Release 0.2.13 (#335)

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>

.coderabbit.yaml

@@ -0,0 +1,6 @@
+# coderabbit.yml
+reviews:
+  path_instructions:
+    - path: "templates/**"
+      instructions: |
+        For files under the `templates` folder, do not report 'Missing Dependencies Detected' errors.

.github/workflows/e2e.yml

@@ -9,16 +9,16 @@ env:
   POETRY_VERSION: "1.6.1"
 
 jobs:
-  e2e:
-    name: create-llama
+  e2e-python:
+    name: python
     timeout-minutes: 60
     strategy:
       fail-fast: true
       matrix:
-        node-version: [18, 20]
+        node-version: [20]
         python-version: ["3.11"]
         os: [macos-latest, windows-latest, ubuntu-22.04]
-        frameworks: ["nextjs", "express", "fastapi"]
+        frameworks: ["fastapi"]
         datasources: ["--no-files", "--example-file"]
     defaults:
       run:
@@ -60,8 +60,8 @@ jobs:
         run: pnpm run pack-install
         working-directory: .
 
-      - name: Run Playwright tests
-        run: pnpm run e2e
+      - name: Run Playwright tests for Python
+        run: pnpm run e2e:python
         env:
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
           LLAMA_CLOUD_API_KEY: ${{ secrets.LLAMA_CLOUD_API_KEY }}
@@ -72,6 +72,73 @@ jobs:
       - uses: actions/upload-artifact@v3
         if: always()
         with:
-          name: playwright-report
+          name: playwright-report-python
+          path: ./playwright-report/
+          retention-days: 30
+
+  e2e-typescript:
+    name: typescript
+    timeout-minutes: 60
+    strategy:
+      fail-fast: true
+      matrix:
+        node-version: [18, 20]
+        python-version: ["3.11"]
+        os: [macos-latest, windows-latest, ubuntu-22.04]
+        frameworks: ["nextjs", "express"]
+        datasources: ["--no-files", "--example-file"]
+    defaults:
+      run:
+        shell: bash
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up python ${{ matrix.python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+
+      - name: Install Poetry
+        uses: snok/install-poetry@v1
+        with:
+          version: ${{ env.POETRY_VERSION }}
+
+      - uses: pnpm/action-setup@v3
+
+      - name: Setup Node.js ${{ matrix.node-version }}
+        uses: actions/setup-node@v4
+        with:
+          node-version: ${{ matrix.node-version }}
+          cache: "pnpm"
+
+      - name: Install dependencies
+        run: pnpm install
+
+      - name: Install Playwright Browsers
+        run: pnpm exec playwright install --with-deps
+        working-directory: .
+
+      - name: Build create-llama
+        run: pnpm run build
+        working-directory: .
+
+      - name: Install
+        run: pnpm run pack-install
+        working-directory: .
+
+      - name: Run Playwright tests for TypeScript
+        run: pnpm run e2e:typescript
+        env:
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          LLAMA_CLOUD_API_KEY: ${{ secrets.LLAMA_CLOUD_API_KEY }}
+          FRAMEWORK: ${{ matrix.frameworks }}
+          DATASOURCE: ${{ matrix.datasources }}
+        working-directory: .
+
+      - uses: actions/upload-artifact@v3
+        if: always()
+        with:
+          name: playwright-report-typescript
           path: ./playwright-report/
           retention-days: 30

CHANGELOG.md

@@ -1,5 +1,51 @@
 # create-llama
 
+## 0.2.13
+
+### Patch Changes
+
+- 04ddebc: Add publisher agent to multi-agents for generating documents (PDF and HTML)
+- 04ddebc: Allow tool selection for multi-agents (Python and TS)
+
+## 0.2.12
+
+### Patch Changes
+
+- 70f7dca: feat: add test deps for llamaparse
+- ef070c0: Add multi agents template for Typescript
+
+## 0.2.11
+
+### Patch Changes
+
+- 7c2a3f6: fix: postgres import
+
+## 0.2.10
+
+### Patch Changes
+
+- cb8d535: Fix only produces one agent event
+
+## 0.2.9
+
+### Patch Changes
+
+- 0213fe0: Update dependencies for vector stores and add e2e test to ensure that they work as expected.
+
+## 0.2.8
+
+### Patch Changes
+
+- 0031e67: Bump llama-index to 0.11.11 for the multi-agent template
+
+## 0.2.7
+
+### Patch Changes
+
+- 505b8e9: bump: use latest ai package version
+- cf3ec97: Dynamically select model for Groq
+- 8c1087f: feat: enhance style for markdown
+
 ## 0.2.6
 
 ### Patch Changes

e2e/python/resolve_dependencies.spec.ts

@@ -0,0 +1,138 @@
+import { expect, test } from "@playwright/test";
+import { exec } from "child_process";
+import fs from "fs";
+import path from "path";
+import util from "util";
+import { TemplateFramework, TemplateVectorDB } from "../../helpers/types";
+import { createTestDir, runCreateLlama } from "../utils";
+
+const execAsync = util.promisify(exec);
+
+const templateFramework: TemplateFramework = process.env.FRAMEWORK
+  ? (process.env.FRAMEWORK as TemplateFramework)
+  : "fastapi";
+const dataSource: string = process.env.DATASOURCE
+  ? process.env.DATASOURCE
+  : "--example-file";
+
+if (
+  dataSource === "--example-file" // XXX: this test provides its own data source - only trigger it on one data source (usually the CI matrix will trigger multiple data sources)
+) {
+  // vectorDBs, tools, and data source combinations to test
+  const vectorDbs: TemplateVectorDB[] = [
+    "mongo",
+    "pg",
+    "pinecone",
+    "milvus",
+    "astra",
+    "qdrant",
+    "chroma",
+    "weaviate",
+  ];
+
+  const toolOptions = [
+    "wikipedia.WikipediaToolSpec",
+    "google.GoogleSearchToolSpec",
+    "document_generator",
+  ];
+
+  const dataSources = [
+    "--example-file",
+    "--web-source https://www.example.com",
+    "--db-source mysql+pymysql://user:pass@localhost:3306/mydb",
+  ];
+
+  test.describe("Test resolve python dependencies", () => {
+    for (const vectorDb of vectorDbs) {
+      for (const tool of toolOptions) {
+        for (const dataSource of dataSources) {
+          const dataSourceType = dataSource.split(" ")[0];
+          const optionDescription = `vectorDb: ${vectorDb}, tools: ${tool}, dataSource: ${dataSourceType}`;
+
+          test(`options: ${optionDescription}`, async () => {
+            const cwd = await createTestDir();
+
+            const result = await runCreateLlama({
+              cwd,
+              templateType: "streaming",
+              templateFramework,
+              dataSource,
+              vectorDb,
+              port: 3000, // port
+              externalPort: 8000, // externalPort
+              postInstallAction: "none", // postInstallAction
+              templateUI: undefined, // ui
+              appType: "--no-frontend", // appType
+              llamaCloudProjectName: undefined, // llamaCloudProjectName
+              llamaCloudIndexName: undefined, // llamaCloudIndexName
+              tools: tool,
+            });
+            const name = result.projectName;
+
+            // Check if the app folder exists
+            const dirExists = fs.existsSync(path.join(cwd, name));
+            expect(dirExists).toBeTruthy();
+
+            // Check if pyproject.toml exists
+            const pyprojectPath = path.join(cwd, name, "pyproject.toml");
+            const pyprojectExists = fs.existsSync(pyprojectPath);
+            expect(pyprojectExists).toBeTruthy();
+
+            // Run poetry lock
+            try {
+              const { stdout, stderr } = await execAsync(
+                "poetry config virtualenvs.in-project true && poetry lock --no-update",
+                {
+                  cwd: path.join(cwd, name),
+                },
+              );
+              console.log("poetry lock stdout:", stdout);
+              console.error("poetry lock stderr:", stderr);
+            } catch (error) {
+              console.error("Error running poetry lock:", error);
+              throw error;
+            }
+
+            // Check if poetry.lock file was created
+            const poetryLockExists = fs.existsSync(
+              path.join(cwd, name, "poetry.lock"),
+            );
+            expect(poetryLockExists).toBeTruthy();
+
+            // Verify that specific dependencies are in pyproject.toml
+            const pyprojectContent = fs.readFileSync(pyprojectPath, "utf-8");
+            if (vectorDb !== "none") {
+              if (vectorDb === "pg") {
+                expect(pyprojectContent).toContain(
+                  "llama-index-vector-stores-postgres",
+                );
+              } else {
+                expect(pyprojectContent).toContain(
+                  `llama-index-vector-stores-${vectorDb}`,
+                );
+              }
+            }
+            if (tool !== "none") {
+              if (tool === "wikipedia.WikipediaToolSpec") {
+                expect(pyprojectContent).toContain("wikipedia");
+              }
+              if (tool === "google.GoogleSearchToolSpec") {
+                expect(pyprojectContent).toContain("google");
+              }
+            }
+
+            // Check for data source specific dependencies
+            if (dataSource.includes("--web-source")) {
+              expect(pyprojectContent).toContain("llama-index-readers-web");
+            }
+            if (dataSource.includes("--db-source")) {
+              expect(pyprojectContent).toContain(
+                "llama-index-readers-database ",
+              );
+            }
+          });
+        }
+      }
+    }
+  });
+}

e2e/shared/extractor_template.spec.ts

@@ -3,8 +3,8 @@ import { expect, test } from "@playwright/test";
 import { ChildProcess } from "child_process";
 import fs from "fs";
 import path from "path";
-import { TemplateFramework } from "../helpers";
-import { createTestDir, runCreateLlama } from "./utils";
+import { TemplateFramework } from "../../helpers";
+import { createTestDir, runCreateLlama } from "../utils";
 
 const templateFramework: TemplateFramework = process.env.FRAMEWORK
   ? (process.env.FRAMEWORK as TemplateFramework)
@@ -16,9 +16,8 @@ const dataSource: string = process.env.DATASOURCE
 // The extractor template currently only works with FastAPI and files (and not on Windows)
 if (
   process.platform !== "win32" &&
-  templateFramework !== "nextjs" &&
-  templateFramework !== "express" &&
-  dataSource !== "--no-files"
+  templateFramework === "fastapi" &&
+  dataSource === "--example-file"
 ) {
   test.describe("Test extractor template", async () => {
     let frontendPort: number;
@@ -32,16 +31,16 @@ if (
       cwd = await createTestDir();
       frontendPort = Math.floor(Math.random() * 10000) + 10000;
       backendPort = frontendPort + 1;
-      const result = await runCreateLlama(
+      const result = await runCreateLlama({
         cwd,
-        "extractor",
-        "fastapi",
-        "--example-file",
-        "none",
-        frontendPort,
-        backendPort,
-        "runApp",
-      );
+        templateType: "extractor",
+        templateFramework: "fastapi",
+        dataSource: "--example-file",
+        vectorDb: "none",
+        port: frontendPort,
+        externalPort: backendPort,
+        postInstallAction: "runApp",
+      });
       name = result.projectName;
       appProcess = result.appProcess;
     });

e2e/shared/multiagent_template.spec.ts

@@ -7,22 +7,22 @@ import type {
   TemplateFramework,
   TemplatePostInstallAction,
   TemplateUI,
-} from "../helpers";
-import { createTestDir, runCreateLlama, type AppType } from "./utils";
+} from "../../helpers";
+import { createTestDir, runCreateLlama, type AppType } from "../utils";
 
-const templateFramework: TemplateFramework = "fastapi";
+const templateFramework: TemplateFramework = process.env.FRAMEWORK
+  ? (process.env.FRAMEWORK as TemplateFramework)
+  : "fastapi";
 const dataSource: string = "--example-file";
 const templateUI: TemplateUI = "shadcn";
 const templatePostInstallAction: TemplatePostInstallAction = "runApp";
-const appType: AppType = "--frontend";
+const appType: AppType = templateFramework === "nextjs" ? "" : "--frontend";
 const userMessage = "Write a blog post about physical standards for letters";
 
 test.describe(`Test multiagent template ${templateFramework} ${dataSource} ${templateUI} ${appType} ${templatePostInstallAction}`, async () => {
   test.skip(
-    process.platform !== "linux" ||
-      process.env.FRAMEWORK !== "fastapi" ||
-      process.env.DATASOURCE === "--no-files",
-    "The multiagent template currently only works with FastAPI and files. We also only run on Linux to speed up tests.",
+    process.platform !== "linux" || process.env.DATASOURCE === "--no-files",
+    "The multiagent template currently only works with files. We also only run on Linux to speed up tests.",
   );
   let port: number;
   let externalPort: number;
@@ -36,18 +36,18 @@ test.describe(`Test multiagent template ${templateFramework} ${dataSource} ${tem
     port = Math.floor(Math.random() * 10000) + 10000;
     externalPort = port + 1;
     cwd = await createTestDir();
-    const result = await runCreateLlama(
+    const result = await runCreateLlama({
       cwd,
-      "multiagent",
+      templateType: "multiagent",
       templateFramework,
       dataSource,
       vectorDb,
       port,
       externalPort,
-      templatePostInstallAction,
+      postInstallAction: templatePostInstallAction,
       templateUI,
       appType,
-    );
+    });
     name = result.projectName;
     appProcess = result.appProcess;
   });

e2e/shared/streaming_template.spec.ts

@@ -7,8 +7,8 @@ import type {
   TemplateFramework,
   TemplatePostInstallAction,
   TemplateUI,
-} from "../helpers";
-import { createTestDir, runCreateLlama, type AppType } from "./utils";
+} from "../../helpers";
+import { createTestDir, runCreateLlama, type AppType } from "../utils";
 
 const templateFramework: TemplateFramework = process.env.FRAMEWORK
   ? (process.env.FRAMEWORK as TemplateFramework)
@@ -39,20 +39,20 @@ test.describe(`Test streaming template ${templateFramework} ${dataSource} ${temp
     port = Math.floor(Math.random() * 10000) + 10000;
     externalPort = port + 1;
     cwd = await createTestDir();
-    const result = await runCreateLlama(
+    const result = await runCreateLlama({
       cwd,
-      "streaming",
+      templateType: "streaming",
       templateFramework,
       dataSource,
       vectorDb,
       port,
       externalPort,
-      templatePostInstallAction,
+      postInstallAction: templatePostInstallAction,
       templateUI,
       appType,
       llamaCloudProjectName,
       llamaCloudIndexName,
-    );
+    });
     name = result.projectName;
     appProcess = result.appProcess;
   });

e2e/typescript/resolve_dependencies.spec.ts

@@ -0,0 +1,106 @@
+import { expect, test } from "@playwright/test";
+import { exec } from "child_process";
+import fs from "fs";
+import path from "path";
+import util from "util";
+import { TemplateFramework, TemplateVectorDB } from "../../helpers/types";
+import { createTestDir, runCreateLlama } from "../utils";
+
+const execAsync = util.promisify(exec);
+
+const templateFramework: TemplateFramework = process.env.FRAMEWORK
+  ? (process.env.FRAMEWORK as TemplateFramework)
+  : "nextjs";
+const dataSource: string = process.env.DATASOURCE
+  ? process.env.DATASOURCE
+  : "--example-file";
+
+// vectorDBs combinations to test
+const vectorDbs: TemplateVectorDB[] = [
+  "mongo",
+  "pg",
+  "qdrant",
+  "pinecone",
+  "milvus",
+  "astra",
+  "chroma",
+  "llamacloud",
+  "weaviate",
+];
+
+test.describe("Test resolve TS dependencies", () => {
+  // Test vector DBs without LlamaParse
+  for (const vectorDb of vectorDbs) {
+    const optionDescription = `vectorDb: ${vectorDb}, dataSource: ${dataSource}`;
+
+    test(`Vector DB test - ${optionDescription}`, async () => {
+      await runTest(vectorDb, false);
+    });
+  }
+
+  // Test LlamaParse with vectorDB 'none'
+  test(`LlamaParse test - vectorDb: none, dataSource: ${dataSource}, llamaParse: true`, async () => {
+    await runTest("none", true);
+  });
+
+  async function runTest(
+    vectorDb: TemplateVectorDB | "none",
+    useLlamaParse: boolean,
+  ) {
+    const cwd = await createTestDir();
+
+    const result = await runCreateLlama({
+      cwd: cwd,
+      templateType: "streaming",
+      templateFramework: templateFramework,
+      dataSource: dataSource,
+      vectorDb: vectorDb,
+      port: 3000,
+      externalPort: 8000,
+      postInstallAction: "none",
+      templateUI: undefined,
+      appType: templateFramework === "nextjs" ? "" : "--no-frontend",
+      llamaCloudProjectName: undefined,
+      llamaCloudIndexName: undefined,
+      tools: undefined,
+      useLlamaParse: useLlamaParse,
+    });
+    const name = result.projectName;
+
+    // Check if the app folder exists
+    const appDir = path.join(cwd, name);
+    const dirExists = fs.existsSync(appDir);
+    expect(dirExists).toBeTruthy();
+
+    // Install dependencies using pnpm
+    try {
+      const { stderr: installStderr } = await execAsync(
+        "pnpm install --prefer-offline",
+        {
+          cwd: appDir,
+        },
+      );
+    } catch (error) {
+      console.error("Error installing dependencies:", error);
+      throw error;
+    }
+
+    // Run tsc type check and capture the output
+    try {
+      const { stdout, stderr } = await execAsync(
+        "pnpm exec tsc -b --diagnostics",
+        {
+          cwd: appDir,
+        },
+      );
+      // Check if there's any error output
+      expect(stderr).toBeFalsy();
+
+      // Log the stdout for debugging purposes
+      console.log("TypeScript type-check output:", stdout);
+    } catch (error) {
+      console.error("Error running tsc:", error);
+      throw error;
+    }
+  }
+});

e2e/utils.ts

@@ -18,21 +18,39 @@ export type CreateLlamaResult = {
   appProcess: ChildProcess;
 };
 
-// eslint-disable-next-line max-params
-export async function runCreateLlama(
-  cwd: string,
-  templateType: TemplateType,
-  templateFramework: TemplateFramework,
-  dataSource: string,
-  vectorDb: TemplateVectorDB,
-  port: number,
-  externalPort: number,
-  postInstallAction: TemplatePostInstallAction,
-  templateUI?: TemplateUI,
-  appType?: AppType,
-  llamaCloudProjectName?: string,
-  llamaCloudIndexName?: string,
-): Promise<CreateLlamaResult> {
+export type RunCreateLlamaOptions = {
+  cwd: string;
+  templateType: TemplateType;
+  templateFramework: TemplateFramework;
+  dataSource: string;
+  vectorDb: TemplateVectorDB;
+  port: number;
+  externalPort: number;
+  postInstallAction: TemplatePostInstallAction;
+  templateUI?: TemplateUI;
+  appType?: AppType;
+  llamaCloudProjectName?: string;
+  llamaCloudIndexName?: string;
+  tools?: string;
+  useLlamaParse?: boolean;
+};
+
+export async function runCreateLlama({
+  cwd,
+  templateType,
+  templateFramework,
+  dataSource,
+  vectorDb,
+  port,
+  externalPort,
+  postInstallAction,
+  templateUI,
+  appType,
+  llamaCloudProjectName,
+  llamaCloudIndexName,
+  tools,
+  useLlamaParse,
+}: RunCreateLlamaOptions): Promise<CreateLlamaResult> {
   if (!process.env.OPENAI_API_KEY || !process.env.LLAMA_CLOUD_API_KEY) {
     throw new Error(
       "Setting the OPENAI_API_KEY and LLAMA_CLOUD_API_KEY is mandatory to run tests",
@@ -41,10 +59,23 @@ export async function runCreateLlama(
   const name = [
     templateType,
     templateFramework,
-    dataSource,
+    dataSource.split(" ")[0],
     templateUI,
     appType,
   ].join("-");
+
+  // Handle different data source types
+  let dataSourceArgs = [];
+  if (dataSource.includes("--web-source" || "--db-source")) {
+    const webSource = dataSource.split(" ")[1];
+    dataSourceArgs.push("--web-source", webSource);
+  } else if (dataSource.includes("--db-source")) {
+    const dbSource = dataSource.split(" ")[1];
+    dataSourceArgs.push("--db-source", dbSource);
+  } else {
+    dataSourceArgs.push(dataSource);
+  }
+
   const commandArgs = [
     "create-llama",
     name,
@@ -52,7 +83,7 @@ export async function runCreateLlama(
     templateType,
     "--framework",
     templateFramework,
-    dataSource,
+    ...dataSourceArgs,
     "--vector-db",
     vectorDb,
     "--open-ai-key",
@@ -65,8 +96,7 @@ export async function runCreateLlama(
     "--post-install-action",
     postInstallAction,
     "--tools",
-    "none",
-    "--no-llama-parse",
+    tools ?? "none",
     "--observability",
     "none",
     "--llama-cloud-key",
@@ -79,6 +109,11 @@ export async function runCreateLlama(
   if (appType) {
     commandArgs.push(appType);
   }
+  if (useLlamaParse) {
+    commandArgs.push("--use-llama-parse");
+  } else {
+    commandArgs.push("--no-llama-parse");
+  }
 
   const command = commandArgs.join(" ");
   console.log(`running command '${command}' in ${cwd}`);
@@ -91,11 +126,11 @@ export async function runCreateLlama(
     },
   });
   appProcess.stderr?.on("data", (data) => {
-    console.log(data.toString());
+    console.error(data.toString());
   });
   appProcess.on("exit", (code) => {
     if (code !== 0 && code !== null) {
-      throw new Error(`create-llama command was failed!`);
+      throw new Error(`create-llama command failed with exit code ${code}`);
     }
   });
 
