetching config", error));
}
}, [starterQuestions, backend]);
// build a map of message id to artifact version
const artifactVersionMap = useMemo(() => {
const map = new Map<string, number | undefined>();
let versionIndex = 1;
props.messages.forEach((m) => {
m.annotations?.forEach((annotation) => {
if (
typeof annotation === "object" &&
annotation != null &&
"type" in annotation &&
annotation.type === "tools"
) {
const data = annotation.data as ToolData;
if (data?.toolCall?.name === "artifact") {
map.set(m.id, versionIndex);
versionIndex++;
}
}
});
});
return map;
}, [props.messages]);
import { ChatMessage, ChatMessages, useChatUI } from "@llamaindex/chat-ui";
import { ChatMessageAvatar } from "./chat-avatar";
import { ChatMessageContent } from "./chat-message-content";
import { ChatStarter } from "./chat-starter";
export default function CustomChatMessages() {
const { messages } = useChatUI();
return (
<div
className="flex-1 w-full rounded-xl bg-white p-4 shadow-xl relative overflow-y-auto"
ref={scrollableChatContainerRef}
>
<div className="flex flex-col gap-5 divide-y">
{props.messages.map((m, i) => {
const isLoadingMessage = i === messageLength - 1 && props.isLoading;
return (
<ChatMessage
key={m.id}
chatMessage={m}
isLoading={isLoadingMessage}
append={props.append!}
isLastMessage={i === messageLength - 1}
artifactVersion={artifactVersionMap.get(m.id)}
/>
);
})}
{isPending && (
<div className="flex justify-center items-center pt-10">
<Loader2 className="h-4 w-4 animate-spin" />
</div>
)}
</div>
{(showReload || showStop) && (
<div className="flex justify-end py-4">
<ChatActions
reload={props.reload}
stop={props.stop}
showReload={showReload}
showStop={showStop}
/>
</div>
)}
{!messageLength && starterQuestions?.length && props.append && (
<div className="absolute bottom-6 left-0 w-full">
<div className="grid grid-cols-2 gap-2 mx-20">
{starterQuestions.map((question, i) => (
<Button
variant="outline"
key={i}
onClick={() =>
props.append!({ role: "user", content: question })
}
>
{question}
</Button>
))}
</div>
</div>
)}
</div>
<ChatMessages className="shadow-xl rounded-xl">
<ChatMessages.List>
{messages.map((message, index) => (
<ChatMessage
key={message.id}
message={message}
isLast={index === messages.length - 1}
>
<ChatMessageAvatar />
<ChatMessageContent />
<ChatMessage.Actions />
</ChatMessage>
))}
<ChatMessages.Loading />
</ChatMessages.List>
<ChatMessages.Actions />
<ChatStarter />
</ChatMessages>
);
}
@@ -0,0 +1,26 @@
import { useChatUI } from "@llamaindex/chat-ui";
import { StarterQuestions } from "@llamaindex/chat-ui/widgets";
import { useEffect, useState } from "react";
import { useClientConfig } from "./hooks/use-config";
export function ChatStarter() {
const { append } = useChatUI();
const { backend } = useClientConfig();
const [starterQuestions, setStarterQuestions] = useState<string[]>();
useEffect(() => {
if (!starterQuestions) {
fetch(`${backend}/api/chat/config`)
.then((response) => response.json())
.then((data) => {
if (data?.starterQuestions) {
setStarterQuestions(data.starterQuestions);
}
})
.catch((error) => console.error("Error fetching config", error));
}
}, [starterQuestions, backend]);
if (!starterQuestions?.length) return null;
return <StarterQuestions append={append} questions={starterQuestions} />;
}
@@ -1,25 +0,0 @@
import { Message } from "ai";
export interface ChatHandler {
messages: Message[];
input: string;
isLoading: boolean;
handleSubmit: (
e: React.FormEvent<HTMLFormElement>,
ops?: {
data?: any;
},
) => void;
