feat: Add publisher agent, merge code with streaming template (#324)

---------
Co-authored-by: Marcus Schiesser <[email protected]>

Author: leehuwuj

.changeset/flat-singers-share.md

@@ -0,0 +1,5 @@
+---
+"create-llama": patch
+---
+
+Add publisher agent to multi-agents for generating documents (PDF and HTML)

.changeset/gorgeous-penguins-shout.md

@@ -0,0 +1,5 @@
+---
+"create-llama": patch
+---
+
+Allow tool selection for multi-agents (Python and TS)

e2e/python/resolve_dependencies.spec.ts

@@ -33,6 +33,7 @@ if (
   const toolOptions = [
     "wikipedia.WikipediaToolSpec",
     "google.GoogleSearchToolSpec",
+    "document_generator",
   ];
 
   const dataSources = [

e2e/utils.ts

@@ -109,7 +109,9 @@ export async function runCreateLlama({
   if (appType) {
     commandArgs.push(appType);
   }
-  if (!useLlamaParse) {
+  if (useLlamaParse) {
+    commandArgs.push("--use-llama-parse");
+  } else {
     commandArgs.push("--no-llama-parse");
   }
 

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/python.ts

@@ -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

@@ -157,7 +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 {

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"
@@ -734,8 +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.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 importlib
 import os
+
 import yaml
-import importlib
-from llama_index.core.tools.tool_spec.base import BaseToolSpec
 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 @@ def load_tools(tool_type: str, tool_name: str, config: dict) -> list[FunctionToo
             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