Toolpacks

Toolpacks provide a way to group related tools together. This is useful for organizing your tools and controlling their availability within a conversation. Instead of listing every single tool directly in the system prompt, you can expose logical groups that the LLM can choose to activate.

Defining a Toolpack

Similar to Tool, you use a fluent API starting with Toolpack.make().

import { Tool, Toolpack } from '@obayd/agentic';

// Define some tools first
const searchFilesTool = Tool.make("search_files")
    .description("Searches user's files by keyword.")
    .param("keyword", "Search keyword", { required: true })
    .action(async (params) => `File A matching ${params.keyword}`);

const readFileTool = Tool.make("read_file")
    .description("Reads the content of a specific file.")
    .param("filename", "The name of the file to read.", { required: true })
    .action(async (params) => ({ content: `Content of ${params.filename}...` }));

// Create a Toolpack and add the tools
const fileManagementPack = Toolpack.make("file_management") // Unique name
    .description("Tools for searching and reading user files.") // Description of the pack
    .add(searchFilesTool, readFileTool, /* ... */) // Add a single tool, or multiple tools

// You can also chain tool definitions directly
const webToolsPack = Toolpack.make("web_tools")
    .description("Tools for interacting with the web.")
    .add(
        Tool.make("browse_page")
            .description("Gets the content of a webpage.")
            .param("url", "URL to browse", { required: true })
            .action(async (params) => ({ content: `Content of ${params.url}...` }))
    );

Configuration Methods

  • .make(name: string): (Static) Starts defining a toolpack with a unique name (letters, numbers, underscores).

  • .description(description: string): Sets a description for the entire toolpack, explaining the capabilities it provides. Alias: .desc().

  • .addTool(tool: Tool): Adds a single, already defined Tool instance to the pack.

  • .addTools(tools: Tool[]): Adds an array of Tool instances to the pack.

Using Toolpacks in Conversation

  • When you include a Toolpack instance in your conversation.content() definition, the framework handles it differently than a direct Tool:

  • Automatic enable_toolpack Tool: If any Toolpack is defined in the content, @obayd/agentic automatically adds a built-in tool named enable_toolpack.

  • Toolpack Listing in Prompt: The system prompt will list the available toolpacks (along with their descriptions and whether they are currently enabled or disabled) instead of listing every tool within those packs initially.

  • Enabling Workflow:

    • The LLM decides it needs a capability provided by a disabled toolpack (based on the pack's description).

    • The LLM calls the enable_toolpack tool, providing the pack_name parameter.

    • The Conversation executes this internal tool, adding the specified pack name to the conversation.enabledToolpacks set.

    • A success message is returned to the LLM in the next turn .

  • Crucially: The LLM must wait for this success message before attempting to use tools from the newly enabled pack.

Using Enabled Tools

Once a toolpack is enabled (either initially via initialEnabledToolpacks option or via the enable_toolpack tool), its contained tools become available for the LLM to call directly in subsequent turns, just like tools added directly to the content.

import { Conversation, Tool, Toolpack } from '@obayd/agentic';
// ... (llmCallback, fileManagementPack definitions from above) ...

const conversation = new Conversation(llmCallback);

conversation.content([
    "You are an assistant that can manage files.",
    fileManagementPack // Include the toolpack
]);

async function run() {
    // Turn 1: User asks to search files
    const stream1 = conversation.send("Search my files for 'report'");
    for await (const event of stream1) {
        // Expect LLM to call enable_toolpack(pack_name="file_management")
        // ... process events ...
    }

    // Turn 2: After enable_toolpack succeeds, user confirms or LLM directly calls search_files
    // (Assuming LLM was instructed or decides to proceed)
    // If the LLM needs another prompt, you might send a confirmation or just wait for it to call the tool.
    // For this example, let's assume the LLM calls search_files in the next reasoning step
    // (This depends heavily on the LLM's behavior and prompting)

    // The conversation loop internally handles sending the result of enable_toolpack back.
    // Now, the LLM receives the success message and *can* call search_files.
    // Let's simulate the LLM calling the tool in the next step (often it will respond first)

    // Example: If the previous turn's assistant message was just confirmation,
    // send an empty message or specific instruction to trigger the tool use.
    // For simplicity, we assume the LLM calls it directly after getting the enable confirmation.
    // (In reality, the assistant might reply "Okay, I've enabled the file tools. Now searching..."
    // and then immediately issue the search_files call if prompted well)

    // Let's assume the second turn resulted in the LLM calling search_files
    // The stream processing would look like:
    // 1. tool.calling event for search_files
    // 2. tool event with search_files results
    // 3. assistant event with the final answer using the search results.
}

run();

Benefits of Toolpacks

  • Scalability: Keeps the initial system prompt cleaner when you have many tools.

  • Organization: Groups related functionalities logically.

  • Control: Allows for dynamic enabling/disabling of capabilities during a conversation.

  • Contextual Relevance: The LLM only activates tools when needed for a specific task group.

PreviousToolsNextLLM Integration

Last updated