Other Tools: think, serve, and native

This document covers three distinct types of tools: a cognitive tool for the LLM, a simple web server, and a way to access the native, built-in capabilities of the model provider.

The think Tool

The think tool gives the LLM a private “scratch space” to pause and reason about a prompt before formulating its final response. This can improve the quality and thoughtfulness of the output, especially for complex or ambiguous questions.

This technique was inspired by a post on Anthropic’s engineering blog. The output of the think tool is hidden from the user by default in the editor plugins, though it is still present in the .lec file.

Configuration

tools:
  - think_about: >
      What the user is really asking for, and what hidden assumptions they
      might have.
    name: scratchpad # Optional name

Example

What's the best city in the world?

:::Assistant

<tool-call with="scratchpad">
<arguments>
<thought>
┆"Best" is subjective. The user could mean best for travel, for
┆food, for work, etc. I need to ask for clarification.
</thought>
</arguments>
<results>
<result type="text">
┆thought complete.
</result>
</results>
</tool-call>

That depends on what you're looking for! Are you interested in the best city
for tourism, career opportunities, or something else?
:::

The serve Tool

The serve tool allows the LLM to spin up a simple, single-use web server to present content, such as an HTML file or a small web application it has generated.

When the LLM uses this tool, Lectic starts a server on the specified port. It will then attempt to open the page in your default web browser. The server shuts down automatically after the first request is served. While the page is loading, Lectic waits for the first request—so the conversation resumes once your browser has loaded the page.

Configuration

tools:
  - serve_on_port: 8080
    name: web_server # Optional name

Example

Generate a simple tic-tac-toe game in HTML and serve it to me.

:::Assistant

<tool-call with="web_server">
<arguments>
<pageHtml>
┆<!DOCTYPE html>
┆<html>
┆<head>
┆<title>Tic-Tac-Toe</title>
┆... (rest of the HTML/JS/CSS) ...
┆</head>
┆<body>
┆...
┆</body>
┆</html>
</pageHtml>
</arguments>
<results>
<result type="text">
┆page is now available
</result>
</results>
</tool-call>


I have generated the game for you. It should be opening in your browser at
http://localhost:8080.
:::

native Tools

Native tools allow you to access functionality that is built directly into the LLM provider’s backend, such as web search or a code interpreter environment for data analysis.

Support for native tools varies by provider.

Configuration

You enable native tools by specifying their type.

tools:
  - native: search # Enable the provider's built-in web search.
  - native: code   # Enable the provider's built-in code interpreter.

Provider Support

  • Gemini: Supports both search and code. Note that the Gemini API has a limitation where you can only use one native tool at a time, and it cannot be combined with other (non-native) tools.
  • Anthropic: Supports search only.
  • OpenAI: Supports both search and code via the openai provider (not the legacy openai/chat provider).