@@ -107,6 +142,8 @@ export async function runCreateLlama(
       port,
       externalPort,
     );
+  } else if (postInstallAction === "dependencies") {
+    await waitForProcess(appProcess, 1000 * 60); // wait 1 min for dependencies to be resolved
   } else {
     // wait 10 seconds for create-llama to exit
     await waitForProcess(appProcess, 1000 * 10);

helpers/env-variables.ts

@@ -426,34 +426,35 @@ const getToolEnvs = (tools?: Tool[]): EnvVar[] => {
 const getSystemPromptEnv = (
   tools?: Tool[],
   dataSources?: TemplateDataSource[],
-  framework?: TemplateFramework,
+  template?: TemplateType,
 ): EnvVar[] => {
   const defaultSystemPrompt =
     "You are a helpful assistant who helps users with their questions.";
 
+  const systemPromptEnv: EnvVar[] = [];
   // build tool system prompt by merging all tool system prompts
-  let toolSystemPrompt = "";
-  tools?.forEach((tool) => {
-    const toolSystemPromptEnv = tool.envVars?.find(
-      (env) => env.name === TOOL_SYSTEM_PROMPT_ENV_VAR,
-    );
-    if (toolSystemPromptEnv) {
-      toolSystemPrompt += toolSystemPromptEnv.value + "\n";
-    }
-  });
+  // multiagent template doesn't need system prompt
+  if (template !== "multiagent") {
+    let toolSystemPrompt = "";
+    tools?.forEach((tool) => {
+      const toolSystemPromptEnv = tool.envVars?.find(
+        (env) => env.name === TOOL_SYSTEM_PROMPT_ENV_VAR,
+      );
+      if (toolSystemPromptEnv) {
+        toolSystemPrompt += toolSystemPromptEnv.value + "\n";
+      }
+    });
 
-  const systemPrompt = toolSystemPrompt
-    ? `\"${toolSystemPrompt}\"`
-    : defaultSystemPrompt;
+    const systemPrompt = toolSystemPrompt
+      ? `\"${toolSystemPrompt}\"`
+      : defaultSystemPrompt;
 
-  const systemPromptEnv = [
-    {
+    systemPromptEnv.push({
       name: "SYSTEM_PROMPT",
       description: "The system prompt for the AI model.",
       value: systemPrompt,
-    },
-  ];
-
+    });
+  }
   if (tools?.length == 0 && (dataSources?.length ?? 0 > 0)) {
     const citationPrompt = `'You have provided information from a knowledge base that has been passed to you in nodes of information.
 Each node has useful metadata such as node ID, file name, page, etc.
@@ -559,7 +560,7 @@ export const createBackendEnvFile = async (
     ...getToolEnvs(opts.tools),
     ...getTemplateEnvs(opts.template),
     ...getObservabilityEnvs(opts.observability),
-    ...getSystemPromptEnv(opts.tools, opts.dataSources, opts.framework),
+    ...getSystemPromptEnv(opts.tools, opts.dataSources, opts.template),
   ];
   // Render and write env file
   const content = renderEnvVar(envVars);

helpers/providers/groq.ts

@@ -3,8 +3,55 @@ import prompts from "prompts";
 import { ModelConfigParams } from ".";
 import { questionHandlers, toChoice } from "../../questions";
 
-const MODELS = ["llama3-8b", "llama3-70b", "mixtral-8x7b"];
-const DEFAULT_MODEL = MODELS[0];
+import got from "got";
+import ora from "ora";
+import { red } from "picocolors";
+
+const GROQ_API_URL = "https://api.groq.com/openai/v1";
+
+async function getAvailableModelChoicesGroq(apiKey: string) {
+  if (!apiKey) {
+    throw new Error("Need Groq API key to retrieve model choices");
+  }
+
+  const spinner = ora("Fetching available models from Groq").start();
+  try {
+    const response = await got(`${GROQ_API_URL}/models`, {
+      headers: {
+        Authorization: `Bearer ${apiKey}`,
+      },
+      timeout: 5000,
+      responseType: "json",
+    });
+    const data: any = await response.body;
+    spinner.stop();
+
+    // Filter out the Whisper models
+    return data.data
+      .filter((model: any) => !model.id.toLowerCase().includes("whisper"))
+      .map((el: any) => {
+        return {
+          title: el.id,
+          value: el.id,
+        };
+      });
+  } catch (error: unknown) {
+    spinner.stop();
+    console.log(error);
+    if ((error as any).response?.statusCode === 401) {
+      console.log(
+        red(
+          "Invalid Groq API key provided! Please provide a valid key and try again!",
+        ),
+      );
+    } else {
+      console.log(red("Request failed: " + error));
+    }
+    process.exit(1);
+  }
+}
+
+const DEFAULT_MODEL = "llama3-70b-8192";
 
 // Use huggingface embedding models for now as Groq doesn't support embedding models
 enum HuggingFaceEmbeddingModelType {
@@ -66,12 +113,14 @@ export async function askGroqQuestions({
   // use default model values in CI or if user should not be asked
   const useDefaults = ciInfo.isCI || !askModels;
   if (!useDefaults) {
+    const modelChoices = await getAvailableModelChoicesGroq(config.apiKey!);
+
     const { model } = await prompts(
       {
         type: "select",
         name: "model",
         message: "Which LLM model would you like to use?",
-        choices: MODELS.map(toChoice),
+        choices: modelChoices,
         initial: 0,
       },
       questionHandlers,

helpers/python.ts

@@ -36,28 +36,28 @@ const getAdditionalDependencies = (
     case "mongo": {
       dependencies.push({
         name: "llama-index-vector-stores-mongodb",
-        version: "^0.1.3",
+        version: "^0.3.1",
       });
       break;
     }
     case "pg": {
       dependencies.push({
         name: "llama-index-vector-stores-postgres",
-        version: "^0.1.1",
+        version: "^0.2.5",
       });
       break;
     }
     case "pinecone": {
       dependencies.push({
         name: "llama-index-vector-stores-pinecone",
-        version: "^0.1.3",
+        version: "^0.2.1",
       });
       break;
     }
     case "milvus": {
       dependencies.push({
         name: "llama-index-vector-stores-milvus",
-        version: "^0.1.20",
+        version: "^0.2.0",
       });
       dependencies.push({
         name: "pymilvus",
@@ -68,28 +68,28 @@ const getAdditionalDependencies = (
     case "astra": {
       dependencies.push({
         name: "llama-index-vector-stores-astra-db",
-        version: "^0.1.5",
+        version: "^0.2.0",
       });
       break;
     }
     case "qdrant": {
       dependencies.push({
         name: "llama-index-vector-stores-qdrant",
-        version: "^0.2.8",
+        version: "^0.3.0",
       });
       break;
     }
     case "chroma": {
       dependencies.push({
         name: "llama-index-vector-stores-chroma",
-        version: "^0.1.8",
+        version: "^0.2.0",
       });
       break;
     }
     case "weaviate": {
       dependencies.push({
         name: "llama-index-vector-stores-weaviate",
-        version: "^1.0.2",
+        version: "^1.1.1",
       });
       break;
     }
@@ -130,7 +130,7 @@ const getAdditionalDependencies = (
         case "llamacloud":
           dependencies.push({
             name: "llama-index-indices-managed-llama-cloud",
-            version: "^0.3.0",
+            version: "^0.3.1",
           });
           break;
       }
@@ -364,7 +364,12 @@ export const installPythonTemplate = async ({
   | "modelConfig"
 >) => {
   console.log("\nInitializing Python project with template:", template, "\n");
-  const templatePath = path.join(templatesDir, "types", template, framework);
+  let templatePath;
+  if (template === "extractor") {
+    templatePath = path.join(templatesDir, "types", "extractor", framework);
+  } else {
+    templatePath = path.join(templatesDir, "types", "streaming", framework);
+  }
   await copy("**", root, {
     parents: true,
     cwd: templatePath,
@@ -401,23 +406,42 @@ export const installPythonTemplate = async ({
       cwd: path.join(compPath, "services", "python"),
     });
   }
-
-  if (template === "streaming") {
-    // For the streaming template only:
+  // Copy engine code
+  if (template === "streaming" || template === "multiagent") {
     // Select and copy engine code based on data sources and tools
     let engine;
-    if (dataSources.length > 0 && (!tools || tools.length === 0)) {
-      console.log("\nNo tools selected - use optimized context chat engine\n");
-      engine = "chat";
-    } else {
+    // Multiagent always uses agent engine
+    if (template === "multiagent") {
       engine = "agent";
+    } else {
+      // For streaming, use chat engine by default
+      // Unless tools are selected, in which case use agent engine
+      if (dataSources.length > 0 && (!tools || tools.length === 0)) {
+        console.log(
+          "\nNo tools selected - use optimized context chat engine\n",
+        );
+        engine = "chat";
+      } else {
+        engine = "agent";
+      }
     }
+
+    // Copy engine code
     await copy("**", enginePath, {
       parents: true,
       cwd: path.join(compPath, "engines", "python", engine),
     });
   }
 
+  if (template === "multiagent") {
+    // Copy multi-agent code
+    await copy("**", path.join(root), {
+      parents: true,
+      cwd: path.join(compPath, "multiagent", "python"),
+      rename: assetRelocator,
+    });
+  }
+
   console.log("Adding additional dependencies");
 
   const addOnDependencies = getAdditionalDependencies(

helpers/tools.ts

@@ -110,6 +110,29 @@ For better results, you can specify the region parameter to get results from a s
       },
     ],
   },
+  {
+    display: "Document generator",
+    name: "document_generator",
+    supportedFrameworks: ["fastapi", "nextjs", "express"],
+    dependencies: [
+      {
+        name: "xhtml2pdf",
+        version: "^0.2.14",
+      },
+      {
+        name: "markdown",
+        version: "^3.7",
+      },
+    ],
+    type: ToolType.LOCAL,
+    envVars: [
+      {
+        name: TOOL_SYSTEM_PROMPT_ENV_VAR,
+        description: "System prompt for document generator tool.",
+        value: `If user request for a report or a post, use document generator tool to create a file and reply with the link to the file.`,
+      },
+    ],
+  },
   {
     display: "Code Interpreter",
     name: "interpreter",

helpers/typescript.ts

@@ -33,8 +33,7 @@ export const installTSTemplate = async ({
    * Copy the template files to the target directory.
    */
   console.log("\nInitializing project with template:", template, "\n");
-  const type = template === "multiagent" ? "streaming" : template; // use nextjs streaming template for multiagent
-  const templatePath = path.join(templatesDir, "types", type, framework);
+  const templatePath = path.join(templatesDir, "types", "streaming", framework);
   const copySource = ["**"];
 
   await copy(copySource, root, {
@@ -124,6 +123,30 @@ export const installTSTemplate = async ({
     cwd: path.join(compPath, "vectordbs", "typescript", vectorDb ?? "none"),
   });
 
+  if (template === "multiagent") {
+    const multiagentPath = path.join(compPath, "multiagent", "typescript");
+
+    // copy workflow code for multiagent template
+    await copy("**", path.join(root, relativeEngineDestPath, "workflow"), {
+      parents: true,
+      cwd: path.join(multiagentPath, "workflow"),
+    });
+
+    if (framework === "nextjs") {
+      // patch route.ts file
+      await copy("**", path.join(root, relativeEngineDestPath), {
+        parents: true,
+        cwd: path.join(multiagentPath, "nextjs"),
+      });
+    } else if (framework === "express") {
+      // patch chat.controller.ts file
+      await copy("**", path.join(root, relativeEngineDestPath), {
+        parents: true,
+        cwd: path.join(multiagentPath, "express"),
+      });
+    }
+  }
+
   // copy loader component (TS only supports llama_parse and file for now)
   const loaderFolder = useLlamaParse ? "llama_parse" : "file";
   await copy("**", enginePath, {
@@ -134,7 +157,10 @@ export const installTSTemplate = async ({
   // Select and copy engine code based on data sources and tools
   let engine;
   tools = tools ?? [];
-  if (dataSources.length > 0 && tools.length === 0) {
+  // multiagent template always uses agent engine
+  if (template === "multiagent") {
+    engine = "agent";
+  } else if (dataSources.length > 0 && tools.length === 0) {
     console.log("\nNo tools selected - use optimized context chat engine\n");
     engine = "chat";
   } else {
@@ -145,6 +171,11 @@ export const installTSTemplate = async ({
     cwd: path.join(compPath, "engines", "typescript", engine),
   });
 
+  // copy settings to engine folder
+  await copy("**", enginePath, {
+    cwd: path.join(compPath, "settings", "typescript"),
+  });
+
   /**
    * Copy the selected UI files to the target directory and reference it.
    */
@@ -180,6 +211,7 @@ export const installTSTemplate = async ({
     framework,
     ui,
     observability,
+    vectorDb,
   });
 
   if (postInstallAction === "runApp" || postInstallAction === "dependencies") {
@@ -200,9 +232,16 @@ async function updatePackageJson({
   framework,
   ui,
   observability,
+  vectorDb,
 }: Pick<
   InstallTemplateArgs,
-  "root" | "appName" | "dataSources" | "framework" | "ui" | "observability"
+  | "root"
+  | "appName"
+  | "dataSources"
+  | "framework"
+  | "ui"
+  | "observability"
+  | "vectorDb"
 > & {
   relativeEngineDestPath: string;
 }): Promise<any> {
@@ -249,6 +288,34 @@ async function updatePackageJson({
     };
   }
 
+  if (vectorDb === "pg") {
+    packageJson.dependencies = {
+      ...packageJson.dependencies,
+      pg: "^8.12.0",
+      pgvector: "^0.2.0",
+    };
+  }
+
+  if (vectorDb === "qdrant") {
+    packageJson.dependencies = {
+      ...packageJson.dependencies,
+      "@qdrant/js-client-rest": "^1.11.0",
+    };
+  }
+  if (vectorDb === "mongo") {
+    packageJson.dependencies = {
+      ...packageJson.dependencies,
+      mongodb: "^6.7.0",
+    };
+  }
+
+  if (vectorDb === "milvus") {
+    packageJson.dependencies = {
+      ...packageJson.dependencies,
+      "@zilliz/milvus2-sdk-node": "^2.4.6",
+    };
+  }
+
   if (observability === "traceloop") {
     packageJson.dependencies = {
       ...packageJson.dependencies,

index.ts

@@ -90,6 +90,20 @@ const program = new Commander.Command(packageJson.name)
     `
 
   Select to use an example PDF as data source.
+`,
+  )
+  .option(
+    "--web-source <url>",
+    `
+  
+  Specify a website URL to use as a data source.
+`,
+  )
+  .option(
+    "--db-source <connection-string>",
+    `
+  
+  Specify a database connection string to use as a data source.
 `,
   )
   .option(
@@ -215,6 +229,27 @@ if (process.argv.includes("--no-files")) {
     },
     EXAMPLE_FILE,
   ];
+} else if (process.argv.includes("--web-source")) {
+  program.dataSources = [
+    {
+      type: "web",
+      config: {
+        baseUrl: program.webSource,
+        prefix: program.webSource,
+        depth: 1,
+      },
+    },
+  ];
+} else if (process.argv.includes("--db-source")) {
+  program.dataSources = [
+    {
+      type: "db",
+      config: {
+        uri: program.dbSource,
+        queries: program.dbQuery || "SELECT * FROM mytable",
+      },
+    },
+  ];
 }
 
 const packageManager = !!program.useNpm

package.json

@@ -1,6 +1,6 @@
 {
   "name": "create-llama",
-  "version": "0.2.6",
+  "version": "0.2.13",
   "description": "Create LlamaIndex-powered apps with one command",
   "keywords": [
     "rag",
@@ -25,6 +25,8 @@
     "clean": "rimraf --glob ./dist ./templates/**/__pycache__ ./templates/**/node_modules ./templates/**/poetry.lock",
     "dev": "ncc build ./index.ts -w -o dist/",
     "e2e": "playwright test",
+    "e2e:python": "playwright test e2e/shared e2e/python",
+    "e2e:typescript": "playwright test e2e/shared e2e/typescript",
     "format": "prettier --ignore-unknown --cache --check .",
     "format:write": "prettier --ignore-unknown --write .",
     "lint": "eslint . --ignore-pattern dist --ignore-pattern e2e/cache",

questions.ts

@@ -141,12 +141,10 @@ export const getDataSourceChoices = (
     });
   }
   if (selectedDataSource === undefined || selectedDataSource.length === 0) {
-    if (template !== "multiagent") {
-      choices.push({
-        title: "No datasource",
-        value: "none",
-      });
-    }
+    choices.push({
+      title: "No datasource",
+      value: "none",
+    });
     choices.push({
       title:
         process.platform !== "linux"
@@ -410,10 +408,7 @@ export const askQuestions = async (
     return; // early return - no further questions needed for llamapack projects
   }
 
-  if (program.template === "multiagent") {
-    // TODO: multi-agents currently only supports FastAPI
-    program.framework = preferences.framework = "fastapi";
-  } else if (program.template === "extractor") {
+  if (program.template === "extractor") {
     // Extractor template only supports FastAPI, empty data sources, and llamacloud
     // So we just use example file for extractor template, this allows user to choose vector database later
     program.dataSources = [EXAMPLE_FILE];
@@ -737,8 +732,10 @@ export const askQuestions = async (
     }
   }
 
-  if (!program.tools && program.template === "streaming") {
-    // TODO: allow to select tools also for multi-agent framework
+  if (
+    !program.tools &&
+    (program.template === "streaming" || program.template === "multiagent")
+  ) {
     if (ciInfo.isCI) {
       program.tools = getPrefOrDefault("tools");
     } else {

templates/components/engines/python/agent/engine.py

@@ -8,7 +8,7 @@ from llama_index.core.settings import Settings
 from llama_index.core.tools.query_engine import QueryEngineTool
 
 
-def get_chat_engine(filters=None, params=None, event_handlers=None):
+def get_chat_engine(filters=None, params=None, event_handlers=None, **kwargs):
     system_prompt = os.getenv("SYSTEM_PROMPT")
     top_k = int(os.getenv("TOP_K", 0))
     tools = []

templates/components/engines/python/agent/tools/__init__.py

@@ -1,8 +1,9 @@
-import os
-import yaml
 import importlib
-from llama_index.core.tools.tool_spec.base import BaseToolSpec
+import os
+
+import yaml
 from llama_index.core.tools.function_tool import FunctionTool
+from llama_index.core.tools.tool_spec.base import BaseToolSpec
 
 
 class ToolType:
@@ -40,14 +41,26 @@ class ToolFactory:
             raise ValueError(f"Failed to load tool {tool_name}: {e}")
 