handleInputChange: (e: React.ChangeEvent<HTMLTextAreaElement>) => void;
reload?: () => void;
stop?: () => void;
onFileUpload?: (file: File) => Promise<void>;
onFileError?: (errMsg: string) => void;
setInput?: (input: string) => void;
append?: (
message: Message | Omit<Message, "id">,
ops?: {
data: any;
},
) => Promise<string | null | undefined>;
}
@@ -1,3 +1,4 @@
import { useChatUI } from "@llamaindex/chat-ui";
import { Loader2 } from "lucide-react";
import { useCallback, useEffect, useState } from "react";
import {
@@ -35,19 +36,18 @@ type LlamaCloudConfig = {
};
export interface LlamaCloudSelectorProps {
setRequestData?: React.Dispatch<any>;
onSelect?: (pipeline: PipelineConfig | undefined) => void;
defaultPipeline?: PipelineConfig;
shouldCheckValid?: boolean;
}
export function LlamaCloudSelector({
setRequestData,
onSelect,
defaultPipeline,
shouldCheckValid = false,
}: LlamaCloudSelectorProps) {
const { backend } = useClientConfig();
const { setRequestData } = useChatUI();
const [config, setConfig] = useState<LlamaCloudConfig>();
const updateRequestParams = useCallback(
@@ -97,6 +97,10 @@ export function LlamaCloudSelector({
setPipeline(JSON.parse(value) as PipelineConfig);
};
if (process.env.NEXT_PUBLIC_USE_LLAMACLOUD !== "true") {
return null;
}
if (!config) {
return (
<div className="flex justify-center items-center p-3">
@@ -0,0 +1,27 @@
import { SourceData } from "@llamaindex/chat-ui";
import { Markdown as MarkdownUI } from "@llamaindex/chat-ui/widgets";
import { useClientConfig } from "../hooks/use-config";
const preprocessMedia = (content: string) => {
// Remove `sandbox:` from the beginning of the URL before rendering markdown
// OpenAI models sometimes prepend `sandbox:` to relative URLs - this fixes it
return content.replace(/(sandbox|attachment|snt):/g, "");
};
export function Markdown({
content,
sources,
}: {
content: string;
sources?: SourceData;
}) {
const { backend } = useClientConfig();
const processedContent = preprocessMedia(content);
return (
<MarkdownUI
content={processedContent}
backend={backend}
sources={sources}
/>
);
}
@@ -1,121 +0,0 @@
"use client";
import { JSONValue } from "llamaindex";
import { useState } from "react";
import {
DocumentFile,
DocumentFileType,
MessageAnnotation,
MessageAnnotationType,
} from "..";
import { useClientConfig } from "./use-config";
const docMineTypeMap: Record<string, DocumentFileType> = {
"text/csv": "csv",
"application/pdf": "pdf",
"text/plain": "txt",
"application/vnd.openxmlformats-officedocument.wordprocessingml.document":
"docx",
};
export function useFile() {
const { backend } = useClientConfig();
const [imageUrl, setImageUrl] = useState<string | null>(null);
const [files, setFiles] = useState<DocumentFile[]>([]);
const addDoc = (file: DocumentFile) => {
const existedFile = files.find((f) => f.id === file.id);
if (!existedFile) {
setFiles((prev) => [...prev, file]);
return true;
}
return false;
};
const removeDoc = (file: DocumentFile) => {
setFiles((prev) => prev.filter((f) => f.id !== file.id));
};
const reset = () => {
imageUrl && setImageUrl(null);
files.length && setFiles([]);
};
const uploadContent = async (
file: File,
requestParams: any = {},
): Promise<DocumentFile> => {
const base64 = await readContent({ file, asUrl: true });
const uploadAPI = `${backend}/api/chat/upload`;
const response = await fetch(uploadAPI, {
method: "POST",
headers: {
"Content-Type": "application/json",
},
body: JSON.stringify({
...requestParams,
base64,
name: file.name,
}),
});
