Groq-Proxy: A reverse-engineered API for Groq using NodeJS

Turn GroqChat webui into a free OpenAI Compatible API. Note: you need to login to your own Groq account

❌: (2024/05/28) This project is pretty pointless now because Groq API has implemented a rather generous rate limit on its API. In addition, I highly suspect that the API you get from this project (chat website reverse engineering) shares the same rate limit as the official API. So...

⚠️ This project is still under active development. Some features (like sign-in) can be unstable.

⚠️ all content in this repository is for experimental purposes

This program offers free self-hosted API access to Groq (which has super fast inference for open source LLM like LLaMA3 70b, Mixtral 8x7b, and some other models) with OpenAI compatible format. Just reset the endpoint.

Installation

• Clone this repo and open the repo's folder
• npm install (some puppeteer related settings may need attention)
• node index.js to run this program.
  • If some chrome or puppeteer thing broke, run this npx puppeteer browsers install chrome
  • You will need to sign in to groq.
  • Follow the instructions. Use your own browser to go to groq.com and sign in using your email (don't actually sign in there).
  • After entering your email, a verification email will be sent. Copy the link (don't open it)
  • Paste the link into the program
  • The cookies will be extracted, and you are good to go

Note:

• Cookies will expire. You may need to restart the server and re-authenticate to get the new cookies.

Sample

terminal code to use the api. Because it uses SSE to stream the data back, you will see a bunch of json appearing on your terminal.

curl --request POST \
  --url 'http://localhost:9876/v1/chat/completions?=' \
  -H Accept:text/event-stream
  --data '{
    "model": "llama3-70b-8192",
    "messages": [
        {
            "content": "",
            "role": "system"
        },
        {
            "content": "Hey buddy",
            "role": "user"
        }
    ],
    "temperature": 0.2,
    "max_tokens": 2048,
    "top_p": 0.8,
    "stream": true
}
'

Implemented

• Sign in to the groq account
• Verify Sign-in
• Query LLM
• Start to transition from puppeteer emulation to direct api calls
• Text Completion with Role/System prompt settings 但我这个要怎么设role啊... 还是要抓包吗... (2024.04.21: 是的, 要抓包) (2024.04.22: 报告，抓好了)
• Switch model
• Edit System Prompt
• View Models (now using http get instead of puppeteer)
• Implement API server based on groq api / openai api / ollama, potentially using http-proxy
  • Implemented /v1/chat/completions and /v1/models api, which are OpenAI Compatible. Just have the server running, set your OpenAI Endpoint to http://localhost:9876, and you should be good to go.
  • /v1/chat/completions will be streamed back using SSE
• New way of signing in: The user now only needs to provide the verification link

To-do

• remove unnecessary code
• Fake model name
  • because groq has a rather different model naming scheme, and that the existing tools may have pre-defined the model name, when using this server, the model name might be incorrect. Add a alias feature to redirect some model name to another. For example, map gpt-3.5-turbo -> llama3-70b-8192
• add checks for the validity of cookies for the API
• Create no puppeteer mode (just provide the cookies or something similar)
• dockerize this thing