     @staticmethod
-    def from_env() -> list[FunctionTool]:
-        tools = []
+    def from_env(
+        map_result: bool = False,
+    ) -> list[FunctionTool] | dict[str, FunctionTool]:
+        """
+        Load tools from the configured file.
+        Params:
+            - use_map: if True, return map of tool name and the tool itself
+        """
+        if map_result:
+            tools = {}
+        else:
+            tools = []
         if os.path.exists("config/tools.yaml"):
             with open("config/tools.yaml", "r") as f:
                 tool_configs = yaml.safe_load(f)
                 for tool_type, config_entries in tool_configs.items():
                     for tool_name, config in config_entries.items():
-                        tools.extend(
-                            ToolFactory.load_tools(tool_type, tool_name, config)
-                        )
+                        tool = ToolFactory.load_tools(tool_type, tool_name, config)
+                        if map_result:
+                            tools[tool_name] = tool
+                        else:
+                            tools.extend(tool)
         return tools

templates/components/engines/python/agent/tools/document_generator.py

@@ -0,0 +1,229 @@
+import logging
+import os
+import re
+from enum import Enum
+from io import BytesIO
+
+from llama_index.core.tools.function_tool import FunctionTool
+
+OUTPUT_DIR = "output/tools"
+
+
+class DocumentType(Enum):
+    PDF = "pdf"
+    HTML = "html"
+
+
+COMMON_STYLES = """
+body {
+    font-family: Arial, sans-serif;
+    line-height: 1.3;
+    color: #333;
+}
+h1, h2, h3, h4, h5, h6 {
+    margin-top: 1em;
+    margin-bottom: 0.5em;
+}
+p {
+    margin-bottom: 0.7em;
+}
+code {
+    background-color: #f4f4f4;
+    padding: 2px 4px;
+    border-radius: 4px;
+}
+pre {
+    background-color: #f4f4f4;
+    padding: 10px;
+    border-radius: 4px;
+    overflow-x: auto;
+}
+table {
+    border-collapse: collapse;
+    width: 100%;
+    margin-bottom: 1em;
+}
+th, td {
+    border: 1px solid #ddd;
+    padding: 8px;
+    text-align: left;
+}
+th {
+    background-color: #f2f2f2;
+    font-weight: bold;
+}
+"""
+
+HTML_SPECIFIC_STYLES = """
+body {
+    max-width: 800px;
+    margin: 0 auto;
+    padding: 20px;
+}
+"""
+
+PDF_SPECIFIC_STYLES = """
+@page {
+    size: letter;
+    margin: 2cm;
+}
+body {
+    font-size: 11pt;
+}
+h1 { font-size: 18pt; }
+h2 { font-size: 16pt; }
+h3 { font-size: 14pt; }
+h4, h5, h6 { font-size: 12pt; }
+pre, code {
+    font-family: Courier, monospace;
+    font-size: 0.9em;
+}
+"""
+
+HTML_TEMPLATE = """
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <style>
+        {common_styles}
+        {specific_styles}
+    </style>
+</head>
+<body>
+    {content}
+</body>
+</html>
+"""
+
+
+class DocumentGenerator:
+    @classmethod
+    def _generate_html_content(cls, original_content: str) -> str:
+        """
+        Generate HTML content from the original markdown content.
+        """
+        try:
+            import markdown
+        except ImportError:
+            raise ImportError(
+                "Failed to import required modules. Please install markdown."
+            )
+
+        # Convert markdown to HTML with fenced code and table extensions
+        html_content = markdown.markdown(
+            original_content, extensions=["fenced_code", "tables"]
+        )
+        return html_content
+
+    @classmethod
+    def _generate_pdf(cls, html_content: str) -> BytesIO:
+        """
+        Generate a PDF from the HTML content.
+        """
+        try:
+            from xhtml2pdf import pisa
+        except ImportError:
+            raise ImportError(
+                "Failed to import required modules. Please install xhtml2pdf."
+            )
+
+        pdf_html = HTML_TEMPLATE.format(
+            common_styles=COMMON_STYLES,
+            specific_styles=PDF_SPECIFIC_STYLES,
+            content=html_content,
+        )
+
+        buffer = BytesIO()
+        pdf = pisa.pisaDocument(
+            BytesIO(pdf_html.encode("UTF-8")), buffer, encoding="UTF-8"
+        )
+
+        if pdf.err:
+            logging.error(f"PDF generation failed: {pdf.err}")
+            raise ValueError("PDF generation failed")
+
+        buffer.seek(0)
+        return buffer
+
+    @classmethod
+    def _generate_html(cls, html_content: str) -> str:
+        """
+        Generate a complete HTML document with the given HTML content.
+        """
+        return HTML_TEMPLATE.format(
+            common_styles=COMMON_STYLES,
+            specific_styles=HTML_SPECIFIC_STYLES,
+            content=html_content,
+        )
+
+    @classmethod
+    def generate_document(
+        cls, original_content: str, document_type: str, file_name: str
+    ) -> str:
+        """
+        To generate document as PDF or HTML file.
+        Parameters:
+            original_content: str (markdown style)
+            document_type: str (pdf or html) specify the type of the file format based on the use case
+            file_name: str (name of the document file) must be a valid file name, no extensions needed
+        Returns:
+            str (URL to the document file): A file URL ready to serve.
+        """
+        try:
+            document_type = DocumentType(document_type.lower())
+        except ValueError:
+            raise ValueError(
+                f"Invalid document type: {document_type}. Must be 'pdf' or 'html'."
+            )
+        # Always generate html content first
+        html_content = cls._generate_html_content(original_content)
+
+        # Based on the type of document, generate the corresponding file
+        if document_type == DocumentType.PDF:
+            content = cls._generate_pdf(html_content)
+            file_extension = "pdf"
+        elif document_type == DocumentType.HTML:
+            content = BytesIO(cls._generate_html(html_content).encode("utf-8"))
+            file_extension = "html"
+        else:
+            raise ValueError(f"Unexpected document type: {document_type}")
+
+        file_name = cls._validate_file_name(file_name)
+        file_path = os.path.join(OUTPUT_DIR, f"{file_name}.{file_extension}")
+
+        cls._write_to_file(content, file_path)
+
+        file_url = f"{os.getenv('FILESERVER_URL_PREFIX')}/{file_path}"
+        return file_url
+
+    @staticmethod
+    def _write_to_file(content: BytesIO, file_path: str):
+        """
+        Write the content to a file.
+        """
+        try:
+            os.makedirs(os.path.dirname(file_path), exist_ok=True)
+            with open(file_path, "wb") as file:
+                file.write(content.getvalue())
+        except Exception as e:
+            raise e
+
+    @staticmethod
+    def _validate_file_name(file_name: str) -> str:
+        """
+        Validate the file name.
+        """
+        # Don't allow directory traversal
+        if os.path.isabs(file_name):
+            raise ValueError("File name is not allowed.")
+        # Don't allow special characters
+        if re.match(r"^[a-zA-Z0-9_.-]+$", file_name):
+            return file_name
+        else:
+            raise ValueError("File name is not allowed to contain special characters.")
+
+
+def get_tools(**kwargs):
+    return [FunctionTool.from_defaults(DocumentGenerator.generate_document)]

templates/components/engines/python/agent/tools/duckduckgo.py

@@ -32,5 +32,37 @@ def duckduckgo_search(
     return results
 
 
+def duckduckgo_image_search(
+    query: str,
+    region: str = "wt-wt",
+    max_results: int = 10,
+):
+    """
+    Use this function to search for images in DuckDuckGo.
+    Args:
+        query (str): The query to search in DuckDuckGo.
+        region Optional(str): The region to be used for the search in [country-language] convention, ex us-en, uk-en, ru-ru, etc...
+        max_results Optional(int): The maximum number of results to be returned. Default is 10.
+    """
+    try:
+        from duckduckgo_search import DDGS
+    except ImportError:
+        raise ImportError(
+            "duckduckgo_search package is required to use this function."
+            "Please install it by running: `poetry add duckduckgo_search` or `pip install duckduckgo_search`"
+        )
+    params = {
+        "keywords": query,
+        "region": region,
+        "max_results": max_results,
+    }
+    with DDGS() as ddg:
+        results = list(ddg.images(**params))
+    return results
+
+
 def get_tools(**kwargs):
-    return [FunctionTool.from_defaults(duckduckgo_search)]
+    return [
+        FunctionTool.from_defaults(duckduckgo_search),
+        FunctionTool.from_defaults(duckduckgo_image_search),
+    ]

templates/components/engines/python/agent/tools/img_gen.py

@@ -1,10 +1,11 @@
+import logging
 import os
 import uuid
-import logging
-import requests
 from typing import Optional
-from pydantic import BaseModel, Field
+
+import requests
 from llama_index.core.tools import FunctionTool
+from pydantic import BaseModel, Field
 
 logger = logging.getLogger(__name__)
 
@@ -26,7 +27,7 @@ class ImageGeneratorToolOutput(BaseModel):
 
 class ImageGeneratorTool:
     _IMG_OUTPUT_FORMAT = "webp"
-    _IMG_OUTPUT_DIR = "output/tool"
+    _IMG_OUTPUT_DIR = "output/tools"
     _IMG_GEN_API = "https://api.stability.ai/v2beta/stable-image/generate/core"
 
     def __init__(self, api_key: str = None):

templates/components/engines/python/agent/tools/interpreter.py

@@ -1,13 +1,13 @@
-import os
-import logging
 import base64
+import logging
+import os
 import uuid
-from pydantic import BaseModel
-from typing import List, Dict, Optional
-from llama_index.core.tools import FunctionTool
+from typing import Dict, List, Optional
+
 from e2b_code_interpreter import CodeInterpreter
 from e2b_code_interpreter.models import Logs
-
+from llama_index.core.tools import FunctionTool
+from pydantic import BaseModel
 
 logger = logging.getLogger(__name__)
 
@@ -26,7 +26,7 @@ class E2BToolOutput(BaseModel):
 
 
 class E2BCodeInterpreter:
-    output_dir = "output/tool"
+    output_dir = "output/tools"
 
     def __init__(self, api_key: str = None):
         if api_key is None:

templates/components/engines/python/chat/engine.py

@@ -9,7 +9,7 @@ from llama_index.core.memory import ChatMemoryBuffer
 from llama_index.core.settings import Settings
 
 
-def get_chat_engine(filters=None, params=None, event_handlers=None):
+def get_chat_engine(filters=None, params=None, event_handlers=None, **kwargs):
     system_prompt = os.getenv("SYSTEM_PROMPT")
     citation_prompt = os.getenv("SYSTEM_CITATION_PROMPT", None)
     top_k = int(os.getenv("TOP_K", 0))

templates/components/engines/typescript/agent/chat.ts

@@ -1,4 +1,9 @@
-import { BaseToolWithCall, OpenAIAgent, QueryEngineTool } from "llamaindex";
+import {
+  BaseToolWithCall,
+  ChatEngine,
+  OpenAIAgent,
+  QueryEngineTool,
+} from "llamaindex";
 import fs from "node:fs/promises";
 import path from "node:path";
 import { getDataSource } from "./index";
@@ -37,8 +42,10 @@ export async function createChatEngine(documentIds?: string[], params?: any) {
     tools.push(...(await createTools(toolConfig)));
   }
 
-  return new OpenAIAgent({
+  const agent = new OpenAIAgent({
     tools,
     systemPrompt: process.env.SYSTEM_PROMPT,
-  });
+  }) as unknown as ChatEngine;
+
+  return agent;
 }

templates/components/engines/typescript/agent/tools/document-generator.ts

@@ -0,0 +1,142 @@
+import { JSONSchemaType } from "ajv";
+import { BaseTool, ToolMetadata } from "llamaindex";
+import { marked } from "marked";
+import path from "node:path";
+import { saveDocument } from "../../llamaindex/documents/helper";
+
+const OUTPUT_DIR = "output/tools";
+
+type DocumentParameter = {
+  originalContent: string;
+  fileName: string;
+};
+
+const DEFAULT_METADATA: ToolMetadata<JSONSchemaType<DocumentParameter>> = {
+  name: "document_generator",
+  description:
+    "Generate HTML document from markdown content. Return a file url to the document",
+  parameters: {
+    type: "object",
+    properties: {
+      originalContent: {
+        type: "string",
+        description: "The original markdown content to convert.",
+      },
+      fileName: {
+        type: "string",
+        description: "The name of the document file (without extension).",
+      },
+    },
+    required: ["originalContent", "fileName"],
+  },
+};
+
+const COMMON_STYLES = `
+  body {
+    font-family: Arial, sans-serif;
+    line-height: 1.3;
+    color: #333;
+  }
+  h1, h2, h3, h4, h5, h6 {
+    margin-top: 1em;
+    margin-bottom: 0.5em;
+  }
+  p {
+    margin-bottom: 0.7em;
+  }
+  code {
+    background-color: #f4f4f4;
+    padding: 2px 4px;
+    border-radius: 4px;
+  }
+  pre {
+    background-color: #f4f4f4;
+    padding: 10px;
+    border-radius: 4px;
+    overflow-x: auto;
+  }
+  table {
+    border-collapse: collapse;
+    width: 100%;
+    margin-bottom: 1em;
+  }
+  th, td {
+    border: 1px solid #ddd;
+    padding: 8px;
+    text-align: left;
+  }
+  th {
+    background-color: #f2f2f2;
+    font-weight: bold;
+  }
+  img {
+    max-width: 90%;
+    height: auto;
+    display: block;
+    margin: 1em auto;
+    border-radius: 10px;
+  }
+`;
+
+const HTML_SPECIFIC_STYLES = `
+  body {
+    max-width: 800px;
+    margin: 0 auto;
+    padding: 20px;
+  }
+`;
+
+const HTML_TEMPLATE = `
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <style>
+        ${COMMON_STYLES}
+        ${HTML_SPECIFIC_STYLES}
+    </style>
+</head>
+<body>
+    {{content}}
+</body>
+</html>
+`;
+
+export interface DocumentGeneratorParams {
+  metadata?: ToolMetadata<JSONSchemaType<DocumentParameter>>;
+}
+
+export class DocumentGenerator implements BaseTool<DocumentParameter> {
+  metadata: ToolMetadata<JSONSchemaType<DocumentParameter>>;
+
+  constructor(params: DocumentGeneratorParams) {
+    this.metadata = params.metadata ?? DEFAULT_METADATA;
+  }
+
+  private static async generateHtmlContent(
+    originalContent: string,
+  ): Promise<string> {
+    return await marked(originalContent);
+  }
+
+  private static generateHtmlDocument(htmlContent: string): string {
+    return HTML_TEMPLATE.replace("{{content}}", htmlContent);
+  }
+
+  async call(input: DocumentParameter): Promise<string> {
+    const { originalContent, fileName } = input;
+
+    const htmlContent =
+      await DocumentGenerator.generateHtmlContent(originalContent);
+    const fileContent = DocumentGenerator.generateHtmlDocument(htmlContent);
+
+    const filePath = path.join(OUTPUT_DIR, `${fileName}.html`);
+
+    return `URL: ${await saveDocument(filePath, fileContent)}`;
+  }
+}
+
+export function getTools(): BaseTool[] {
+  return [new DocumentGenerator({})];
+}

templates/components/engines/typescript/agent/tools/duckduckgo.ts

@@ -5,15 +5,19 @@ import { BaseTool, ToolMetadata } from "llamaindex";
 export type DuckDuckGoParameter = {
   query: string;
   region?: string;
+  maxResults?: number;
 };
 
 export type DuckDuckGoToolParams = {
   metadata?: ToolMetadata<JSONSchemaType<DuckDuckGoParameter>>;
 };
 
-const DEFAULT_META_DATA: ToolMetadata<JSONSchemaType<DuckDuckGoParameter>> = {
-  name: "duckduckgo",
-  description: "Use this function to search for any query in DuckDuckGo.",
+const DEFAULT_SEARCH_METADATA: ToolMetadata<
+  JSONSchemaType<DuckDuckGoParameter>
+> = {
+  name: "duckduckgo_search",
+  description:
+    "Use this function to search for information (only text) in the internet using DuckDuckGo.",
   parameters: {
     type: "object",
     properties: {
@@ -27,6 +31,12 @@ const DEFAULT_META_DATA: ToolMetadata<JSONSchemaType<DuckDuckGoParameter>> = {
           "Optional, The region to be used for the search in [country-language] convention, ex us-en, uk-en, ru-ru, etc...",
         nullable: true,
       },
+      maxResults: {
+        type: "number",
+        description:
+          "Optional, The maximum number of results to be returned. Default is 10.",
+        nullable: true,
+      },
     },
     required: ["query"],
   },
@@ -42,15 +52,18 @@ export class DuckDuckGoSearchTool implements BaseTool<DuckDuckGoParameter> {
   metadata: ToolMetadata<JSONSchemaType<DuckDuckGoParameter>>;
 
   constructor(params: DuckDuckGoToolParams) {
-    this.metadata = params.metadata ?? DEFAULT_META_DATA;
+    this.metadata = params.metadata ?? DEFAULT_SEARCH_METADATA;
   }
 
   async call(input: DuckDuckGoParameter) {
-    const { query, region } = input;
+    const { query, region, maxResults = 10 } = input;
     const options = region ? { region } : {};
+    // Temporarily sleep to reduce overloading the DuckDuckGo
+    await new Promise((resolve) => setTimeout(resolve, 1000));
+
     const searchResults = await search(query, options);
 
-    return searchResults.results.map((result) => {
+    return searchResults.results.slice(0, maxResults).map((result) => {
       return {
         title: result.title,
         description: result.description,
@@ -59,3 +72,7 @@ export class DuckDuckGoSearchTool implements BaseTool<DuckDuckGoParameter> {
     });
   }
 }
+
+export function getTools() {
+  return [new DuckDuckGoSearchTool({})];
+}

templates/components/engines/typescript/agent/tools/img-gen.ts

@@ -37,7 +37,7 @@ const DEFAULT_META_DATA: ToolMetadata<JSONSchemaType<ImgGeneratorParameter>> = {
 
 export class ImgGeneratorTool implements BaseTool<ImgGeneratorParameter> {
   readonly IMG_OUTPUT_FORMAT = "webp";
-  readonly IMG_OUTPUT_DIR = "output/tool";
+  readonly IMG_OUTPUT_DIR = "output/tools";
   readonly IMG_GEN_API =
     "https://api.stability.ai/v2beta/stable-image/generate/core";
 

templates/components/engines/typescript/agent/tools/index.ts

@@ -1,5 +1,9 @@
 import { BaseToolWithCall } from "llamaindex";
 import { ToolsFactory } from "llamaindex/tools/ToolsFactory";
+import {
+  DocumentGenerator,
+  DocumentGeneratorParams,
+} from "./document-generator";
 import { DuckDuckGoSearchTool, DuckDuckGoToolParams } from "./duckduckgo";
 import { ImgGeneratorTool, ImgGeneratorToolParams } from "./img-gen";
 import { InterpreterTool, InterpreterToolParams } from "./interpreter";
@@ -43,6 +47,9 @@ const toolFactory: Record<string, ToolCreator> = {
   img_gen: async (config: unknown) => {
     return [new ImgGeneratorTool(config as ImgGeneratorToolParams)];
   },
+  document_generator: async (config: unknown) => {
+    return [new DocumentGenerator(config as DocumentGeneratorParams)];
+  },
 };
 
 async function createLocalTools(

templates/components/engines/typescript/agent/tools/interpreter.ts

@@ -56,7 +56,7 @@ const DEFAULT_META_DATA: ToolMetadata<JSONSchemaType<InterpreterParameter>> = {
 };
 
 export class InterpreterTool implements BaseTool<InterpreterParameter> {
-  private readonly outputDir = "output/tool";
+  private readonly outputDir = "output/tools";
   private apiKey?: string;
   private fileServerURLPrefix?: string;
   metadata: ToolMetadata<JSONSchemaType<InterpreterParameter>>;

templates/components/llamaindex/typescript/documents/helper.ts

@@ -1,4 +1,5 @@
-import fs from "fs";
+import fs from "node:fs";
+import path from "node:path";
 import { getExtractors } from "../../engine/loader";
 
 const MIME_TYPE_TO_EXT: Record<string, string> = {
@@ -15,8 +16,12 @@ export async function storeAndParseFile(
   fileBuffer: Buffer,
   mimeType: string,
 ) {
+  const fileExt = MIME_TYPE_TO_EXT[mimeType];
+  if (!fileExt) throw new Error(`Unsupported document type: ${mimeType}`);
+
   const documents = await loadDocuments(fileBuffer, mimeType);
-  await saveDocument(filename, fileBuffer, mimeType);
+  const filepath = path.join(UPLOADED_FOLDER, filename);
+  await saveDocument(filepath, fileBuffer);
   for (const document of documents) {
     document.metadata = {
       ...document.metadata,
@@ -38,26 +43,31 @@ async function loadDocuments(fileBuffer: Buffer, mimeType: string) {
   return await reader.loadDataAsContent(fileBuffer);
 }
 
-async function saveDocument(
-  filename: string,
-  fileBuffer: Buffer,
-  mimeType: string,
-) {
-  const fileExt = MIME_TYPE_TO_EXT[mimeType];
-  if (!fileExt) throw new Error(`Unsupported document type: ${mimeType}`);
-
-  const filepath = `${UPLOADED_FOLDER}/${filename}`;
-  const fileurl = `${process.env.FILESERVER_URL_PREFIX}/${filepath}`;
-
-  if (!fs.existsSync(UPLOADED_FOLDER)) {
-    fs.mkdirSync(UPLOADED_FOLDER, { recursive: true });
+// Save document to file server and return the file url
+export async function saveDocument(filepath: string, content: string | Buffer) {
+  if (path.isAbsolute(filepath)) {
+    throw new Error("Absolute file paths are not allowed.");
+  }
+  const fileName = path.basename(filepath);
+  if (!/^[a-zA-Z0-9_.-]+$/.test(fileName)) {
+    throw new Error(
+      "File name is not allowed to contain any special characters.",
+    );
+  }
+  if (!process.env.FILESERVER_URL_PREFIX) {
+    throw new Error("FILESERVER_URL_PREFIX environment variable is not set.");
   }
-  await fs.promises.writeFile(filepath, fileBuffer);
 