if (!response.ok) throw new Error("Failed to upload document.");
return (await response.json()) as DocumentFile;
};
const getAnnotations = () => {
const annotations: MessageAnnotation[] = [];
if (imageUrl) {
annotations.push({
type: MessageAnnotationType.IMAGE,
data: { url: imageUrl },
});
}
if (files.length > 0) {
annotations.push({
type: MessageAnnotationType.DOCUMENT_FILE,
data: { files },
});
}
return annotations as JSONValue[];
};
const readContent = async (input: {
file: File;
asUrl?: boolean;
}): Promise<string> => {
const { file, asUrl } = input;
const content = await new Promise<string>((resolve, reject) => {
const reader = new FileReader();
if (asUrl) {
reader.readAsDataURL(file);
} else {
reader.readAsText(file);
}
reader.onload = () => resolve(reader.result as string);
reader.onerror = (error) => reject(error);
});
return content;
};
const uploadFile = async (file: File, requestParams: any = {}) => {
if (file.type.startsWith("image/")) {
const base64 = await readContent({ file, asUrl: true });
return setImageUrl(base64);
}
const filetype = docMineTypeMap[file.type];
if (!filetype) throw new Error("Unsupported document type.");
const newDoc = await uploadContent(file, requestParams);
return addDoc(newDoc);
};
return {
imageUrl,
setImageUrl,
files,
removeDoc,
reset,
getAnnotations,
uploadFile,
};
}
@@ -1,131 +0,0 @@
import { JSONValue } from "ai";
import ChatInput from "./chat-input";
import ChatMessages from "./chat-messages";
export { type ChatHandler } from "./chat.interface";
export { ChatInput, ChatMessages };
export enum MessageAnnotationType {
IMAGE = "image",
DOCUMENT_FILE = "document_file",
SOURCES = "sources",
EVENTS = "events",
TOOLS = "tools",
SUGGESTED_QUESTIONS = "suggested_questions",
AGENT_EVENTS = "agent",
}
export type ImageData = {
url: string;
};
export type DocumentFileType = "csv" | "pdf" | "txt" | "docx";
export const DOCUMENT_FILE_TYPES: DocumentFileType[] = [
"csv",
"pdf",
"txt",
"docx",
];
export type DocumentFile = {
id: string;
name: string; // The uploaded file name in the backend
size: number; // The file size in bytes
type: DocumentFileType;
url: string; // The URL of the uploaded file in the backend
refs?: string[]; // DocumentIDs of the uploaded file in the vector index
};
export type DocumentFileData = {
files: DocumentFile[];
};
export type SourceNode = {
id: string;
metadata: Record<string, unknown>;
score?: number;
text: string;
url: string;
};
export type SourceData = {
nodes: SourceNode[];
};
export type EventData = {
title: string;
};
export type AgentEventData = {
agent: string;
text: string;
};
export type ToolData = {
toolCall: {
id: string;
name: string;
input: {
[key: string]: JSONValue;
};
};
toolOutput: {
output: JSONValue;
isError: boolean;
};
};
export type SuggestedQuestionsData = string[];
export type AnnotationData =
| ImageData
| DocumentFileData
| SourceData
| EventData
| AgentEventData
| ToolData
| SuggestedQuestionsData;
export type MessageAnnotation = {
type: MessageAnnotationType;
data: AnnotationData;
};
const NODE_SCORE_THRESHOLD = 0.25;
export function getAnnotationData<T extends AnnotationData>(
annotations: MessageAnnotation[],
type: MessageAnnotationType,
): T[] {
return annotations.filter((a) => a.type === type).map((a) => a.data as T);
}
export function getSourceAnnotationData(
annotations: MessageAnnotation[],
): SourceData[] {
const data = getAnnotationData<SourceData>(
annotations,
MessageAnnotationType.SOURCES,
);
if (data.length > 0) {
const sourceData = data[0] as SourceData;
if (sourceData.nodes) {
