Generative AI OpenAI API Server

A lightweight server that translates Gemini API calls into OpenAI API-compatible format.

Features

This project provides an alternative to using Google's /v1beta/openai/ endpoint by addressing its limitations, offering enhanced functionality, and extending support for key features.

Why not use /v1beta/openai/ directly?

While Google does provide a partially OpenAI-compatible API, there are significant limitations:

1. Unsupported Endpoints: Many endpoints, such as /v1/models, are not available.
2. Limited Parameters: Important parameters like "frequency_penalty", "presence_penalty", and "stop" are not supported. When unsupported parameters are included, the API throws an error instead of gracefully ignoring them.
3. Missing Advanced Features: Features like context caching and advanced safety configurations from Gemini API are absent.

This server addresses these issues by acting as a middleware between your application and the Gemini API.

Getting Started

Prerequisites

• Download the latest release binary.

Steps to Use

1. Create a configuration file (genai.config.json).
2. Run the server. The default port is 4949.

Custom Config

main.exe -c my_owner.config.json

Configuration

The server uses a JSON-based configuration file. Below is a basic example:

{
  "api_key": "sk-xxx",
  "server": {
    "port": 4949
  }
}

Full Configuration Options

Here is a complete list of configurable parameters:

type Params = {
  api_key: string;
  server?: {
    port?: number;
  };
  no_docs?: boolean;
  retry?: {
    enabled?: boolean;
    retries?: number;
    factor?: number;
    minTimeout?: number;
    maxTimeout?: number;
  };
  debug?: {
    stream?: {
      log?: boolean;
    };
  };
  bodyLimit?: number;
};

Supported Endpoints

The server currently supports the following endpoints:

• /v1/models: Retrieve available model list.
• /v1/embeddings: Generate vector embeddings for input text.
• /v1/chat/completions: Chat-based text completions.

Note: /v1/completions is not supported because Gemini models do not support completion functionality, and Google's PaLM model (which does) is likely to be deprecated.

Roadmap

• v1
• 支持配置多个 api key ，并且可以轮询、分配权限、重试等调度操作
• 支持配置和切换代理

Building the Project

pnpm run build:ci

License

This project is licensed under the MIT License.