-  console.log(`Saved document file to ${filepath}.\nURL: ${fileurl}`);
-  return {
-    filename,
-    filepath,
-    fileurl,
-  };
+  const dirPath = path.dirname(filepath);
+  await fs.promises.mkdir(dirPath, { recursive: true });
+
+  if (typeof content === "string") {
+    await fs.promises.writeFile(filepath, content, "utf-8");
+  } else {
+    await fs.promises.writeFile(filepath, content);
+  }
+
+  const fileurl = `${process.env.FILESERVER_URL_PREFIX}/${filepath}`;
+  console.log(`Saved document to ${filepath}. Reachable at URL: ${fileurl}`);
+  return fileurl;
 }

templates/components/llamaindex/typescript/streaming/stream.ts

@@ -1,57 +0,0 @@
-import {
-  StreamData,
-  createCallbacksTransformer,
-  createStreamDataTransformer,
-  trimStartOfStreamHelper,
-  type AIStreamCallbacksAndOptions,
-} from "ai";
-import { ChatMessage, EngineResponse } from "llamaindex";
-import { generateNextQuestions } from "./suggestion";
-
-export function LlamaIndexStream(
-  response: AsyncIterable<EngineResponse>,
-  data: StreamData,
-  chatHistory: ChatMessage[],
-  opts?: {
-    callbacks?: AIStreamCallbacksAndOptions;
-  },
-): ReadableStream<Uint8Array> {
-  return createParser(response, data, chatHistory)
-    .pipeThrough(createCallbacksTransformer(opts?.callbacks))
-    .pipeThrough(createStreamDataTransformer());
-}
-
-function createParser(
-  res: AsyncIterable<EngineResponse>,
-  data: StreamData,
-  chatHistory: ChatMessage[],
-) {
-  const it = res[Symbol.asyncIterator]();
-  const trimStartOfStream = trimStartOfStreamHelper();
-  let llmTextResponse = "";
-
-  return new ReadableStream<string>({
-    async pull(controller): Promise<void> {
-      const { value, done } = await it.next();
-      if (done) {
-        controller.close();
-        // LLM stream is done, generate the next questions with a new LLM call
-        chatHistory.push({ role: "assistant", content: llmTextResponse });
-        const questions: string[] = await generateNextQuestions(chatHistory);
-        if (questions.length > 0) {
-          data.appendMessageAnnotation({
-            type: "suggested_questions",
-            data: questions,
-          });
-        }
-        data.close();
-        return;
-      }
-      const text = trimStartOfStream(value.delta ?? "");
-      if (text) {
-        llmTextResponse += text;
-        controller.enqueue(text);
-      }
-    },
-  });
-}

templates/components/loaders/typescript/llama_parse/loader.ts

@@ -1,4 +1,4 @@
-import { LlamaParseReader } from "llamaindex/readers/LlamaParseReader";
+import { LlamaParseReader } from "llamaindex";
 import {
   FILE_EXT_TO_READER,
   SimpleDirectoryReader,

templates/components/multiagent/python/app/agents/multi.py

@@ -1,16 +1,14 @@
-import asyncio
 from typing import Any, List
 
-from llama_index.core.tools.types import ToolMetadata, ToolOutput
-from llama_index.core.tools.utils import create_schema_from_function
-from llama_index.core.workflow import Context, Workflow
-
+from app.agents.planner import StructuredPlannerAgent
 from app.agents.single import (
     AgentRunResult,
     ContextAwareTool,
     FunctionCallingAgent,
 )
-from app.agents.planner import StructuredPlannerAgent
+from llama_index.core.tools.types import ToolMetadata, ToolOutput
+from llama_index.core.tools.utils import create_schema_from_function
+from llama_index.core.workflow import Context, StopEvent, Workflow
 
 
 class AgentCallTool(ContextAwareTool):
@@ -27,18 +25,23 @@ class AgentCallTool(ContextAwareTool):
             name=name,
             description=(
                 f"Use this tool to delegate a sub task to the {agent.name} agent."
-                + (f" The agent is an {agent.role}." if agent.role else "")
+                + (
+                    f" The agent is an {agent.description}."
+                    if agent.description
+                    else ""
+                )
             ),
             fn_schema=fn_schema,
         )
 
     # overload the acall function with the ctx argument as it's needed for bubbling the events
     async def acall(self, ctx: Context, input: str) -> ToolOutput:
-        task = asyncio.create_task(self.agent.run(input=input))
+        handler = self.agent.run(input=input)
         # bubble all events while running the agent to the calling agent
-        async for ev in self.agent.stream_events():
-            ctx.write_event_to_stream(ev)
-        ret: AgentRunResult = await task
+        async for ev in handler.stream_events():
+            if type(ev) is not StopEvent:
+                ctx.write_event_to_stream(ev)
+        ret: AgentRunResult = await handler
         response = ret.response.message.content
         return ToolOutput(
             content=str(response),

templates/components/multiagent/python/app/agents/planner.py

@@ -1,8 +1,8 @@
-import asyncio
 import uuid
 from enum import Enum
 from typing import Any, AsyncGenerator, Dict, List, Optional, Tuple, Union
 
+from app.agents.single import AgentRunEvent, AgentRunResult, FunctionCallingAgent
 from llama_index.core.agent.runner.planner import (
     DEFAULT_INITIAL_PLAN_PROMPT,
     DEFAULT_PLAN_REFINE_PROMPT,
@@ -11,6 +11,7 @@ from llama_index.core.agent.runner.planner import (
     SubTask,
 )
 from llama_index.core.bridge.pydantic import ValidationError
+from llama_index.core.chat_engine.types import ChatMessage
 from llama_index.core.llms.function_calling import FunctionCallingLLM
 from llama_index.core.prompts import PromptTemplate
 from llama_index.core.settings import Settings
@@ -24,7 +25,17 @@ from llama_index.core.workflow import (
     step,
 )
 
-from app.agents.single import AgentRunEvent, AgentRunResult, FunctionCallingAgent
+INITIAL_PLANNER_PROMPT = """\
+Think step-by-step. Given a conversation, set of tools and a user request. Your responsibility is to create a plan to complete the task.
+The plan must adapt with the user request and the conversation. It's fine to just start with needed tasks first and asking user for the next step approval.
+
+The tools available are:
+{tools_str}
+
+Conversation: {chat_history}
+
+Overall Task: {task}
+"""
 
 
 class ExecutePlanEvent(Event):
@@ -64,14 +75,21 @@ class StructuredPlannerAgent(Workflow):
         tools: List[BaseTool] | None = None,
         timeout: float = 360.0,
         refine_plan: bool = False,
+        chat_history: Optional[List[ChatMessage]] = None,
         **kwargs: Any,
     ) -> None:
         super().__init__(*args, timeout=timeout, **kwargs)
         self.name = name
         self.refine_plan = refine_plan
+        self.chat_history = chat_history
 
         self.tools = tools or []
-        self.planner = Planner(llm=llm, tools=self.tools, verbose=self._verbose)
+        self.planner = Planner(
+            llm=llm,
+            tools=self.tools,
+            initial_plan_prompt=INITIAL_PLANNER_PROMPT,
+            verbose=self._verbose,
+        )
         # The executor is keeping the memory of all tool calls and decides to call the right tool for the task
         self.executor = FunctionCallingAgent(
             name="executor",
@@ -91,7 +109,9 @@ class StructuredPlannerAgent(Workflow):
         ctx.data["streaming"] = getattr(ev, "streaming", False)
         ctx.data["task"] = ev.input
 
-        plan_id, plan = await self.planner.create_plan(input=ev.input)
+        plan_id, plan = await self.planner.create_plan(
+            input=ev.input, chat_history=self.chat_history
+        )
         ctx.data["act_plan_id"] = plan_id
 
         # inform about the new plan
@@ -108,11 +128,12 @@ class StructuredPlannerAgent(Workflow):
             ctx.data["act_plan_id"]
         )
 
-        ctx.data["num_sub_tasks"] = len(upcoming_sub_tasks)
-        # send an event per sub task
-        events = [SubTaskEvent(sub_task=sub_task) for sub_task in upcoming_sub_tasks]
-        for event in events:
-            ctx.send_event(event)
+        if upcoming_sub_tasks:
+            # Execute only the first sub-task
+            # otherwise the executor will get over-lapping messages
+            # alternatively, we could use one executor for all sub tasks
+            next_sub_task = upcoming_sub_tasks[0]
+            return SubTaskEvent(sub_task=next_sub_task)
 
         return None
 
@@ -122,19 +143,19 @@ class StructuredPlannerAgent(Workflow):
     ) -> SubTaskResultEvent:
         if self._verbose:
             print(f"=== Executing sub task: {ev.sub_task.name} ===")
-        is_last_tasks = ctx.data["num_sub_tasks"] == self.get_remaining_subtasks(ctx)
+        is_last_tasks = self.get_remaining_subtasks(ctx) == 1
         # TODO: streaming only works without plan refining
         streaming = is_last_tasks and ctx.data["streaming"] and not self.refine_plan
-        task = asyncio.create_task(
-            self.executor.run(
-                input=ev.sub_task.input,
-                streaming=streaming,
-            )
+        handler = self.executor.run(
+            input=ev.sub_task.input,
+            streaming=streaming,
         )
         # bubble all events while running the executor to the planner
-        async for event in self.executor.stream_events():
-            ctx.write_event_to_stream(event)
-        result = await task
+        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)
+        result: AgentRunResult = await handler
         if self._verbose:
             print("=== Done executing sub task ===\n")
         self.planner.state.add_completed_sub_task(ctx.data["act_plan_id"], ev.sub_task)
@@ -144,22 +165,17 @@ class StructuredPlannerAgent(Workflow):
     async def gather_results(
         self, ctx: Context, ev: SubTaskResultEvent
     ) -> ExecutePlanEvent | StopEvent:
-        # wait for all sub tasks to finish
-        num_sub_tasks = ctx.data["num_sub_tasks"]
-        results = ctx.collect_events(ev, [SubTaskResultEvent] * num_sub_tasks)
-        if results is None:
-            return None
+        result = ev
 
         upcoming_sub_tasks = self.get_upcoming_sub_tasks(ctx)
         # if no more tasks to do, stop workflow and send result of last step
         if upcoming_sub_tasks == 0:
-            return StopEvent(result=results[-1].result)
+            return StopEvent(result=result.result)
 
         if self.refine_plan:
-            # store all results for refining the plan
+            # store the result for refining the plan
             ctx.data["results"] = ctx.data.get("results", {})
-            for result in results:
-                ctx.data["results"][result.sub_task.name] = result.result
+            ctx.data["results"][result.sub_task.name] = result.result
 
             new_plan = await self.planner.refine_plan(
                 ctx.data["task"], ctx.data["act_plan_id"], ctx.data["results"]
@@ -215,7 +231,9 @@ class Planner:
             plan_refine_prompt = PromptTemplate(plan_refine_prompt)
         self.plan_refine_prompt = plan_refine_prompt
 
-    async def create_plan(self, input: str) -> Tuple[str, Plan]:
+    async def create_plan(
+        self, input: str, chat_history: Optional[List[ChatMessage]] = None
+    ) -> Tuple[str, Plan]:
         tools = self.tools
         tools_str = ""
         for tool in tools:
@@ -227,6 +245,7 @@ class Planner:
                 self.initial_plan_prompt,
                 tools_str=tools_str,
                 task=input,
+                chat_history=chat_history,
             )
         except (ValueError, ValidationError):
             if self.verbose:

templates/components/multiagent/python/app/agents/single.py

@@ -5,10 +5,8 @@ from llama_index.core.llms import ChatMessage, ChatResponse
 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 import ToolOutput, ToolSelection
+from llama_index.core.tools import FunctionTool, ToolOutput, ToolSelection
 from llama_index.core.tools.types import BaseTool
-from llama_index.core.tools import FunctionTool
-
 from llama_index.core.workflow import (
     Context,
     Event,
@@ -64,14 +62,14 @@ class FunctionCallingAgent(Workflow):
         timeout: float = 360.0,
         name: str,
         write_events: bool = True,
-        role: Optional[str] = None,
+        description: str | None = None,
         **kwargs: Any,
     ) -> None:
         super().__init__(*args, verbose=verbose, timeout=timeout, **kwargs)
         self.tools = tools or []
         self.name = name
-        self.role = role
         self.write_events = write_events
+        self.description = description
 
         if llm is None:
             llm = Settings.llm

templates/components/multiagent/python/app/api/routers/chat.py

@@ -0,0 +1,46 @@
+import logging
+
+from app.api.routers.events import EventCallbackHandler
+from app.api.routers.models import (
+    ChatData,
+)
+from app.api.routers.vercel_response import VercelStreamResponse
+from app.engine import get_chat_engine
+from fastapi import APIRouter, BackgroundTasks, HTTPException, Request, status
+
+chat_router = r = APIRouter()
+
+logger = logging.getLogger("uvicorn")
+
+
+@r.post("")
+async def chat(
+    request: Request,
+    data: ChatData,
+    background_tasks: BackgroundTasks,
+):
+    try:
+        last_message_content = data.get_last_message_content()
+        messages = data.get_history_messages(include_agent_messages=True)
+
+        event_handler = EventCallbackHandler()
+        # 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
+        # TODO: use chat params
+        engine = get_chat_engine(chat_history=messages)
+
+        event_handler = engine.run(input=last_message_content, streaming=True)
+        return VercelStreamResponse(
+            request=request,
+            chat_data=data,
+            event_handler=event_handler,
+            events=engine.stream_events(),
+        )
+    except Exception as e:
+        logger.exception("Error in chat engine", exc_info=True)
+        raise HTTPException(
+            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
+            detail=f"Error in chat engine: {e}",
+        ) from e

templates/components/multiagent/python/app/api/routers/vercel_response.py

@@ -1,6 +1,6 @@
 import json
 import logging
-from asyncio import Task
+from abc import ABC
 from typing import AsyncGenerator, List
 
 from aiostream import stream
@@ -13,14 +13,88 @@ from fastapi.responses import StreamingResponse
 logger = logging.getLogger("uvicorn")
 
 
-class VercelStreamResponse(StreamingResponse):
+class VercelStreamResponse(StreamingResponse, ABC):
     """
-    Class to convert the response from the chat engine to the streaming format expected by Vercel
+    Base class to convert the response from the chat engine to the streaming format expected by Vercel
     """
 
     TEXT_PREFIX = "0:"
     DATA_PREFIX = "8:"
 
+    def __init__(self, request: Request, chat_data: ChatData, *args, **kwargs):
+        self.request = request
+
+        stream = self._create_stream(request, chat_data, *args, **kwargs)
+        content = self.content_generator(stream)
+
+        super().__init__(content=content)
+
+    async def content_generator(self, stream):
+        is_stream_started = False
+
+        async with stream.stream() as streamer:
+            async for output in streamer:
+                if not is_stream_started:
+                    is_stream_started = True
+                    # Stream a blank message to start the stream
+                    yield self.convert_text("")
+
+                yield output
+
+                if await self.request.is_disconnected():
+                    break
+
+    def _create_stream(
+        self,
+        request: Request,
+        chat_data: ChatData,
+        event_handler: AgentRunResult | AsyncGenerator,
+        events: AsyncGenerator[AgentRunEvent, None],
+        verbose: bool = True,
+    ):
+        # Yield the text response
+        async def _chat_response_generator():
+            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
+                    yield self.convert_text(token.delta)
+
+            # Generate next questions if next question prompt is configured
+            question_data = await self._generate_next_questions(
+                chat_data.messages, final_response
+            )
+            if question_data:
+                yield self.convert_data(question_data)
+
+            # TODO: stream sources
+
+        # Yield the events from the event handler
+        async def _event_generator():
+            async for event in events:
+                event_response = self._event_to_response(event)
+                if verbose:
+                    logger.debug(event_response)
+                if event_response is not None:
+                    yield self.convert_data(event_response)
+
+        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
@@ -32,82 +106,6 @@ class VercelStreamResponse(StreamingResponse):
         data_str = json.dumps(data)
         return f"{cls.DATA_PREFIX}[{data_str}]\n"
 
-    def __init__(
-        self,
-        request: Request,
-        task: Task[AgentRunResult | AsyncGenerator],
-        events: AsyncGenerator[AgentRunEvent, None],
-        chat_data: ChatData,
-        verbose: bool = True,
-    ):
-        content = VercelStreamResponse.content_generator(
-            request, task, events, chat_data, verbose
-        )
-        super().__init__(content=content)
-
-    @classmethod
-    async def content_generator(
-        cls,
-        request: Request,
-        task: Task[AgentRunResult | AsyncGenerator],
-        events: AsyncGenerator[AgentRunEvent, None],
-        chat_data: ChatData,
-        verbose: bool = True,
-    ):
-        # Yield the text response
-        async def _chat_response_generator():
-            result = await task
-            final_response = ""
-
-            if isinstance(result, AgentRunResult):
-                for token in result.response.message.content:
-                    final_response += token
-                    yield cls.convert_text(token)
-
-            if isinstance(result, AsyncGenerator):
-                async for token in result:
-                    final_response += token.delta
-                    yield cls.convert_text(token.delta)
-
-            # Generate next questions if next question prompt is configured
-            question_data = await cls._generate_next_questions(
-                chat_data.messages, final_response
-            )
-            if question_data:
-                yield cls.convert_data(question_data)
-
-            # TODO: stream sources
-
-        # Yield the events from the event handler
-        async def _event_generator():
-            async for event in events():
-                event_response = cls._event_to_response(event)
-                if verbose:
-                    logger.debug(event_response)
-                if event_response is not None:
-                    yield cls.convert_data(event_response)
-
-        combine = stream.merge(_chat_response_generator(), _event_generator())
-
-        is_stream_started = False
-        async with combine.stream() as streamer:
-            if not is_stream_started:
-                is_stream_started = True
-                # Stream a blank message to start the stream
-                yield cls.convert_text("")
-
-            async for output in streamer:
-                yield output
-                if await request.is_disconnected():
-                    break
-
-    @staticmethod
-    def _event_to_response(event: AgentRunEvent) -> dict:
-        return {
-            "type": "agent",
-            "data": {"agent": event.name, "text": event.msg},
-        }
-
     @staticmethod
     async def _generate_next_questions(chat_history: List[Message], response: str):
         questions = await NextQuestionSuggestion.suggest_next_questions(

templates/components/multiagent/python/app/engine/engine.py

@@ -1,20 +1,20 @@
 import logging
+import os
 from typing import List, Optional
+
 from app.examples.choreography import create_choreography
 from app.examples.orchestrator import create_orchestrator
 from app.examples.workflow import create_workflow
-
-
-from llama_index.core.workflow import Workflow
 from llama_index.core.chat_engine.types import ChatMessage
-
-
-import os
+from llama_index.core.workflow import Workflow
 
 logger = logging.getLogger("uvicorn")
 