sourceData.nodes = preprocessSourceNodes(sourceData.nodes);
}
}
return data;
}
function preprocessSourceNodes(nodes: SourceNode[]): SourceNode[] {
// Filter source nodes has lower score
nodes = nodes
.filter((node) => (node.score ?? 1) > NODE_SCORE_THRESHOLD)
.filter((node) => node.url && node.url.trim() !== "")
.sort((a, b) => (b.score ?? 1) - (a.score ?? 1))
.map((node) => {
// remove trailing slash for node url if exists
node.url = node.url.replace(/\/$/, "");
return node;
});
return nodes;
}
@@ -10,7 +10,7 @@ import {
} from "../../collapsible";
import { cn } from "../../lib/utils";
import { Tabs, TabsContent, TabsList, TabsTrigger } from "../../tabs";
import Markdown from "../chat-message/markdown";
import { Markdown } from "../custom/markdown";
import { useClientConfig } from "../hooks/use-config";
import { useCopyToClipboard } from "../hooks/use-copy-to-clipboard";
@@ -29,12 +29,17 @@ export type CodeArtifact = {
files?: string[];
};
type OutputUrl = {
url: string;
filename: string;
};
type ArtifactResult = {
template: string;
stdout: string[];
stderr: string[];
runtimeError?: { name: string; value: string; tracebackRaw: string[] };
outputUrls: Array<{ url: string; filename: string }>;
outputUrls: OutputUrl[];
url: string;
};
@@ -272,11 +277,7 @@ function CodeSandboxPreview({ url }: { url: string }) {
);
}
function InterpreterOutput({
outputUrls,
}: {
outputUrls: Array<{ url: string; filename: string }>;
}) {
function InterpreterOutput({ outputUrls }: { outputUrls: OutputUrl[] }) {
return (
<ul className="flex flex-col gap-2 mt-4">
{outputUrls.map((url) => (
@@ -0,0 +1,89 @@
import {
getAnnotationData,
MessageAnnotation,
useChatMessage,
useChatUI,
} from "@llamaindex/chat-ui";
import { JSONValue, Message } from "ai";
import { useMemo } from "react";
import { Artifact, CodeArtifact } from "./artifact";
import { WeatherCard, WeatherData } from "./weather-card";
export function ToolAnnotations({ message }: { message: Message }) {
const annotations = message.annotations as MessageAnnotation[] | undefined;
const toolData = annotations
? (getAnnotationData(annotations, "tools") as unknown as ToolData[])
: null;
return toolData?.[0] ? <ChatTools data={toolData[0]} /> : null;
}
// TODO: Used to render outputs of tools. If needed, add more renderers here.
function ChatTools({ data }: { data: ToolData }) {
const { messages } = useChatUI();
const { message } = useChatMessage();
// build a map of message id to artifact version
const artifactVersionMap = useMemo(() => {
const map = new Map<string, number | undefined>();
let versionIndex = 1;
messages.forEach((m) => {
m.annotations?.forEach((annotation: any) => {
if (
typeof annotation === "object" &&
annotation != null &&
"type" in annotation &&
annotation.type === "tools"
) {
const data = annotation.data as ToolData;
if (data?.toolCall?.name === "artifact") {
map.set(m.id, versionIndex);
versionIndex++;
}
}
});
});
return map;
}, [messages]);
if (!data) return null;
const { toolCall, toolOutput } = data;
if (toolOutput.isError) {
return (
<div className="border-l-2 border-red-400 pl-2">
There was an error when calling the tool {toolCall.name} with input:{" "}
<br />
{JSON.stringify(toolCall.input)}
</div>
);
}
switch (toolCall.name) {
case "get_weather_information":
const weatherData = toolOutput.output as unknown as WeatherData;
return <WeatherCard data={weatherData} />;
case "artifact":
return (
<Artifact
artifact={toolOutput.output as CodeArtifact}
version={artifactVersionMap.get(message.id)}
/>
