Skip to content

This project demonstrates a basic chain-of-thought interaction with any LLM (Large Language Model)

Notifications You must be signed in to change notification settings

antibitcoin/ReflectionAnyLLM

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

40 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ReflectionAnyLLM

A preview of it used with Gemma 2 9b LLM

Online Demo: freemindgpt.com/reflectionanyllm
Important: This demo is running on OpenRouter's free tier, utilizing Gemini-pro-1.5, . While it showcases basic functionality, for more advanced use cases, models like GPT-4o, LLaMA 3.1 70b, or 405b would offer far better performance. The demo may also occasionally stop working if usage limits are exceeded.


Project Overview

This was requested to share with the public on the following Reddit thread by me: https://www.reddit.com/r/LocalLLaMA/comments/1fgo671/openai_sent_me_an_email_threatening_a_ban_if_i/

ReflectionAnyLLM is a lightweight proof-of-concept that enables basic chain-of-thought (CoT) reasoning with any Large Language Model (LLM) that supports the OpenAI-compatible API. The flexibility of this project allows it to interface with local or remote LLMs, giving you the ability to switch between providers with minimal setup.

The project can be integrated with multiple LLM providers, including but not limited to:

  • LM Studio
  • Groq
  • OpenRouter
  • OpenAI API

By replacing the API key and endpoint, you can easily adapt this project to work with other LLMs that support OpenAI's API format.


Features

  • Chain-of-Thought Reasoning: ReflectionAnyLLM implements a reasoning process with steps dynamically determined by the LLM itself. The model can create up to 10 reasoning steps, and this limit can be increased by editing the chat.php file.
  • Thought Process Summary: The detailed thought process is hidden by default to keep the page and history neat, but a summary is displayed for quick reference.
  • History Management: The project keeps track of the last 30 messages to maintain a concise history and optimize performance.
  • Strawberry Count Problem: The model can solve tasks like the "strawberry count" problem on almost any model larger than 8 billion parameters.
  • API Flexibility: Easily switch between different LLM providers by modifying the API endpoint and key.
  • Simple Setup: The project is lightweight and doesn't require a complex database or backend setup.
  • Downloadable Chat History: Users can download their chat history as JSON for reference or further processing.
  • Clear Chat: Clear your chat history with a simple click.

Getting Started

To get started with ReflectionAnyLLM, you’ll need a local web server. The following servers are recommended based on your operating system:

  • Windows: XAMPP
  • Mac: MAMP
  • Linux: Use the LAMP stack.

Note: MySQL is not required, but your server must have PHP installed with the curl extension enabled. Running php -S (PHP's built-in server) will not work out of the box, as this project relies on the curl extension to handle API requests.

Prerequisites

Before running the project, ensure you have the following:

  1. Web Server: Download and install one of the following:
    • XAMPP for Windows
    • MAMP for Mac
    • LAMP for Linux
  2. PHP with curl: Ensure the curl extension is enabled in your PHP configuration. This is critical for handling API requests.

Installation Instructions

  1. Clone this repository or download the project files.

  2. Move the index.html and chat.php files to your web server’s public directory:

    • XAMPP: htdocs/
    • MAMP: htdocs/
    • LAMP: /var/www/
  3. Make sure the curl extension is active on your PHP server:

    • In your php.ini file, ensure that extension=curl is not commented out.
  4. Once set up, navigate to the project directory through your browser:

    • Local Setup: Open http://localhost/ in your web browser.
    • Online Hosting: Upload the files to your remote server and access it via your domain name.

Directory Structure

ReflectionAnyLLM/
├── index.html   # Front-end HTML file
└── chat.php     # Backend PHP script for handling API requests

Usage

The front-end interacts with the back-end PHP script, which makes API requests to the LLM. You can modify the API settings by editing the chat.php file, where you can:

  • Change the API endpoint: Replace the OpenAI-compatible API URL.
  • Update the API key: Input your API key for the LLM you are using (e.g., OpenRouter, Groq, etc.).

This flexibility allows the project to be used with any compatible LLM service, enabling quick testing and iteration.


Notes & Considerations

This project was developed as a quick demo and should be considered a prototype. Here are some key things to note:

  • Reasoning Steps: The number of reasoning steps is dynamically decided by the LLM and can go up to 10. You can increase this limit by editing the logic in chat.php.
  • Chain of Thought (CoT) Reasoning: The current implementation uses dynamic reasoning steps to simulate how CoT works with larger models like GPT-4o or LLaMA 3.1. It can perform simple reasoning tasks, such as solving the "strawberry count" problem on models larger than 8 billion parameters.
  • Thought Process Summary: To keep the history and page clean, the full thought process is hidden, and only a summary is shown. Users can toggle to see the detailed process if needed.
  • Security Warning: The code, especially the back-end chat.php, is not secured and is not suitable for production environments. It was created quickly to demonstrate basic functionality and lacks robust security practices.
  • Improvement Potential: Although this project is functional, it is by no means polished or optimized. If you're interested in improving it, feel free to fork the repository and submit a pull request. I'm open to reviewing and merging community contributions.

Future Improvements

While this project serves as a simple demonstration of basic LLM interactions, it can be expanded in several ways:

  1. Enhanced Chain-of-Thought: Extend the CoT process to allow for more complex reasoning and interaction.
  2. UI/UX Improvements: Improve the front-end design and add features like loading indicators, download history functionality, clear history button, and better error handling.
  3. Security Enhancements: Implement more secure coding practices, especially on the back-end, to prevent potential vulnerabilities.

If there’s interest from the community, I may continue developing and improving this project. Feedback and suggestions are welcome!


Available Ports to Other Languages

ReflectionAnyLLM can be easily adapted to various programming languages. If you create a port in another language, feel free to submit it by opening an issue on the repository, and I'll gladly list it here. Don't forget to attribute the original project when submitting your port!

Please note that different languages provide different results but the basic is the same, for example some might use a termnial, some might be a webui.

Here are the currently available ports:

  1. Python Port by devinambron: PyThoughtChain - A Python implementation of the ReflectionAnyLLM chain-of-thought reasoning.
  2. CoT-at-Home by andrewginns: Cot-at-Home - This project enables any OpenAI-compatible API to automatically have Chain-of-Thought (CoT) reasoning. Allowing for transparently added CoT to use with any model.

Contributing

I created this project based on a request from Reddit, and I’m happy to share it with the community. If you find ways to improve it or want to contribute new features, please fork the repository and submit a pull request.


License

This project is licensed under the MIT License, meaning you're free to use, modify, and distribute it, as long as attribution is provided.


Feel free to explore, modify, and share your improvements! If you have any questions or encounter issues, don’t hesitate to reach out.

About

This project demonstrates a basic chain-of-thought interaction with any LLM (Large Language Model)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published