 
-def create_agent(chat_history: Optional[List[ChatMessage]] = None) -> Workflow:
+def get_chat_engine(
+    chat_history: Optional[List[ChatMessage]] = None, **kwargs
+) -> Workflow:
+    # TODO: the EXAMPLE_TYPE could be passed as a chat config parameter?
     agent_type = os.getenv("EXAMPLE_TYPE", "").lower()
     match agent_type:
         case "choreography":

templates/components/multiagent/python/app/examples/choreography.py

@@ -0,0 +1,32 @@
+from textwrap import dedent
+from typing import List, Optional
+
+from app.agents.multi import AgentCallingAgent
+from app.agents.single import FunctionCallingAgent
+from app.examples.publisher import create_publisher
+from app.examples.researcher import create_researcher
+from llama_index.core.chat_engine.types import ChatMessage
+
+
+def create_choreography(chat_history: Optional[List[ChatMessage]] = None):
+    researcher = create_researcher(chat_history)
+    publisher = create_publisher(chat_history)
+    reviewer = FunctionCallingAgent(
+        name="reviewer",
+        description="expert in reviewing blog posts, needs a written post to review",
+        system_prompt="You are an expert in reviewing blog posts. You are given a task to review a blog post. Review the post for logical inconsistencies, ask critical questions, and provide suggestions for improvement. Furthermore, proofread the post for grammar and spelling errors. If the post is good, you can say 'The post is good.'",
+        chat_history=chat_history,
+    )
+    return AgentCallingAgent(
+        name="writer",
+        agents=[researcher, reviewer, publisher],
+        description="expert in writing blog posts, needs researched information and images to write a blog post",
+        system_prompt=dedent("""
+            You are an expert in writing blog posts. You are given a task to write a blog post. Before starting to write the post, consult the researcher agent to get the information you need. Don't make up any information yourself.
+            After creating a draft for the post, send it to the reviewer agent to receive feedback and make sure to incorporate the feedback from the reviewer.
+            You can consult the reviewer and researcher a maximum of two times. Your output should contain only the blog post.
+            Finally, always request the publisher to create a document (PDF, HTML) and publish the blog post.
+        """),
+        # TODO: add chat_history support to AgentCallingAgent
+        # chat_history=chat_history,
+    )

templates/components/multiagent/python/app/examples/orchestrator.py

@@ -0,0 +1,40 @@
+from textwrap import dedent
+from typing import List, Optional
+
+from app.agents.multi import AgentOrchestrator
+from app.agents.single import FunctionCallingAgent
+from app.examples.publisher import create_publisher
+from app.examples.researcher import create_researcher
+from llama_index.core.chat_engine.types import ChatMessage
+
+
+def create_orchestrator(chat_history: Optional[List[ChatMessage]] = None):
+    researcher = create_researcher(chat_history)
+    writer = FunctionCallingAgent(
+        name="writer",
+        description="expert in writing blog posts, need information and images to write a post",
+        system_prompt=dedent("""
+            You are an expert in writing blog posts.
+            You are given a task to write a blog post. Do not make up any information yourself.
+            If you don't have the necessary information to write a blog post, reply "I need information about the topic to write the blog post".
+            If you need to use images, reply "I need images about the topic to write the blog post". Do not use any dummy images made up by you.
+            If you have all the information needed, write the blog post.
+        """),
+        chat_history=chat_history,
+    )
+    reviewer = FunctionCallingAgent(
+        name="reviewer",
+        description="expert in reviewing blog posts, needs a written blog post to review",
+        system_prompt=dedent("""
+            You are an expert in reviewing blog posts. You are given a task to review a blog post. Review the post and fix any issues found yourself. You must output a final blog post.
+            A post must include at least one valid image. If not, reply "I need images about the topic to write the blog post". An image URL starting with "example" or "your website" is not valid.
+            Especially check for logical inconsistencies and proofread the post for grammar and spelling errors.
+        """),
+        chat_history=chat_history,
+    )
+    publisher = create_publisher(chat_history)
+    return AgentOrchestrator(
+        agents=[writer, reviewer, researcher, publisher],
+        refine_plan=False,
+        chat_history=chat_history,
+    )

templates/components/multiagent/python/app/examples/publisher.py

@@ -0,0 +1,35 @@
+from textwrap import dedent
+from typing import List, Tuple
+
+from app.agents.single import FunctionCallingAgent
+from app.engine.tools import ToolFactory
+from llama_index.core.chat_engine.types import ChatMessage
+from llama_index.core.tools import FunctionTool
+
+
+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"])
+        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.
+        """)
+        description = "Expert in publishing the blog post, able to publish the blog post in PDF or HTML format."
+    else:
+        prompt_instructions = "You don't have a tool to generate document. Please reply the content directly."
+        description = "Expert in publishing the blog post"
+    return tools, prompt_instructions, description
+
+
+def create_publisher(chat_history: List[ChatMessage]):
+    tools, prompt_instructions, description = get_publisher_tools()
+    return FunctionCallingAgent(
+        name="publisher",
+        tools=tools,
+        description=description,
+        system_prompt=prompt_instructions,
+        chat_history=chat_history,
+    )

templates/components/multiagent/python/app/examples/researcher.py

@@ -0,0 +1,82 @@
+import os
+from textwrap import dedent
+from typing import List
+
+from app.agents.single import FunctionCallingAgent
+from app.engine.index import get_index
+from app.engine.tools import ToolFactory
+from llama_index.core.chat_engine.types import ChatMessage
+from llama_index.core.tools import QueryEngineTool, ToolMetadata
+
+
+def _create_query_engine_tool() -> QueryEngineTool:
+    """
+    Provide an agent worker that can be used to query the index.
+    """
+    index = get_index()
+    if index is None:
+        return None
+    top_k = int(os.getenv("TOP_K", 0))
+    query_engine = index.as_query_engine(
+        **({"similarity_top_k": top_k} if top_k != 0 else {})
+    )
+    return QueryEngineTool(
+        query_engine=query_engine,
+        metadata=ToolMetadata(
+            name="query_index",
+            description="""
+                Use this tool to retrieve information about the text corpus from the index.
+            """,
+        ),
+    )
+
+
+def _get_research_tools() -> QueryEngineTool:
+    """
+    Researcher take responsibility for retrieving information.
+    Try init wikipedia or duckduckgo tool if available.
+    """
+    tools = []
+    query_engine_tool = _create_query_engine_tool()
+    if query_engine_tool is not None:
+        tools.append(query_engine_tool)
+    researcher_tool_names = ["duckduckgo", "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)
+    return tools
+
+
+def create_researcher(chat_history: List[ChatMessage]):
+    """
+    Researcher is an agent that take responsibility for using tools to complete a given task.
+    """
+    tools = _get_research_tools()
+    return FunctionCallingAgent(
+        name="researcher",
+        tools=tools,
+        description="expert in retrieving any unknown content or searching for images from the internet",
+        system_prompt=dedent("""
+            You are a researcher agent. You are given a research task.
+            
+            If the conversation already includes the information and there is no new request for additional information from the user, you should return the appropriate content to the writer.
+            Otherwise, you must use tools to retrieve information or images needed for the task.
+
+            It's normal for the task to include some ambiguity. You must always think carefully about the context of the user's request to understand what are the main content needs to be retrieved.
+            Example:
+                Request: "Create a blog post about the history of the internet, write in English and publish in PDF format."
+                ->Though: The main content is "history of the internet", while "write in English and publish in PDF format" is a requirement for other agents.
+                Your task: Look for information in English about the history of the Internet.
+                This is not your task: Create a blog post or look for how to create a PDF.
+
+                Next request: "Publish the blog post in HTML format."
+                ->Though: User just asking for a format change, the previous content is still valid.
+                Your task: Return the previous content of the post to the writer. No need to do any research.
+                This is not your task: Look for how to create an HTML file.
+
+            If you use the tools but don't find any related information, please return "I didn't find any new information for {the topic}." along with the content you found. Don't try to make up information yourself.
+            If the request doesn't need any new information because it was in the conversation history, please return "The task doesn't need any new information. Please reuse the existing content in the conversation history."
+        """),
+        chat_history=chat_history,
+    )

templates/components/multiagent/python/app/examples/workflow.py

@@ -0,0 +1,250 @@
+from textwrap import dedent
+from typing import AsyncGenerator, List, Optional
+from llama_index.core.settings import Settings
+from llama_index.core.prompts import PromptTemplate
+
+from app.agents.single import AgentRunEvent, AgentRunResult, FunctionCallingAgent
+from app.examples.publisher import create_publisher
+from app.examples.researcher import create_researcher
+from llama_index.core.chat_engine.types import ChatMessage
+from llama_index.core.workflow import (
+    Context,
+    Event,
+    StartEvent,
+    StopEvent,
+    Workflow,
+    step,
+)
+
+
+def create_workflow(chat_history: Optional[List[ChatMessage]] = None):
+    researcher = create_researcher(
+        chat_history=chat_history,
+    )
+    publisher = create_publisher(
+        chat_history=chat_history,
+    )
+    writer = FunctionCallingAgent(
+        name="writer",
+        description="expert in writing blog posts, need information and images to write a post.",
+        system_prompt=dedent(
+            """
+            You are an expert in writing blog posts.
+            You are given the task of writing a blog post based on research content provided by the researcher agent. Do not invent any information yourself. 
+            It's important to read the entire conversation history to write the blog post accurately.
+            If you receive a review from the reviewer, update the post according to the feedback and return the new post content.
+            If the user requests an update with new information but no research content is provided, you must respond with: "I don't have any research content to write about."
+            If the content is not valid (e.g., broken link, broken image, etc.), do not use it.
+            It's normal for the task to include some ambiguity, so you must define the user's initial request to write the post correctly.
+            If you update the post based on the reviewer's feedback, first explain what changes you made to the post, then provide the new post content. Do not include the reviewer's comments.
+            Example:
+                Task: "Here is the information I found about the history of the internet: 
+                Create a blog post about the history of the internet, write in English, and publish in PDF format."
+                -> Your task: Use the research content {...} to write a blog post in English.
+                -> This is not your task: Create a PDF
+                Please note that a localhost link is acceptable, but dummy links like "example.com" or "your-website.com" are not valid.
+        """
+        ),
+        chat_history=chat_history,
+    )
+    reviewer = FunctionCallingAgent(
+        name="reviewer",
+        description="expert in reviewing blog posts, needs a written blog post to review.",
+        system_prompt=dedent(
+            """
+            You are an expert in reviewing blog posts.
+            You are given a task to review a blog post. As a reviewer, it's important that your review aligns with the user's request. Please focus on the user's request when reviewing the post.
+            Review the post for logical inconsistencies, ask critical questions, and provide suggestions for improvement.
+            Furthermore, proofread the post for grammar and spelling errors.
+            Only if the post is good enough for publishing should you return 'The post is good.' In all other cases, return your review.
+            It's normal for the task to include some ambiguity, so you must define the user's initial request to review the post correctly.
+            Please note that a localhost link is acceptable, but dummy links like "example.com" or "your-website.com" are not valid.
+            Example:
+                Task: "Create a blog post about the history of the internet, write in English and publish in PDF format."
+                -> Your task: Review whether the main content of the post is about the history of the internet and if it is written in English.
+                -> This is not your task: Create blog post, create PDF, write in English.
+        """
+        ),
+        chat_history=chat_history,
+    )
+    workflow = BlogPostWorkflow(
+        timeout=360, chat_history=chat_history
+    )  # Pass chat_history here
+    workflow.add_workflows(
+        researcher=researcher,
+        writer=writer,
+        reviewer=reviewer,
+        publisher=publisher,
+    )
+    return workflow
+
+
+class ResearchEvent(Event):
+    input: str
+
+
+class WriteEvent(Event):
+    input: str
+    is_good: bool = False
+
+
+class ReviewEvent(Event):
+    input: str
+
+
+class PublishEvent(Event):
+    input: str
+
+
+class BlogPostWorkflow(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 | PublishEvent:
+        # 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 PublishEvent(
+                input=f"Please publish content 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:
+        prompt_template = PromptTemplate(
+            "Given the following chat history and new task, decide whether to publish based on existing information.\n"
+            "Chat history:\n{chat_history}\n"
+            "New task: {input}\n"
+            "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
+    ) -> WriteEvent:
+        result: AgentRunResult = await self.run_agent(ctx, researcher, ev.input)
+        content = result.response.message.content
+        return WriteEvent(
+            input=f"Write a blog post given this task: {ctx.data['task']} using this research content: {content}"
+        )
+
+    @step()
+    async def write(
+        self, ctx: Context, ev: WriteEvent, writer: FunctionCallingAgent
+    ) -> ReviewEvent | StopEvent:
+        MAX_ATTEMPTS = 2
+        ctx.data["attempts"] = ctx.data.get("attempts", 0) + 1
+        too_many_attempts = ctx.data["attempts"] > MAX_ATTEMPTS
+        if too_many_attempts:
+            ctx.write_event_to_stream(
+                AgentRunEvent(
+                    name=writer.name,
+                    msg=f"Too many attempts ({MAX_ATTEMPTS}) to write the blog post. Proceeding with the current version.",
+                )
+            )
+        if ev.is_good or too_many_attempts:
+            # too many attempts or the blog post is good - stream final response if requested
+            result = await self.run_agent(
+                ctx, writer, ev.input, streaming=ctx.data["streaming"]
+            )
+            return StopEvent(result=result)
+        result: AgentRunResult = await self.run_agent(ctx, writer, ev.input)
+        ctx.data["result"] = result
+        return ReviewEvent(input=result.response.message.content)
+
+    @step()
+    async def review(
+        self, ctx: Context, ev: ReviewEvent, reviewer: FunctionCallingAgent
+    ) -> WriteEvent:
+        result: AgentRunResult = await self.run_agent(ctx, reviewer, ev.input)
+        review = result.response.message.content
+        old_content = ctx.data["result"].response.message.content
+        post_is_good = "post is good" in review.lower()
+        ctx.write_event_to_stream(
+            AgentRunEvent(
+                name=reviewer.name,
+                msg=f"The post is {'not ' if not post_is_good else ''}good enough for publishing. Sending back to the writer{' for publication.' if post_is_good else '.'}",
+            )
+        )
+        if post_is_good:
+            return WriteEvent(
+                input=f"You're blog post is ready for publication. Please respond with just the blog post. Blog post: ```{old_content}```",
+                is_good=True,
+            )
+        else:
+            return WriteEvent(
+                input=dedent(
+                    f"""
+                    Improve the writing of a given blog post by using a given review.
+                    Blog post:
+                    ```
+                    {old_content}
+                    ``` 
+
+                    Review: 
+                    ```
+                    {review}
+                    ```
+                    """
+                ),
+            )
+
+    @step()
+    async def publish(
+        self,
+        ctx: Context,
+        ev: PublishEvent,
+        publisher: FunctionCallingAgent,
+    ) -> StopEvent:
+        try:
+            result: AgentRunResult = await self.run_agent(ctx, publisher, ev.input)
+            return StopEvent(result=result)
+        except Exception as e:
+            ctx.write_event_to_stream(
+                AgentRunEvent(
+                    name=publisher.name,
+                    msg=f"Error publishing: {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

templates/components/multiagent/typescript/express/chat.controller.ts

@@ -0,0 +1,41 @@
+import { StopEvent } from "@llamaindex/core/workflow";
+import { Message, streamToResponse } from "ai";
+import { Request, Response } from "express";
+import { ChatMessage, ChatResponseChunk } from "llamaindex";
+import { createWorkflow } from "./workflow/factory";
+import { toDataStream, workflowEventsToStreamData } from "./workflow/stream";
+
+export const chat = async (req: Request, res: Response) => {
+  try {
+    const { messages }: { messages: Message[] } = req.body;
+    const userMessage = messages.pop();
+    if (!messages || !userMessage || userMessage.role !== "user") {
+      return res.status(400).json({
+        error:
+          "messages are required in the request body and the last message must be from the user",
+      });
+    }
+
+    const chatHistory = messages as ChatMessage[];
+    const agent = createWorkflow(chatHistory);
+    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(),
+    });
+
+    return streamToResponse(stream, res, {}, agentStreamData);
+  } catch (error) {
+    console.error("[LlamaIndex]", error);
+    return res.status(500).json({
+      detail: (error as Error).message,
+    });
+  }
+};

templates/components/multiagent/typescript/nextjs/route.ts

@@ -0,0 +1,57 @@
+import { initObservability } from "@/app/observability";
+import { StopEvent } from "@llamaindex/core/workflow";
+import { Message, StreamingTextResponse } from "ai";
+import { ChatMessage, ChatResponseChunk } from "llamaindex";
+import { NextRequest, NextResponse } from "next/server";
+import { initSettings } from "./engine/settings";
+import { createWorkflow } from "./workflow/factory";
+import { toDataStream, workflowEventsToStreamData } from "./workflow/stream";
+
+initObservability();
+initSettings();
+
+export const runtime = "nodejs";
+export const dynamic = "force-dynamic";
+
+export async function POST(request: NextRequest) {
+  try {
+    const body = await request.json();
+    const { messages }: { messages: Message[] } = body;
+    const userMessage = messages.pop();
+    if (!messages || !userMessage || userMessage.role !== "user") {
+      return NextResponse.json(
+        {
+          error:
+            "messages are required in the request body and the last message must be from the user",
+        },
+        { status: 400 },
+      );
+    }
+
+    const chatHistory = messages as ChatMessage[];
+    const agent = createWorkflow(chatHistory);
+    // 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(),
+    });
+    return new StreamingTextResponse(stream, {}, agentStreamData);
+  } catch (error) {
+    console.error("[LlamaIndex]", error);
+    return NextResponse.json(
+      {
+        detail: (error as Error).message,
+      },
+      {
+        status: 500,
+      },
+    );
+  }
+}

templates/components/multiagent/typescript/workflow/agents.ts

@@ -0,0 +1,94 @@
+import { ChatMessage } from "llamaindex";
+import { FunctionCallingAgent } from "./single-agent";
+import { lookupTools } from "./tools";
+
+export const createResearcher = async (chatHistory: ChatMessage[]) => {
+  const tools = await lookupTools([
+    "query_index",
+    "wikipedia_tool",
+    "duckduckgo_search",
+    "image_generator",
+  ]);
+
+  return new FunctionCallingAgent({
+    name: "researcher",
+    tools: tools,
+    systemPrompt: `You are a researcher agent. You are given a research task.
+            
+If the conversation already includes the information and there is no new request for additional information from the user, you should return the appropriate content to the writer.
+Otherwise, you must use tools to retrieve information or images needed for the task.
+
+It's normal for the task to include some ambiguity. You must always think carefully about the context of the user's request to understand what are the main content needs to be retrieved.
+Example:
+    Request: "Create a blog post about the history of the internet, write in English and publish in PDF format."
+    ->Though: The main content is "history of the internet", while "write in English and publish in PDF format" is a requirement for other agents.
+    Your task: Look for information in English about the history of the Internet.
+    This is not your task: Create a blog post or look for how to create a PDF.
+
+    Next request: "Publish the blog post in HTML format."
+    ->Though: User just asking for a format change, the previous content is still valid.
+    Your task: Return the previous content of the post to the writer. No need to do any research.
+    This is not your task: Look for how to create an HTML file.
+
+If you use the tools but don't find any related information, please return "I didn't find any new information for {the topic}." along with the content you found. Don't try to make up information yourself.
+If the request doesn't need any new information because it was in the conversation history, please return "The task doesn't need any new information. Please reuse the existing content in the conversation history.
+`,
+    chatHistory,
+  });
+};
+
+export const createWriter = (chatHistory: ChatMessage[]) => {
+  return new FunctionCallingAgent({
+    name: "writer",
+    systemPrompt: `You are an expert in writing blog posts.
+You are given the task of writing a blog post based on research content provided by the researcher agent. Do not invent any information yourself. 
+It's important to read the entire conversation history to write the blog post accurately.
+If you receive a review from the reviewer, update the post according to the feedback and return the new post content.
+If the user requests an update with new information but no research content is provided, you must respond with: "I don't have any research content to write about."
+If the content is not valid (e.g., broken link, broken image, etc.), do not use it.
+It's normal for the task to include some ambiguity, so you must define the user's initial request to write the post correctly.
+If you update the post based on the reviewer's feedback, first explain what changes you made to the post, then provide the new post content. Do not include the reviewer's comments.
+Example:
+    Task: "Here is the information I found about the history of the internet: 
+    Create a blog post about the history of the internet, write in English, and publish in PDF format."
+    -> Your task: Use the research content {...} to write a blog post in English.
+    -> This is not your task: Create a PDF
+    Please note that a localhost link is acceptable, but dummy links like "example.com" or "your-website.com" are not valid.`,
+    chatHistory,
+  });
+};
+
+export const createReviewer = (chatHistory: ChatMessage[]) => {
+  return new FunctionCallingAgent({
+    name: "reviewer",
+    systemPrompt: `You are an expert in reviewing blog posts.
+You are given a task to review a blog post. As a reviewer, it's important that your review aligns with the user's request. Please focus on the user's request when reviewing the post.
+Review the post for logical inconsistencies, ask critical questions, and provide suggestions for improvement.
+Furthermore, proofread the post for grammar and spelling errors.
+Only if the post is good enough for publishing should you return 'The post is good.' In all other cases, return your review.
+It's normal for the task to include some ambiguity, so you must define the user's initial request to review the post correctly.
+Please note that a localhost link is acceptable, but dummy links like "example.com" or "your-website.com" are not valid.
+Example:
+    Task: "Create a blog post about the history of the internet, write in English and publish in PDF format."
+    -> Your task: Review whether the main content of the post is about the history of the internet and if it is written in English.
+    -> This is not your task: Create blog post, create PDF, write in English.`,
+    chatHistory,
+  });
+};
+
+export const createPublisher = async (chatHistory: ChatMessage[]) => {
+  const tools = await lookupTools(["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) {
+    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,
+    systemPrompt: systemPrompt,
+    chatHistory,
+  });
+};