);
default:
return null;
}
}
type ToolData = {
toolCall: {
id: string;
name: string;
input: {
[key: string]: JSONValue;
};
};
toolOutput: {
output: JSONValue;
isError: boolean;
};
};
@@ -1,67 +0,0 @@
import dynamic from "next/dynamic";
import { Button } from "../../button";
import {
Drawer,
DrawerClose,
DrawerContent,
DrawerDescription,
DrawerHeader,
DrawerTitle,
DrawerTrigger,
} from "../../drawer";
export interface PdfDialogProps {
documentId: string;
url: string;
trigger: React.ReactNode;
}
// Dynamic imports for client-side rendering only
const PDFViewer = dynamic(
() => import("@llamaindex/pdf-viewer").then((module) => module.PDFViewer),
{ ssr: false },
);
const PdfFocusProvider = dynamic(
() =>
import("@llamaindex/pdf-viewer").then((module) => module.PdfFocusProvider),
{ ssr: false },
);
export default function PdfDialog(props: PdfDialogProps) {
return (
<Drawer direction="left">
<DrawerTrigger asChild>{props.trigger}</DrawerTrigger>
<DrawerContent className="w-3/5 mt-24 h-full max-h-[96%] ">
<DrawerHeader className="flex justify-between">
<div className="space-y-2">
<DrawerTitle>PDF Content</DrawerTitle>
<DrawerDescription>
File URL:{" "}
<a
className="hover:text-blue-900"
href={props.url}
target="_blank"
>
{props.url}
</a>
</DrawerDescription>
</div>
<DrawerClose asChild>
<Button variant="outline">Close</Button>
</DrawerClose>
</DrawerHeader>
<div className="m-4">
<PdfFocusProvider>
<PDFViewer
file={{
id: props.documentId,
url: props.url,
}}
/>
</PdfFocusProvider>
</div>
</DrawerContent>
</Drawer>
);
}
@@ -1,129 +0,0 @@
import { XCircleIcon } from "lucide-react";
import Image from "next/image";
import DocxIcon from "../ui/icons/docx.svg";
import PdfIcon from "../ui/icons/pdf.svg";
import SheetIcon from "../ui/icons/sheet.svg";
import TxtIcon from "../ui/icons/txt.svg";
import { Button } from "./button";
import { DocumentFile, DocumentFileType } from "./chat";
import {
Drawer,
DrawerClose,
DrawerContent,
DrawerDescription,
DrawerHeader,
DrawerTitle,
DrawerTrigger,
} from "./drawer";
import { cn } from "./lib/utils";
export interface DocumentPreviewProps {
file: DocumentFile;
onRemove?: () => void;
}
export function DocumentPreview(props: DocumentPreviewProps) {
const { name, size, type, refs } = props.file;
if (refs?.length) {
return (
<div title={`Document IDs: ${refs.join(", ")}`}>
<PreviewCard {...props} />
</div>
);
}
return (
<Drawer direction="left">
<DrawerTrigger asChild>
<div>
<PreviewCard className="cursor-pointer" {...props} />
</div>
</DrawerTrigger>
<DrawerContent className="w-3/5 mt-24 h-full max-h-[96%] ">
<DrawerHeader className="flex justify-between">
<div className="space-y-2">
<DrawerTitle>{type.toUpperCase()} Raw Content</DrawerTitle>
<DrawerDescription>
{name} ({inKB(size)} KB)
</DrawerDescription>
</div>
<DrawerClose asChild>
<Button variant="outline">Close</Button>
</DrawerClose>
</DrawerHeader>
<div className="m-4 max-h-[80%] overflow-auto">
{refs?.length && (
<pre className="bg-secondary rounded-md p-4 block text-sm">
{refs.join(", ")}
</pre>
)}
</div>
</DrawerContent>
</Drawer>
);
}
export const FileIcon: Record<DocumentFileType, string> = {
csv: SheetIcon,
pdf: PdfIcon,
docx: DocxIcon,
txt: TxtIcon,
};
export function PreviewCard(props: {
file: {
name: string;
size?: number;
type: DocumentFileType;
};
onRemove?: () => void;
className?: string;
}) {
const { onRemove, file, className } = props;
return (
<div
className={cn(
"p-2 w-60 max-w-60 bg-secondary rounded-lg text-sm relative",
className,
)}
>