templates/components/multiagent/typescript/workflow/factory.ts

@@ -0,0 +1,187 @@
+import {
+  Context,
+  StartEvent,
+  StopEvent,
+  Workflow,
+  WorkflowEvent,
+} from "@llamaindex/core/workflow";
+import { ChatMessage, ChatResponseChunk } from "llamaindex";
+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: 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;
+
+  // Construct a new agent message from agent messages
+  // Get annotations from assistant messages
+  const agentAnnotations = chatHistory
+    .filter((msg) => msg.role === "assistant")
+    .flatMap((msg) => msg.annotations || [])
+    .filter((annotation) => annotation.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],
+    ];
+  }
+  return chatHistory;
+};
+
+export const createWorkflow = (chatHistory: ChatMessage[]) => {
+  const chatHistoryWithAgentMessages = prepareChatHistory(chatHistory);
+  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);
+    return new ResearchEvent({
+      input: `Research for this task: ${ev.data.input}`,
+    });
+  };
+
+  const research = async (context: Context, ev: ResearchEvent) => {
+    const researcher = await createResearcher(chatHistoryWithAgentMessages);
+    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: ${researchResult}`,
+      isGood: false,
+    });
+  };
+
+  const write = async (context: Context, ev: WriteEvent) => {
+    context.set("attempts", context.get("attempts", 0) + 1);
+    const tooManyAttempts = context.get("attempts") > MAX_ATTEMPTS;
+    if (tooManyAttempts) {
+      context.writeEventToStream(
+        new AgentRunEvent({
+          name: "writer",
+          msg: `Too many attempts (${MAX_ATTEMPTS}) to write the blog post. Proceeding with the current version.`,
+        }),
+      );
+    }
+
+    if (ev.data.isGood || tooManyAttempts) {
+      return new PublishEvent({
+        input: "Please help me to publish the blog post.",
+      });
+    }
+
+    const writer = createWriter(chatHistoryWithAgentMessages);
+    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 });
+  workflow.addStep(ResearchEvent, research, { outputs: WriteEvent });
+  workflow.addStep(WriteEvent, write, { outputs: [ReviewEvent, PublishEvent] });
+  workflow.addStep(ReviewEvent, review, { outputs: WriteEvent });
+  workflow.addStep(PublishEvent, publish, { outputs: StopEvent });
+
+  return workflow;
+};

templates/components/multiagent/typescript/workflow/single-agent.ts

@@ -0,0 +1,236 @@
+import {
+  Context,
+  StartEvent,
+  StopEvent,
+  Workflow,
+  WorkflowEvent,
+} from "@llamaindex/core/workflow";
+import {
+  BaseToolWithCall,
+  ChatMemoryBuffer,
+  ChatMessage,
+  ChatResponse,
+  ChatResponseChunk,
+  Settings,
+  ToolCall,
+  ToolCallLLM,
+  ToolCallLLMMessageOptions,
+  callTool,
+} from "llamaindex";
+import { AgentInput, AgentRunEvent } from "./type";
+
+class InputEvent extends WorkflowEvent<{
+  input: ChatMessage[];
+}> {}
+
+class ToolCallEvent extends WorkflowEvent<{
+  toolCalls: ToolCall[];
+}> {}
+
+export class FunctionCallingAgent extends Workflow {
+  name: string;
+  llm: ToolCallLLM;
+  memory: ChatMemoryBuffer;
+  tools: BaseToolWithCall[];
+  systemPrompt?: string;
+  writeEvents: boolean;
+  role?: string;
+
+  constructor(options: {
+    name: string;
+    llm?: ToolCallLLM;
+    chatHistory?: ChatMessage[];
+    tools?: BaseToolWithCall[];
+    systemPrompt?: string;
+    writeEvents?: boolean;
+    role?: string;
+    verbose?: boolean;
+    timeout?: number;
+  }) {
+    super({
+      verbose: options?.verbose ?? false,
+      timeout: options?.timeout ?? 360,
+    });
+    this.name = options?.name;
+    this.llm = options.llm ?? (Settings.llm as ToolCallLLM);
+    this.checkToolCallSupport();
+    this.memory = new ChatMemoryBuffer({
+      llm: this.llm,
+      chatHistory: options.chatHistory,
+    });
+    this.tools = options?.tools ?? [];
+    this.systemPrompt = options.systemPrompt;
+    this.writeEvents = options?.writeEvents ?? true;
+    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,
+    });
+  }
+
+  private get chatHistory() {
+    return this.memory.getMessages();
+  }
+
+  private async prepareChatHistory(
+    ctx: Context,
+    ev: StartEvent<AgentInput>,
+  ): Promise<InputEvent> {
+    const { message, streaming } = ev.data.input;
+    ctx.set("streaming", streaming);
+    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,
+    ev: InputEvent,
+  ): Promise<StopEvent<string | AsyncGenerator> | ToolCallEvent> {
+    if (ctx.get("streaming")) {
+      return await this.handleLLMInputStream(ctx, ev);
+    }
+
+    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 });
+    }
+    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 (fullResponse) {
+        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 });
+    }
+
+    this.writeEvent("Finished task", context);
+    return new StopEvent({ result: generator });
+  }
+
+  private async handleToolCalls(
+    ctx: Context,
+    ev: ToolCallEvent,
+  ): 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: console.error.bind(console),
+        warn: () => {},
+      });
+      toolMsgs.push({
+        content: JSON.stringify(toolOutput.output),
+        role: "user",
+        options: {
+          toolResult: {
+            result: toolOutput.output,
+            isError: toolOutput.isError,
+            id: call.id,
+          },
+        },
+      });
+    }
+
+    for (const msg of toolMsgs) {
+      this.memory.put(msg);
+    }
+
+    return new InputEvent({ input: this.memory.getMessages() });
+  }
+
+  private writeEvent(msg: string, context: Context) {
+    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 [];
+  }
+}

templates/components/multiagent/typescript/workflow/stream.ts

@@ -0,0 +1,65 @@
+import { StopEvent } from "@llamaindex/core/workflow";
+import {
+  createCallbacksTransformer,
+  createStreamDataTransformer,
+  StreamData,
+  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>>>,
+) {
+  const trimStartOfStream = trimStartOfStreamHelper();
+  return new ReadableStream<string>({
+    start(controller) {
+      controller.enqueue(""); // Kickstart the stream
+    },
+    async pull(controller): Promise<void> {
+      const stopEvent = await result;
+      const generator = stopEvent.data.result;
+      const { value, done } = await generator.next();
+      if (done) {
+        controller.close();
+        return;
+      }
+
+      const text = trimStartOfStream(value.delta ?? "");
+      if (text) controller.enqueue(text);
+    },
+  });
+}
+
+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;
+}

templates/components/multiagent/typescript/workflow/tools.ts

@@ -0,0 +1,52 @@
+import fs from "fs/promises";
+import { BaseToolWithCall, QueryEngineTool } from "llamaindex";
+import path from "path";
+import { getDataSource } from "../engine";
+import { createTools } from "../engine/tools/index";
+
+const getQueryEngineTool = async (): Promise<QueryEngineTool | null> => {
+  const index = await getDataSource();
+  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),
+  );
+};

templates/components/multiagent/typescript/workflow/type.ts

@@ -0,0 +1,11 @@
+import { WorkflowEvent } from "@llamaindex/core/workflow";
+
+export type AgentInput = {
+  message: string;
+  streaming?: boolean;
+};
+
+export class AgentRunEvent extends WorkflowEvent<{
+  name: string;
+  msg: string;
+}> {}

templates/components/settings/python/settings.py

@@ -126,13 +126,7 @@ def init_fastembed():
 def init_groq():
     from llama_index.llms.groq import Groq
 
-    model_map: Dict[str, str] = {
-        "llama3-8b": "llama3-8b-8192",
-        "llama3-70b": "llama3-70b-8192",
-        "mixtral-8x7b": "mixtral-8x7b-32768",
-    }
-
-    Settings.llm = Groq(model=model_map[os.getenv("MODEL")])
+    Settings.llm = Groq(model=os.getenv("MODEL"))
     # Groq does not provide embeddings, so we use FastEmbed instead
     init_fastembed()
 

templates/components/settings/typescript/settings.ts

@@ -138,14 +138,8 @@ function initGroq() {
     "all-mpnet-base-v2": "Xenova/all-mpnet-base-v2",
   };
 
-  const modelMap: Record<string, string> = {
-    "llama3-8b": "llama3-8b-8192",
-    "llama3-70b": "llama3-70b-8192",
-    "mixtral-8x7b": "mixtral-8x7b-32768",
-  };
-
   Settings.llm = new Groq({
-    model: modelMap[process.env.MODEL!],
+    model: process.env.MODEL!,
   });
 
   Settings.embedModel = new HuggingFaceEmbedding({

templates/components/vectordbs/typescript/astra/generate.ts

@@ -1,7 +1,7 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import * as dotenv from "dotenv";
 import { VectorStoreIndex, storageContextFromDefaults } from "llamaindex";
-import { AstraDBVectorStore } from "llamaindex/storage/vectorStore/AstraDBVectorStore";
+import { AstraDBVectorStore } from "llamaindex/vector-store/AstraDBVectorStore";
 import { getDocuments } from "./loader";
 import { initSettings } from "./settings";
 import { checkRequiredEnvVars } from "./shared";
@@ -15,13 +15,12 @@ async function loadAndIndex() {
   // create vector store and a collection
   const collectionName = process.env.ASTRA_DB_COLLECTION!;
   const vectorStore = new AstraDBVectorStore();
-  await vectorStore.create(collectionName, {
+  await vectorStore.createAndConnect(collectionName, {
     vector: {
       dimension: parseInt(process.env.EMBEDDING_DIM!),
       metric: "cosine",
     },
   });
-  await vectorStore.connect(collectionName);
 
   // create index from documents and store them in Astra
   console.log("Start creating embeddings...");

templates/components/vectordbs/typescript/astra/index.ts

@@ -1,6 +1,6 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import { VectorStoreIndex } from "llamaindex";
-import { AstraDBVectorStore } from "llamaindex/storage/vectorStore/AstraDBVectorStore";
+import { AstraDBVectorStore } from "llamaindex/vector-store/AstraDBVectorStore";
 import { checkRequiredEnvVars } from "./shared";
 
 export async function getDataSource(params?: any) {

templates/components/vectordbs/typescript/chroma/generate.ts

@@ -1,7 +1,7 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import * as dotenv from "dotenv";
 import { VectorStoreIndex, storageContextFromDefaults } from "llamaindex";
-import { ChromaVectorStore } from "llamaindex/storage/vectorStore/ChromaVectorStore";
+import { ChromaVectorStore } from "llamaindex/vector-store/ChromaVectorStore";
 import { getDocuments } from "./loader";
 import { initSettings } from "./settings";
 import { checkRequiredEnvVars } from "./shared";
@@ -16,7 +16,7 @@ async function loadAndIndex() {
   const chromaUri = `http://${process.env.CHROMA_HOST}:${process.env.CHROMA_PORT}`;
 
   const vectorStore = new ChromaVectorStore({
-    collectionName: process.env.CHROMA_COLLECTION,
+    collectionName: process.env.CHROMA_COLLECTION!,
     chromaClientParams: { path: chromaUri },
   });
 

templates/components/vectordbs/typescript/chroma/index.ts

@@ -1,6 +1,6 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import { VectorStoreIndex } from "llamaindex";
-import { ChromaVectorStore } from "llamaindex/storage/vectorStore/ChromaVectorStore";
+import { ChromaVectorStore } from "llamaindex/vector-store/ChromaVectorStore";
 import { checkRequiredEnvVars } from "./shared";
 
 export async function getDataSource(params?: any) {
@@ -8,7 +8,7 @@ export async function getDataSource(params?: any) {
   const chromaUri = `http://${process.env.CHROMA_HOST}:${process.env.CHROMA_PORT}`;
 
   const store = new ChromaVectorStore({
-    collectionName: process.env.CHROMA_COLLECTION,
+    collectionName: process.env.CHROMA_COLLECTION!,
     chromaClientParams: { path: chromaUri },
   });
 

templates/components/vectordbs/typescript/llamacloud/queryFilter.ts

@@ -4,7 +4,6 @@ export function generateFilters(documentIds: string[]): MetadataFilters {
   // public documents don't have the "private" field or it's set to "false"
   const publicDocumentsFilter: MetadataFilter = {
     key: "private",
-    value: null,
     operator: "is_empty",
   };
 

templates/components/vectordbs/typescript/milvus/generate.ts

@@ -1,7 +1,7 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import * as dotenv from "dotenv";
 import { VectorStoreIndex, storageContextFromDefaults } from "llamaindex";
-import { MilvusVectorStore } from "llamaindex/storage/vectorStore/MilvusVectorStore";
+import { MilvusVectorStore } from "llamaindex/vector-store/MilvusVectorStore";
 import { getDocuments } from "./loader";
 import { initSettings } from "./settings";
 import { checkRequiredEnvVars, getMilvusClient } from "./shared";

templates/components/vectordbs/typescript/milvus/index.ts

@@ -1,5 +1,5 @@
 import { VectorStoreIndex } from "llamaindex";
-import { MilvusVectorStore } from "llamaindex/storage/vectorStore/MilvusVectorStore";
+import { MilvusVectorStore } from "llamaindex/vector-store/MilvusVectorStore";
 import { checkRequiredEnvVars, getMilvusClient } from "./shared";
 
 export async function getDataSource(params?: any) {

templates/components/vectordbs/typescript/mongo/generate.ts

@@ -1,7 +1,7 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import * as dotenv from "dotenv";
 import { storageContextFromDefaults, VectorStoreIndex } from "llamaindex";
-import { MongoDBAtlasVectorSearch } from "llamaindex/storage/vectorStore/MongoDBAtlasVectorStore";
+import { MongoDBAtlasVectorSearch } from "llamaindex/vector-store/MongoDBAtlasVectorStore";
 import { MongoClient } from "mongodb";
 import { getDocuments } from "./loader";
 import { initSettings } from "./settings";

templates/components/vectordbs/typescript/mongo/index.ts

@@ -1,6 +1,6 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import { VectorStoreIndex } from "llamaindex";
-import { MongoDBAtlasVectorSearch } from "llamaindex/storage/vectorStore/MongoDBAtlasVectorStore";
+import { MongoDBAtlasVectorSearch } from "llamaindex/vector-store/MongoDBAtlasVectorStore";
 import { MongoClient } from "mongodb";
 import { checkRequiredEnvVars, POPULATED_METADATA_FIELDS } from "./shared";
 

templates/components/vectordbs/typescript/pg/generate.ts

@@ -1,7 +1,6 @@
-/* eslint-disable turbo/no-undeclared-env-vars */
 import * as dotenv from "dotenv";
 import { VectorStoreIndex, storageContextFromDefaults } from "llamaindex";
-import { PGVectorStore } from "llamaindex/storage/vectorStore/PGVectorStore";
+import { PGVectorStore } from "llamaindex/vector-store/PGVectorStore";
 import { getDocuments } from "./loader";
 import { initSettings } from "./settings";
 import {

templates/components/vectordbs/typescript/pg/index.ts

@@ -1,6 +1,5 @@
-/* eslint-disable turbo/no-undeclared-env-vars */
 import { VectorStoreIndex } from "llamaindex";
-import { PGVectorStore } from "llamaindex/storage/vectorStore/PGVectorStore";
+import { PGVectorStore } from "llamaindex/vector-store/PGVectorStore";
 import {
   PGVECTOR_SCHEMA,
   PGVECTOR_TABLE,

templates/components/vectordbs/typescript/pinecone/generate.ts

@@ -1,7 +1,7 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import * as dotenv from "dotenv";
 import { VectorStoreIndex, storageContextFromDefaults } from "llamaindex";
-import { PineconeVectorStore } from "llamaindex/storage/vectorStore/PineconeVectorStore";
+import { PineconeVectorStore } from "llamaindex/vector-store/PineconeVectorStore";
 import { getDocuments } from "./loader";
 import { initSettings } from "./settings";
 import { checkRequiredEnvVars } from "./shared";

templates/components/vectordbs/typescript/pinecone/index.ts

@@ -1,6 +1,6 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import { VectorStoreIndex } from "llamaindex";
-import { PineconeVectorStore } from "llamaindex/storage/vectorStore/PineconeVectorStore";
+import { PineconeVectorStore } from "llamaindex/vector-store/PineconeVectorStore";
 import { checkRequiredEnvVars } from "./shared";
 
 export async function getDataSource(params?: any) {

templates/components/vectordbs/typescript/qdrant/generate.ts

@@ -1,7 +1,7 @@
 /* eslint-disable turbo/no-undeclared-env-vars */
 import * as dotenv from "dotenv";
 import { VectorStoreIndex, storageContextFromDefaults } from "llamaindex";
-import { QdrantVectorStore } from "llamaindex/storage/vectorStore/QdrantVectorStore";
+import { QdrantVectorStore } from "llamaindex/vector-store/QdrantVectorStore";
 import { getDocuments } from "./loader";
 import { initSettings } from "./settings";
 import { checkRequiredEnvVars, getQdrantClient } from "./shared";

templates/components/vectordbs/typescript/qdrant/index.ts

@@ -1,6 +1,6 @@
 import * as dotenv from "dotenv";
 import { VectorStoreIndex } from "llamaindex";
-import { QdrantVectorStore } from "llamaindex/storage/vectorStore/QdrantVectorStore";
+import { QdrantVectorStore } from "llamaindex/vector-store/QdrantVectorStore";
 import { checkRequiredEnvVars, getQdrantClient } from "./shared";
 
 dotenv.config();

templates/components/vectordbs/typescript/weaviate/index.ts

@@ -1,6 +1,6 @@
 import * as dotenv from "dotenv";
 import { VectorStoreIndex } from "llamaindex";
-import { WeaviateVectorStore } from "llamaindex/storage/vectorStore/WeaviateVectorStore";
+import { WeaviateVectorStore } from "llamaindex/vector-store/WeaviateVectorStore";
 import { checkRequiredEnvVars, DEFAULT_INDEX_NAME } from "./shared";
 
 dotenv.config();

templates/types/multiagent/fastapi/app/api/routers/chat.py

@@ -1,43 +0,0 @@
-import asyncio
-import logging
-
-from fastapi import APIRouter, HTTPException, Request, status
-from llama_index.core.workflow import Workflow
-
-from app.examples.factory import create_agent
-from app.api.routers.models import (
-    ChatData,
-)
-from app.api.routers.vercel_response import VercelStreamResponse
-
-chat_router = r = APIRouter()
-
-logger = logging.getLogger("uvicorn")
-
-
-@r.post("")
-async def chat(
-    request: Request,
-    data: ChatData,
-):
-    try:
-        last_message_content = data.get_last_message_content()
-        messages = data.get_history_messages()
-        # TODO: generate filters based on doc_ids
-        # for now just use all documents
-        # doc_ids = data.get_chat_document_ids()
-        # TODO: use params
-        # params = data.data or {}
-
-        agent: Workflow = create_agent(chat_history=messages)
-        task = asyncio.create_task(
-            agent.run(input=last_message_content, streaming=True)
-        )
-
-        return VercelStreamResponse(request, task, agent.stream_events, data)
-    except Exception as e:
-        logger.exception("Error in agent", exc_info=True)
-        raise HTTPException(
-            status_code=status.HTTP_500_INTERNAL_SERVER_ERROR,
-            detail=f"Error in agent: {e}",
-        ) from e

templates/types/multiagent/fastapi/app/api/routers/chat_config.py

@@ -1,48 +0,0 @@
-import logging
-import os
-
-from fastapi import APIRouter
-
-from app.api.routers.models import ChatConfig
-
-
-config_router = r = APIRouter()
-
-logger = logging.getLogger("uvicorn")
-
-
-@r.get("")
-async def chat_config() -> ChatConfig:
-    starter_questions = None
-    conversation_starters = os.getenv("CONVERSATION_STARTERS")
-    if conversation_starters and conversation_starters.strip():
-        starter_questions = conversation_starters.strip().split("\n")
-    return ChatConfig(starter_questions=starter_questions)
-
-
-try:
-    from app.engine.service import LLamaCloudFileService
-
-    logger.info("LlamaCloud is configured. Adding /config/llamacloud route.")
-
-    @r.get("/llamacloud")
-    async def chat_llama_cloud_config():
-        projects = LLamaCloudFileService.get_all_projects_with_pipelines()
-        pipeline = os.getenv("LLAMA_CLOUD_INDEX_NAME")
-        project = os.getenv("LLAMA_CLOUD_PROJECT_NAME")
-        pipeline_config = None
-        if pipeline and project:
-            pipeline_config = {
-                "pipeline": pipeline,
-                "project": project,
-            }
-        return {
-            "projects": projects,
-            "pipeline": pipeline_config,
-        }
-
-except ImportError:
-    logger.debug(
-        "LlamaCloud is not configured. Skipping adding /config/llamacloud route."
-    )
-    pass

templates/types/multiagent/fastapi/app/api/routers/models.py

@@ -1,227 +0,0 @@
-import logging
-import os
-from typing import Any, Dict, List, Literal, Optional
-
-from llama_index.core.llms import ChatMessage, MessageRole
-from llama_index.core.schema import NodeWithScore
-from pydantic import BaseModel, Field, validator
-from pydantic.alias_generators import to_camel
-
-from app.config import DATA_DIR
-
-logger = logging.getLogger("uvicorn")
-
-
-class FileContent(BaseModel):
-    type: Literal["text", "ref"]
-    # If the file is pure text then the value is be a string
-    # otherwise, it's a list of document IDs
-    value: str | List[str]
-
-
-class File(BaseModel):
-    id: str
-    content: FileContent
-    filename: str
-    filesize: int
-    filetype: str
-
-
-class AnnotationFileData(BaseModel):
-    files: List[File] = Field(
-        default=[],
-        description="List of files",
-    )
-
-    class Config:
-        json_schema_extra = {
-            "example": {
-                "csvFiles": [
-                    {
-                        "content": "Name, Age\nAlice, 25\nBob, 30",
-                        "filename": "example.csv",
-                        "filesize": 123,
-                        "id": "123",
-                        "type": "text/csv",
-                    }
-                ]
-            }
-        }
-        alias_generator = to_camel
-
-
-class Annotation(BaseModel):
-    type: str
-    data: AnnotationFileData | List[str]
-
-    def to_content(self) -> str | None:
-        if self.type == "document_file":
-            # We only support generating context content for CSV files for now
-            csv_files = [file for file in self.data.files if file.filetype == "csv"]
-            if len(csv_files) > 0:
-                return "Use data from following CSV raw content\n" + "\n".join(
-                    [f"```csv\n{csv_file.content.value}\n```" for csv_file in csv_files]
-                )
-        else:
-            logger.warning(
-                f"The annotation {self.type} is not supported for generating context content"
-            )
-        return None
-
-
-class Message(BaseModel):
-    role: MessageRole
-    content: str
-    annotations: List[Annotation] | None = None
-
-
-class ChatData(BaseModel):
-    messages: List[Message]
-    data: Any = None
-
-    class Config:
-        json_schema_extra = {
-            "example": {
-                "messages": [
-                    {
-                        "role": "user",
-                        "content": "What standards for letters exist?",
-                    }
-                ]
-            }
-        }
-
-    @validator("messages")
-    def messages_must_not_be_empty(cls, v):
-        if len(v) == 0:
-            raise ValueError("Messages must not be empty")
-        return v
-
-    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
-        """
-        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):
-            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
-        return message_content
-
-    def get_history_messages(self) -> List[ChatMessage]:
-        """
-        Get the history messages
-        """
-        return [
-            ChatMessage(role=message.role, content=message.content)
-            for message in self.messages[:-1]
-        ]
-
-    def is_last_message_from_user(self) -> bool:
-        return self.messages[-1].role == MessageRole.USER
-
-    def get_chat_document_ids(self) -> List[str]:
-        """
-        Get the document IDs from the chat messages
-        """
-        document_ids: List[str] = []
-        for message in self.messages:
-            if message.role == MessageRole.USER and message.annotations is not None:
-                for annotation in message.annotations:
-                    if (
-                        annotation.type == "document_file"
-                        and annotation.data.files is not None
-                    ):
-                        for fi in annotation.data.files:
-                            if fi.content.type == "ref":
-                                document_ids += fi.content.value
-        return list(set(document_ids))
-
-
-class SourceNodes(BaseModel):
-    id: str
-    metadata: Dict[str, Any]
-    score: Optional[float]
-    text: str
-    url: Optional[str]
-
-    @classmethod
-    def from_source_node(cls, source_node: NodeWithScore):
-        metadata = source_node.node.metadata
-        url = cls.get_url_from_metadata(metadata)
-
-        return cls(
-            id=source_node.node.node_id,
-            metadata=metadata,
-            score=source_node.score,
-            text=source_node.node.text,  # type: ignore
-            url=url,
-        )
-
-    @classmethod
-    def get_url_from_metadata(cls, metadata: Dict[str, Any]) -> str:
-        url_prefix = os.getenv("FILESERVER_URL_PREFIX")
-        if not url_prefix:
-            logger.warning(
-                "Warning: FILESERVER_URL_PREFIX not set in environment variables. Can't use file server"
-            )
-        file_name = metadata.get("file_name")
-
-        if file_name and url_prefix:
-            # file_name exists and file server is configured
-            pipeline_id = metadata.get("pipeline_id")
-            if pipeline_id:
-                # file is from LlamaCloud
-                file_name = f"{pipeline_id}${file_name}"
-                return f"{url_prefix}/output/llamacloud/{file_name}"
-            is_private = metadata.get("private", "false") == "true"
-            if is_private:
-                # file is a private upload
-                return f"{url_prefix}/output/uploaded/{file_name}"
-            # file is from calling the 'generate' script
-            # Get the relative path of file_path to data_dir
-            file_path = metadata.get("file_path")
-            data_dir = os.path.abspath(DATA_DIR)
-            if file_path and data_dir:
-                relative_path = os.path.relpath(file_path, data_dir)
-                return f"{url_prefix}/data/{relative_path}"
-        # fallback to URL in metadata (e.g. for websites)
-        return metadata.get("URL")
-
-    @classmethod
-    def from_source_nodes(cls, source_nodes: List[NodeWithScore]):
-        return [cls.from_source_node(node) for node in source_nodes]
-
-
-class Result(BaseModel):
-    result: Message
-    nodes: List[SourceNodes]
-
-
-class ChatConfig(BaseModel):
-    starter_questions: Optional[List[str]] = Field(
-        default=None,
-        description="List of starter questions",
-        serialization_alias="starterQuestions",
-    )
-
-    class Config:
-        json_schema_extra = {
-            "example": {
-                "starterQuestions": [
-                    "What standards for letters exist?",
-                    "What are the requirements for a letter to be considered a letter?",
-                ]
-            }
-        }

templates/types/multiagent/fastapi/app/api/routers/upload.py

@@ -1,29 +0,0 @@
-import logging
-from typing import List, Any
-
-from fastapi import APIRouter, HTTPException
-from pydantic import BaseModel
-
-from app.api.services.file import PrivateFileService
-
-file_upload_router = r = APIRouter()
-
-logger = logging.getLogger("uvicorn")
-
-
-class FileUploadRequest(BaseModel):
-    base64: str
-    filename: str
-    params: Any = None
-
-
-@r.post("")
-def upload_file(request: FileUploadRequest) -> List[str]:
-    try:
-        logger.info("Processing file")
-        return PrivateFileService.process_file(
-            request.filename, request.base64, request.params
-        )
-    except Exception as e:
-        logger.error(f"Error processing file: {e}", exc_info=True)
-        raise HTTPException(status_code=500, detail="Error processing file")

templates/types/multiagent/fastapi/app/config.py

@@ -1 +0,0 @@
-DATA_DIR = "data"

templates/types/multiagent/fastapi/app/examples/choreography.py

@@ -1,25 +0,0 @@
-from typing import List, Optional
-from app.agents.single import FunctionCallingAgent
-from app.agents.multi import AgentCallingAgent
-from app.examples.researcher import create_researcher
-from llama_index.core.chat_engine.types import ChatMessage
-
-
-def create_choreography(chat_history: Optional[List[ChatMessage]] = None):
-    researcher = create_researcher(chat_history)
-    reviewer = FunctionCallingAgent(
-        name="reviewer",
-        role="expert in reviewing blog posts",
-        system_prompt="You are an expert in reviewing blog posts. You are given a task to review a blog post. Review the post for logical inconsistencies, ask critical questions, and provide suggestions for improvement. Furthermore, proofread the post for grammar and spelling errors. If the post is good, you can say 'The post is good.'",
-        chat_history=chat_history,
-    )
-    return AgentCallingAgent(
-        name="writer",
-        agents=[researcher, reviewer],
-        role="expert in writing blog posts",
-        system_prompt="""You are an expert in writing blog posts. You are given a task to write a blog post. Before starting to write the post, consult the researcher agent to get the information you need. Don't make up any information yourself.
-        After creating a draft for the post, send it to the reviewer agent to receive some feedback and make sure to incorporate the feedback from the reviewer.
-        You can consult the reviewer and researcher maximal two times. Your output should just contain the blog post.""",
-        # TODO: add chat_history support to AgentCallingAgent
-        # chat_history=chat_history,
-    )

templates/types/multiagent/fastapi/app/examples/orchestrator.py

@@ -1,27 +0,0 @@
-from typing import List, Optional
-from app.agents.single import FunctionCallingAgent
-from app.agents.multi import AgentOrchestrator
-from app.examples.researcher import create_researcher
-
-from llama_index.core.chat_engine.types import ChatMessage
-
-
-def create_orchestrator(chat_history: Optional[List[ChatMessage]] = None):
-    researcher = create_researcher(chat_history)
-    writer = FunctionCallingAgent(
-        name="writer",
-        role="expert in writing blog posts",
-        system_prompt="""You are an expert in writing blog posts. You are given a task to write a blog post. Don't make up any information yourself. If you don't have the necessary information to write a blog post, reply "I need information about the topic to write the blog post". If you have all the information needed, write the blog post.""",
-        chat_history=chat_history,
-    )
-    reviewer = FunctionCallingAgent(
-        name="reviewer",
-        role="expert in reviewing blog posts",
-        system_prompt="""You are an expert in reviewing blog posts. You are given a task to review a blog post. Review the post and fix the issues found yourself. You must output a final blog post.
-        Especially check for logical inconsistencies and proofread the post for grammar and spelling errors.""",
-        chat_history=chat_history,
-    )
-    return AgentOrchestrator(
-        agents=[writer, reviewer, researcher],
-        refine_plan=False,
-    )

templates/types/multiagent/fastapi/app/examples/researcher.py

@@ -1,39 +0,0 @@
-import os
-from typing import List
-from llama_index.core.tools import QueryEngineTool, ToolMetadata
-from app.agents.single import FunctionCallingAgent
-from app.engine.index import get_index
-
-from llama_index.core.chat_engine.types import ChatMessage
-
-
-def get_query_engine_tool() -> QueryEngineTool:
-    """
-    Provide an agent worker that can be used to query the index.
-    """
-    index = get_index()
-    if index is None:
-        raise ValueError("Index not found. Please create an index first.")
-    top_k = int(os.getenv("TOP_K", 0))
-    query_engine = index.as_query_engine(
-        **({"similarity_top_k": top_k} if top_k != 0 else {})
-    )
-    return QueryEngineTool(
-        query_engine=query_engine,
-        metadata=ToolMetadata(
-            name="query_index",
-            description="""
-                Use this tool to retrieve information about the text corpus from the index.
-            """,
-        ),
-    )
-
-
-def create_researcher(chat_history: List[ChatMessage]):
-    return FunctionCallingAgent(
-        name="researcher",
-        tools=[get_query_engine_tool()],
-        role="expert in retrieving any unknown content",
-        system_prompt="You are a researcher agent. You are given a researching task. You must use your tools to complete the research.",
-        chat_history=chat_history,
-    )

templates/types/multiagent/fastapi/app/examples/workflow.py

@@ -1,139 +0,0 @@
-import asyncio
-from typing import AsyncGenerator, List, Optional
-
-
-from llama_index.core.workflow import (
-    Context,
-    Event,
-    StartEvent,
-    StopEvent,
-    Workflow,
-    step,
-)
-from llama_index.core.chat_engine.types import ChatMessage
-from app.agents.single import AgentRunEvent, AgentRunResult, FunctionCallingAgent
-from app.examples.researcher import create_researcher
-
-
-def create_workflow(chat_history: Optional[List[ChatMessage]] = None):
-    researcher = create_researcher(
-        chat_history=chat_history,
-    )
-    writer = FunctionCallingAgent(
-        name="writer",
-        role="expert in writing blog posts",
-        system_prompt="""You are an expert in writing blog posts. You are given a task to write a blog post. Don't make up any information yourself.""",
-        chat_history=chat_history,
-    )
-    reviewer = FunctionCallingAgent(
-        name="reviewer",
-        role="expert in reviewing blog posts",
-        system_prompt="You are an expert in reviewing blog posts. You are given a task to review a blog post. Review the post for logical inconsistencies, ask critical questions, and provide suggestions for improvement. Furthermore, proofread the post for grammar and spelling errors. Only if the post is good enough for publishing, then you MUST return 'The post is good.'. In all other cases return your review.",
-        chat_history=chat_history,
-    )
-    workflow = BlogPostWorkflow(timeout=360)
-    workflow.add_workflows(researcher=researcher, writer=writer, reviewer=reviewer)
-    return workflow
-
-
-class ResearchEvent(Event):
-    input: str
-
-
-class WriteEvent(Event):
-    input: str
-    is_good: bool = False
-
-
-class ReviewEvent(Event):
-    input: str
-
-
-class BlogPostWorkflow(Workflow):
-    @step()
-    async def start(self, ctx: Context, ev: StartEvent) -> ResearchEvent:
-        # set streaming
-        ctx.data["streaming"] = getattr(ev, "streaming", False)
-        # start the workflow with researching about a topic
-        ctx.data["task"] = ev.input
-        return ResearchEvent(input=f"Research for this task: {ev.input}")
-
-    @step()
-    async def research(
-        self, ctx: Context, ev: ResearchEvent, researcher: FunctionCallingAgent
-    ) -> WriteEvent:
-        result: AgentRunResult = await self.run_agent(ctx, researcher, ev.input)
-        content = result.response.message.content
-        return WriteEvent(
-            input=f"Write a blog post given this task: {ctx.data['task']} using this research content: {content}"
-        )
-
-    @step()
-    async def write(
-        self, ctx: Context, ev: WriteEvent, writer: FunctionCallingAgent
-    ) -> ReviewEvent | StopEvent:
-        MAX_ATTEMPTS = 2
-        ctx.data["attempts"] = ctx.data.get("attempts", 0) + 1
-        too_many_attempts = ctx.data["attempts"] > MAX_ATTEMPTS
-        if too_many_attempts:
-            ctx.write_event_to_stream(
-                AgentRunEvent(
-                    name=writer.name,
-                    msg=f"Too many attempts ({MAX_ATTEMPTS}) to write the blog post. Proceeding with the current version.",
-                )
-            )
-        if ev.is_good or too_many_attempts:
-            # too many attempts or the blog post is good - stream final response if requested
-            result = await self.run_agent(
-                ctx, writer, ev.input, streaming=ctx.data["streaming"]
-            )
-            return StopEvent(result=result)
-        result: AgentRunResult = await self.run_agent(ctx, writer, ev.input)
-        ctx.data["result"] = result
-        return ReviewEvent(input=result.response.message.content)
-
-    @step()
-    async def review(
-        self, ctx: Context, ev: ReviewEvent, reviewer: FunctionCallingAgent
-    ) -> WriteEvent:
-        result: AgentRunResult = await self.run_agent(ctx, reviewer, ev.input)
-        review = result.response.message.content
-        old_content = ctx.data["result"].response.message.content
-        post_is_good = "post is good" in review.lower()
-        ctx.write_event_to_stream(
-            AgentRunEvent(
-                name=reviewer.name,
-                msg=f"The post is {'not ' if not post_is_good else ''}good enough for publishing. Sending back to the writer{' for publication.' if post_is_good else '.'}",
-            )
-        )
-        if post_is_good:
-            return WriteEvent(
-                input=f"You're blog post is ready for publication. Please respond with just the blog post. Blog post: ```{old_content}```",
-                is_good=True,
-            )
-        else:
-            return WriteEvent(
-                input=f"""Improve the writing of a given blog post by using a given review.
-Blog post:
-```
-{old_content}
-``` 
-
-Review: 
-```
-{review}
-```"""
-            )
-
-    async def run_agent(
-        self,
-        ctx: Context,
-        agent: FunctionCallingAgent,
-        input: str,
-        streaming: bool = False,
-    ) -> AgentRunResult | AsyncGenerator:
-        task = asyncio.create_task(agent.run(input=input, streaming=streaming))
-        # bubble all events while running the executor to the planner
-        async for event in agent.stream_events():
-            ctx.write_event_to_stream(event)
-        return await task

templates/types/multiagent/fastapi/app/observability.py

@@ -1,2 +0,0 @@
-def init_observability():
-    pass

templates/types/multiagent/fastapi/app/utils.py

@@ -1,8 +0,0 @@
-import os
-
-
-def load_from_env(var: str, throw_error: bool = True) -> str:
-    res = os.getenv(var)
-    if res is None and throw_error:
-        raise ValueError(f"Missing environment variable: {var}")
-    return res

templates/types/multiagent/fastapi/gitignore

@@ -1,4 +0,0 @@
-__pycache__
-storage
-.env
-output

templates/types/multiagent/fastapi/main.py

@@ -1,72 +0,0 @@
-# flake8: noqa: E402
-import os
-from dotenv import load_dotenv
-
-from app.config import DATA_DIR
-
-load_dotenv()
-
-import logging
-
-import uvicorn
-from app.api.routers.chat import chat_router
-from app.api.routers.chat_config import config_router
-from app.api.routers.upload import file_upload_router
-from app.observability import init_observability
-from app.settings import init_settings
-from fastapi import FastAPI
-from fastapi.middleware.cors import CORSMiddleware
-from fastapi.responses import RedirectResponse
-from fastapi.staticfiles import StaticFiles
-
-app = FastAPI()
-
-init_settings()
-init_observability()
-
-
-environment = os.getenv("ENVIRONMENT", "dev")  # Default to 'development' if not set
-logger = logging.getLogger("uvicorn")
-
-if environment == "dev":
-    logger.warning("Running in development mode - allowing CORS for all origins")
-    app.add_middleware(
-        CORSMiddleware,
-        allow_origins=["*"],
-        allow_credentials=True,
-        allow_methods=["*"],
-        allow_headers=["*"],
-    )
-
-    # Redirect to documentation page when accessing base URL
-    @app.get("/")
-    async def redirect_to_docs():
-        return RedirectResponse(url="/docs")
-
-
-def mount_static_files(directory, path):
-    if os.path.exists(directory):
-        logger.info(f"Mounting static files '{directory}' at '{path}'")
-        app.mount(
-            path,
-            StaticFiles(directory=directory, check_dir=False),
-            name=f"{directory}-static",
-        )
-
-
-# Mount the data files to serve the file viewer
-mount_static_files(DATA_DIR, "/api/files/data")
-# Mount the output files from tools
-mount_static_files("output", "/api/files/output")
-
-app.include_router(chat_router, prefix="/api/chat")
-app.include_router(config_router, prefix="/api/chat/config")
-app.include_router(file_upload_router, prefix="/api/chat/upload")
-
-
-if __name__ == "__main__":
-    app_host = os.getenv("APP_HOST", "0.0.0.0")
-    app_port = int(os.getenv("APP_PORT", "8000"))
-    reload = True if environment == "dev" else False
-
-    uvicorn.run(app="main:app", host=app_host, port=app_port, reload=reload)

templates/types/multiagent/fastapi/pyproject.toml

@@ -1,27 +0,0 @@
-[tool]
-[tool.poetry]
-name = "app"
-version = "0.1.0"
-description = ""
-authors = ["Marcus Schiesser <mail@marcusschiesser.de>"]
-readme = "README.md"
-
-[tool.poetry.scripts]
-generate = "app.engine.generate:generate_datasource"
-
-[tool.poetry.dependencies]
-python = "^3.11"
-llama-index-agent-openai = ">=0.3.0,<0.4.0"
-llama-index = "^0.11.4"
-fastapi = "^0.112.2"
-python-dotenv = "^1.0.0"
-uvicorn = { extras = ["standard"], version = "^0.23.2" }
-cachetools = "^5.3.3"
-aiostream = "^0.5.2"
-
-[tool.poetry.dependencies.docx2txt]
-version = "^0.8"
-
-[build-system]
-requires = ["poetry-core"]
-build-backend = "poetry.core.masonry.api"

templates/types/streaming/express/package.json

@@ -15,18 +15,20 @@
     "dev": "concurrently \"tsup index.ts --format esm --dts --watch\" \"nodemon --watch dist/index.js\""
   },
   "dependencies": {
-    "ai": "3.0.21",
+    "@llamaindex/core": "^0.2.6",
+    "ai": "3.3.42",
     "cors": "^2.8.5",
     "dotenv": "^16.3.1",
     "duck-duck-scrape": "^2.2.5",
     "express": "^4.18.2",
-    "llamaindex": "0.5.24",
+    "llamaindex": "0.6.2",
     "pdf2json": "3.0.5",
     "ajv": "^8.12.0",
     "@e2b/code-interpreter": "^0.0.5",
     "got": "^14.4.1",
     "@apidevtools/swagger-parser": "^10.1.0",
-    "formdata-node": "^6.0.3"
+    "formdata-node": "^6.0.3",
+    "marked": "^14.1.2"
   },
   "devDependencies": {
     "@types/cors": "^2.8.16",

templates/types/streaming/express/src/controllers/chat-request.controller.ts

@@ -35,7 +35,7 @@ export const chatRequest = async (req: Request, res: Response) => {
     // Convert message content from Vercel/AI format to LlamaIndex/OpenAI format
     // Note: The non-streaming template does not need the Vercel/AI format, we're still using it for consistency with the streaming template
     const userMessageContent = convertMessageContent(
-      userMessage.content,
+      userMessage.content as string,
       data?.imageUrl,
     );
 

templates/types/streaming/express/src/controllers/chat-upload.controller.ts

@@ -14,5 +14,11 @@ export const chatUpload = async (req: Request, res: Response) => {
     });
   }
   const index = await getDataSource(params);
+  if (!index) {
+    return res.status(500).json({
+      error:
+        "StorageContext is empty - call 'npm run generate' to generate the storage first",
+    });
+  }
   return res.status(200).json(await uploadDocument(index, filename, base64));
 };

templates/types/streaming/express/src/controllers/chat.controller.ts

@@ -1,4 +1,10 @@
-import { JSONValue, Message, StreamData, streamToResponse } from "ai";
+import {
+  JSONValue,
+  LlamaIndexAdapter,
+  Message,
+  StreamData,
+  streamToResponse,
+} from "ai";
 import { Request, Response } from "express";
 import { ChatMessage, Settings } from "llamaindex";
 import { createChatEngine } from "./engine/chat";
@@ -10,7 +16,7 @@ import {
   createCallbackManager,
   createStreamTimeout,
 } from "./llamaindex/streaming/events";
-import { LlamaIndexStream } from "./llamaindex/streaming/stream";
+import { generateNextQuestions } from "./llamaindex/streaming/suggestion";
 
 export const chat = async (req: Request, res: Response) => {
   // Init Vercel AI StreamData and timeout
@@ -56,23 +62,34 @@ export const chat = async (req: Request, res: Response) => {
 