<div className="flex flex-row items-center gap-2">
<div className="relative h-8 w-8 shrink-0 overflow-hidden rounded-md flex items-center justify-center">
<Image
className="h-full w-auto object-contain"
priority
src={FileIcon[file.type]}
alt="Icon"
/>
</div>
<div className="overflow-hidden">
<div className="truncate font-semibold">
{file.name} {file.size ? `(${inKB(file.size)} KB)` : ""}
</div>
{file.type && (
<div className="truncate text-token-text-tertiary flex items-center gap-2">
<span>{file.type.toUpperCase()} File</span>
</div>
)}
</div>
</div>
{onRemove && (
<div
className={cn(
"absolute -top-2 -right-2 w-6 h-6 z-10 bg-gray-500 text-white rounded-full",
)}
>
<XCircleIcon
className="w-6 h-6 bg-gray-500 text-white rounded-full"
onClick={onRemove}
/>
</div>
)}
</div>
);
}
function inKB(size: number) {
return Math.round((size / 1024) * 10) / 10;
}
@@ -1,105 +0,0 @@
"use client";
import { Loader2, Paperclip } from "lucide-react";
import { ChangeEvent, useState } from "react";
import { buttonVariants } from "./button";
import { cn } from "./lib/utils";
export interface FileUploaderProps {
config?: {
inputId?: string;
fileSizeLimit?: number;
allowedExtensions?: string[];
checkExtension?: (extension: string) => string | null;
disabled: boolean;
};
onFileUpload: (file: File) => Promise<void>;
onFileError?: (errMsg: string) => void;
}
const DEFAULT_INPUT_ID = "fileInput";
const DEFAULT_FILE_SIZE_LIMIT = 1024 * 1024 * 50; // 50 MB
export default function FileUploader({
config,
onFileUpload,
onFileError,
}: FileUploaderProps) {
const [uploading, setUploading] = useState(false);
const inputId = config?.inputId || DEFAULT_INPUT_ID;
const fileSizeLimit = config?.fileSizeLimit || DEFAULT_FILE_SIZE_LIMIT;
const allowedExtensions = config?.allowedExtensions;
const defaultCheckExtension = (extension: string) => {
if (allowedExtensions && !allowedExtensions.includes(extension)) {
return `Invalid file type. Please select a file with one of these formats: ${allowedExtensions!.join(
",",
)}`;
}
return null;
};
const checkExtension = config?.checkExtension ?? defaultCheckExtension;
const isFileSizeExceeded = (file: File) => {
return file.size > fileSizeLimit;
};
const resetInput = () => {
const fileInput = document.getElementById(inputId) as HTMLInputElement;
fileInput.value = "";
};
const onFileChange = async (e: ChangeEvent<HTMLInputElement>) => {
const file = e.target.files?.[0];
if (!file) return;
setUploading(true);
await handleUpload(file);
resetInput();
setUploading(false);
};
const handleUpload = async (file: File) => {
const onFileUploadError = onFileError || window.alert;
const fileExtension = file.name.split(".").pop() || "";
const extensionFileError = checkExtension(fileExtension);
if (extensionFileError) {
return onFileUploadError(extensionFileError);
}
if (isFileSizeExceeded(file)) {
return onFileUploadError(
`File size exceeded. Limit is ${fileSizeLimit / 1024 / 1024} MB`,
);
}
await onFileUpload(file);
};
return (
<div className="self-stretch">
<input
type="file"
id={inputId}
style={{ display: "none" }}
onChange={onFileChange}
accept={allowedExtensions?.join(",")}
disabled={config?.disabled || uploading}
/>
<label
htmlFor={inputId}
className={cn(
buttonVariants({ variant: "secondary", size: "icon" }),
"cursor-pointer",
uploading && "opacity-50",
)}
>
{uploading ? (
<Loader2 className="h-4 w-4 animate-spin" />
) : (
<Paperclip className="-rotate-45 w-4 h-4" />
)}
</label>
</div>
);
}

Some files were not shown because too many files have changed in this diff Show More