     // Setup callbacks
     const callbackManager = createCallbackManager(vercelStreamData);
+    const chatHistory: ChatMessage[] = messages as ChatMessage[];
 
     // Calling LlamaIndex's ChatEngine to get a streamed response
     const response = await Settings.withCallbackManager(callbackManager, () => {
       return chatEngine.chat({
         message: userMessageContent,
-        chatHistory: messages as ChatMessage[],
+        chatHistory,
         stream: true,
       });
     });
 
-    // Return a stream, which can be consumed by the Vercel/AI client
-    const stream = LlamaIndexStream(
-      response,
-      vercelStreamData,
-      messages as ChatMessage[],
-    );
+    const onFinal = (content: string) => {
+      chatHistory.push({ role: "assistant", content: content });
+      generateNextQuestions(chatHistory)
+        .then((questions: string[]) => {
+          if (questions.length > 0) {
+            vercelStreamData.appendMessageAnnotation({
+              type: "suggested_questions",
+              data: questions,
+            });
+          }
+        })
+        .finally(() => {
+          vercelStreamData.close();
+        });
+    };
 
+    const stream = LlamaIndexAdapter.toDataStream(response, { onFinal });
     return streamToResponse(stream, res, {}, vercelStreamData);
   } catch (error) {
     console.error("[LlamaIndex]", error);

templates/types/streaming/express/src/controllers/engine/settings.ts

@@ -1,185 +0,0 @@
-import {
-  ALL_AVAILABLE_MISTRAL_MODELS,
-  Anthropic,
-  GEMINI_EMBEDDING_MODEL,
-  GEMINI_MODEL,
-  Gemini,
-  GeminiEmbedding,
-  Groq,
-  MistralAI,
-  MistralAIEmbedding,
-  MistralAIEmbeddingModelType,
-  OpenAI,
-  OpenAIEmbedding,
-  Settings,
-} from "llamaindex";
-import { HuggingFaceEmbedding } from "llamaindex/embeddings/HuggingFaceEmbedding";
-import { OllamaEmbedding } from "llamaindex/embeddings/OllamaEmbedding";
-import { ALL_AVAILABLE_ANTHROPIC_MODELS } from "llamaindex/llm/anthropic";
-import { Ollama } from "llamaindex/llm/ollama";
-
-const CHUNK_SIZE = 512;
-const CHUNK_OVERLAP = 20;
-
-export const initSettings = async () => {
-  // HINT: you can delete the initialization code for unused model providers
-  console.log(`Using '${process.env.MODEL_PROVIDER}' model provider`);
-
-  if (!process.env.MODEL || !process.env.EMBEDDING_MODEL) {
-    throw new Error("'MODEL' and 'EMBEDDING_MODEL' env variables must be set.");
-  }
-
-  switch (process.env.MODEL_PROVIDER) {
-    case "ollama":
-      initOllama();
-      break;
-    case "groq":
-      initGroq();
-      break;
-    case "anthropic":
-      initAnthropic();
-      break;
-    case "gemini":
-      initGemini();
-      break;
-    case "mistral":
-      initMistralAI();
-      break;
-    case "azure-openai":
-      initAzureOpenAI();
-      break;
-    default:
-      initOpenAI();
-      break;
-  }
-  Settings.chunkSize = CHUNK_SIZE;
-  Settings.chunkOverlap = CHUNK_OVERLAP;
-};
-
-function initOpenAI() {
-  Settings.llm = new OpenAI({
-    model: process.env.MODEL ?? "gpt-4o-mini",
-    maxTokens: process.env.LLM_MAX_TOKENS
-      ? Number(process.env.LLM_MAX_TOKENS)
-      : undefined,
-  });
-  Settings.embedModel = new OpenAIEmbedding({
-    model: process.env.EMBEDDING_MODEL,
-    dimensions: process.env.EMBEDDING_DIM
-      ? parseInt(process.env.EMBEDDING_DIM)
-      : undefined,
-  });
-}
-
-function initAzureOpenAI() {
-  // Map Azure OpenAI model names to OpenAI model names (only for TS)
-  const AZURE_OPENAI_MODEL_MAP: Record<string, string> = {
-    "gpt-35-turbo": "gpt-3.5-turbo",
-    "gpt-35-turbo-16k": "gpt-3.5-turbo-16k",
-    "gpt-4o": "gpt-4o",
-    "gpt-4": "gpt-4",
-    "gpt-4-32k": "gpt-4-32k",
-    "gpt-4-turbo": "gpt-4-turbo",
-    "gpt-4-turbo-2024-04-09": "gpt-4-turbo",
-    "gpt-4-vision-preview": "gpt-4-vision-preview",
-    "gpt-4-1106-preview": "gpt-4-1106-preview",
-    "gpt-4o-2024-05-13": "gpt-4o-2024-05-13",
-  };
-
-  const azureConfig = {
-    apiKey: process.env.AZURE_OPENAI_KEY,
-    endpoint: process.env.AZURE_OPENAI_ENDPOINT,
-    apiVersion:
-      process.env.AZURE_OPENAI_API_VERSION || process.env.OPENAI_API_VERSION,
-  };
-
-  Settings.llm = new OpenAI({
-    model:
-      AZURE_OPENAI_MODEL_MAP[process.env.MODEL ?? "gpt-35-turbo"] ??
-      "gpt-3.5-turbo",
-    maxTokens: process.env.LLM_MAX_TOKENS
-      ? Number(process.env.LLM_MAX_TOKENS)
-      : undefined,
-    azure: {
-      ...azureConfig,
-      deployment: process.env.AZURE_OPENAI_LLM_DEPLOYMENT,
-    },
-  });
-
-  Settings.embedModel = new OpenAIEmbedding({
-    model: process.env.EMBEDDING_MODEL,
-    dimensions: process.env.EMBEDDING_DIM
-      ? parseInt(process.env.EMBEDDING_DIM)
-      : undefined,
-    azure: {
-      ...azureConfig,
-      deployment: process.env.AZURE_OPENAI_EMBEDDING_DEPLOYMENT,
-    },
-  });
-}
-
-function initOllama() {
-  const config = {
-    host: process.env.OLLAMA_BASE_URL ?? "http://127.0.0.1:11434",
-  };
-  Settings.llm = new Ollama({
-    model: process.env.MODEL ?? "",
-    config,
-  });
-  Settings.embedModel = new OllamaEmbedding({
-    model: process.env.EMBEDDING_MODEL ?? "",
-    config,
-  });
-}
-
-function initGroq() {
-  const embedModelMap: Record<string, string> = {
-    "all-MiniLM-L6-v2": "Xenova/all-MiniLM-L6-v2",
-    "all-mpnet-base-v2": "Xenova/all-mpnet-base-v2",
-  };
-
-  const modelMap: Record<string, string> = {
-    "llama3-8b": "llama3-8b-8192",
-    "llama3-70b": "llama3-70b-8192",
-    "mixtral-8x7b": "mixtral-8x7b-32768",
-  };
-
-  Settings.llm = new Groq({
-    model: modelMap[process.env.MODEL!],
-  });
-
-  Settings.embedModel = new HuggingFaceEmbedding({
-    modelType: embedModelMap[process.env.EMBEDDING_MODEL!],
-  });
-}
-
-function initAnthropic() {
-  const embedModelMap: Record<string, string> = {
-    "all-MiniLM-L6-v2": "Xenova/all-MiniLM-L6-v2",
-    "all-mpnet-base-v2": "Xenova/all-mpnet-base-v2",
-  };
-  Settings.llm = new Anthropic({
-    model: process.env.MODEL as keyof typeof ALL_AVAILABLE_ANTHROPIC_MODELS,
-  });
-  Settings.embedModel = new HuggingFaceEmbedding({
-    modelType: embedModelMap[process.env.EMBEDDING_MODEL!],
-  });
-}
-
-function initGemini() {
-  Settings.llm = new Gemini({
-    model: process.env.MODEL as GEMINI_MODEL,
-  });
-  Settings.embedModel = new GeminiEmbedding({
-    model: process.env.EMBEDDING_MODEL as GEMINI_EMBEDDING_MODEL,
-  });
-}
-
-function initMistralAI() {
-  Settings.llm = new MistralAI({
-    model: process.env.MODEL as keyof typeof ALL_AVAILABLE_MISTRAL_MODELS,
-  });
-  Settings.embedModel = new MistralAIEmbedding({
-    model: process.env.EMBEDDING_MODEL as MistralAIEmbeddingModelType,
-  });
-}

templates/types/streaming/fastapi/app/api/routers/chat.py

@@ -1,8 +1,8 @@
 import logging
 from typing import List
 
-from fastapi import APIRouter, BackgroundTasks, Depends, HTTPException, Request, status
-from llama_index.core.chat_engine.types import BaseChatEngine, NodeWithScore
+from fastapi import APIRouter, BackgroundTasks, HTTPException, Request, status
+from llama_index.core.chat_engine.types import NodeWithScore
 from llama_index.core.llms import MessageRole
 
 from app.api.routers.events import EventCallbackHandler
@@ -58,11 +58,19 @@ async def chat(
 @r.post("/request")
 async def chat_request(
     data: ChatData,
-    chat_engine: BaseChatEngine = Depends(get_chat_engine),
 ) -> Result:
     last_message_content = data.get_last_message_content()
     messages = data.get_history_messages()
 
+    doc_ids = data.get_chat_document_ids()
+    filters = generate_filters(doc_ids)
+    params = data.data or {}
+    logger.info(
+        f"Creating chat engine with filters: {str(filters)}",
+    )
+
+    chat_engine = get_chat_engine(filters=filters, params=params)
+
     response = await chat_engine.achat(last_message_content, messages)
     return Result(
         result=Message(role=MessageRole.ASSISTANT, content=response.response),

templates/types/streaming/fastapi/app/api/routers/models.py

@@ -50,9 +50,14 @@ class AnnotationFileData(BaseModel):
         alias_generator = to_camel
 
 
+class AgentAnnotation(BaseModel):
+    agent: str
+    text: str
+
+
 class Annotation(BaseModel):
     type: str
-    data: AnnotationFileData | List[str]
+    data: AnnotationFileData | List[str] | AgentAnnotation
 
     def to_content(self) -> str | None:
         if self.type == "document_file":
@@ -119,14 +124,48 @@ class ChatData(BaseModel):
                 break
         return message_content
 
-    def get_history_messages(self) -> List[ChatMessage]:
+    def _get_agent_messages(self, max_messages: int = 10) -> List[str]:
+        """
+        Construct agent messages from the annotations in the chat messages
+        """
+        agent_messages = []
+        for message in self.messages:
+            if (
+                message.role == MessageRole.ASSISTANT
+                and message.annotations is not None
+            ):
+                for annotation in message.annotations:
+                    if annotation.type == "agent" and isinstance(
+                        annotation.data, AgentAnnotation
+                    ):
+                        text = annotation.data.text
+                        agent_messages.append(
+                            f"\nAgent: {annotation.data.agent}\nsaid: {text}\n"
+                        )
+                        if len(agent_messages) >= max_messages:
+                            break
+        return agent_messages
+
+    def get_history_messages(
+        self, include_agent_messages: bool = False
+    ) -> List[ChatMessage]:
         """
         Get the history messages
         """
-        return [
+        chat_messages = [
             ChatMessage(role=message.role, content=message.content)
             for message in self.messages[:-1]
         ]
+        if include_agent_messages:
+            agent_messages = self._get_agent_messages(max_messages=5)
+            if len(agent_messages) > 0:
+                message = ChatMessage(
+                    role=MessageRole.ASSISTANT,
+                    content="Previous agent events: \n" + "\n".join(agent_messages),
+                )
+                chat_messages.append(message)
+
+        return chat_messages
 
     def is_last_message_from_user(self) -> bool:
         return self.messages[-1].role == MessageRole.USER

templates/types/streaming/fastapi/pyproject.toml

@@ -9,7 +9,7 @@ readme = "README.md"
 generate = "app.engine.generate:generate_datasource"
 
 [tool.poetry.dependencies]
-python = "^3.11,<4.0"
+python = ">=3.11,<3.13"
 fastapi = "^0.109.1"
 uvicorn = { extras = ["standard"], version = "^0.23.2" }
 python-dotenv = "^1.0.0"

templates/types/streaming/nextjs/app/api/chat/route.ts

@@ -1,5 +1,5 @@
 import { initObservability } from "@/app/observability";
-import { JSONValue, Message, StreamData, StreamingTextResponse } from "ai";
+import { JSONValue, LlamaIndexAdapter, Message, StreamData } from "ai";
 import { ChatMessage, Settings } from "llamaindex";
 import { NextRequest, NextResponse } from "next/server";
 import { createChatEngine } from "./engine/chat";
@@ -12,7 +12,7 @@ import {
   createCallbackManager,
   createStreamTimeout,
 } from "./llamaindex/streaming/events";
-import { LlamaIndexStream } from "./llamaindex/streaming/stream";
+import { generateNextQuestions } from "./llamaindex/streaming/suggestion";
 
 initObservability();
 initSettings();
@@ -69,25 +69,37 @@ export async function POST(request: NextRequest) {
 
     // Setup callbacks
     const callbackManager = createCallbackManager(vercelStreamData);
+    const chatHistory: ChatMessage[] = messages as ChatMessage[];
 
     // Calling LlamaIndex's ChatEngine to get a streamed response
     const response = await Settings.withCallbackManager(callbackManager, () => {
       return chatEngine.chat({
         message: userMessageContent,
-        chatHistory: messages as ChatMessage[],
+        chatHistory,
         stream: true,
       });
     });
 
-    // Transform LlamaIndex stream to Vercel/AI format
-    const stream = LlamaIndexStream(
-      response,
-      vercelStreamData,
-      messages as ChatMessage[],
-    );
+    const onFinal = (content: string) => {
+      chatHistory.push({ role: "assistant", content: content });
+      generateNextQuestions(chatHistory)
+        .then((questions: string[]) => {
+          if (questions.length > 0) {
+            vercelStreamData.appendMessageAnnotation({
+              type: "suggested_questions",
+              data: questions,
+            });
+          }
+        })
+        .finally(() => {
+          vercelStreamData.close();
+        });
+    };
 
-    // Return a StreamingTextResponse, which can be consumed by the Vercel/AI client
-    return new StreamingTextResponse(stream, {}, vercelStreamData);
+    return LlamaIndexAdapter.toDataStreamResponse(response, {
+      data: vercelStreamData,
+      callbacks: { onFinal },
+    });
   } catch (error) {
     console.error("[LlamaIndex]", error);
     return NextResponse.json(

templates/types/streaming/nextjs/app/components/ui/chat/chat-message/chat-agent-events.tsx

@@ -17,6 +17,7 @@ const AgentIcons: Record<string, LucideIcon> = {
   researcher: icons.ScanSearch,
   writer: icons.PenLine,
   reviewer: icons.MessageCircle,
+  publisher: icons.BookCheck,
 };
 
 type MergedEvent = {

templates/types/streaming/nextjs/app/components/ui/chat/chat-message/markdown.tsx

@@ -133,7 +133,11 @@ export default function Markdown({
               return <></>;
             }
           }
-          return <a href={href}>{children}</a>;
+          return (
+            <a href={href} target="_blank">
+              {children}
+            </a>
+          );
         },
       }}
     >

templates/types/streaming/nextjs/app/markdown.css

@@ -2,11 +2,15 @@
 .custom-markdown ul {
   list-style-type: disc;
   margin-left: 20px;
+  margin-top: 20px;
+  margin-bottom: 20px;
 }
 
 .custom-markdown ol {
   list-style-type: decimal;
   margin-left: 20px;
+  margin-top: 20px;
+  margin-bottom: 20px;
 }
 
 .custom-markdown li {
@@ -21,3 +25,55 @@
 .custom-markdown ol ol {
   margin-left: 20px;
 }
+
+.custom-markdown img {
+  border-radius: 8px;
+  box-shadow: 0 0 10px rgba(0, 0, 0, 0.1);
+  margin: 10px 0;
+}
+
+.custom-markdown a {
+  text-decoration: underline;
+  color: #007bff;
+}
+
+.custom-markdown h1,
+h2,
+h3,
+h4,
+h5,
+h6 {
+  font-weight: bold;
+  margin-bottom: 20px;
+  margin-top: 20px;
+}
+
+.custom-markdown h6 {
+  font-size: 16px;
+}
+
+.custom-markdown h5 {
+  font-size: 18px;
+}
+
+.custom-markdown h4 {
+  font-size: 20px;
+}
+
+.custom-markdown h3 {
+  font-size: 22px;
+}
+
+.custom-markdown h2 {
+  font-size: 24px;
+}
+
+.custom-markdown h1 {
+  font-size: 26px;
+}
+
+.custom-markdown hr {
+  border: 0;
+  border-top: 1px solid #e1e4e8;
+  margin: 20px 0;
+}

templates/types/streaming/nextjs/package.json

@@ -12,12 +12,13 @@
   "dependencies": {
     "@apidevtools/swagger-parser": "^10.1.0",
     "@e2b/code-interpreter": "^0.0.5",
+    "@llamaindex/core": "^0.2.6",
     "@llamaindex/pdf-viewer": "^1.1.3",
     "@radix-ui/react-collapsible": "^1.0.3",
     "@radix-ui/react-hover-card": "^1.0.7",
     "@radix-ui/react-select": "^2.1.1",
     "@radix-ui/react-slot": "^1.0.2",
-    "ai": "3.0.21",
+    "ai": "3.3.42",
     "ajv": "^8.12.0",
     "class-variance-authority": "^0.7.0",
     "clsx": "^2.1.1",
@@ -25,7 +26,7 @@
     "duck-duck-scrape": "^2.2.5",
     "formdata-node": "^6.0.3",
     "got": "^14.4.1",
-    "llamaindex": "0.5.24",
+    "llamaindex": "0.6.2",
     "lucide-react": "^0.294.0",
     "next": "^14.2.4",
     "react": "^18.2.0",
@@ -41,7 +42,8 @@
     "tailwind-merge": "^2.1.0",
     "tiktoken": "^1.0.15",
     "uuid": "^9.0.1",
-    "vaul": "^0.9.1"
+    "vaul": "^0.9.1",
+    "marked": "^14.1.2"
   },
   "devDependencies": {
     "@types/node": "^20.10.3",