FastOAI (OpenAI-like API Server)

Motivation

This project is a simple API server that can be used to serve language models. It is designed to be simple to use and easy to deploy. It is built using FastAPI, Pydantic and openai-python.

Quick Start

The following example shows a case to overriding the original streaming response, we append [ollama] to every chunk content.

from fastapi.responses import StreamingResponse
from fastoai import FastOAI
from fastoai.requests import CompletionCreateParams
from fastoai.routers import router
from openai import AsyncOpenAI

app = FastOAI()


@app.post("/chat/completions")
async def create_chat_completions(params: CompletionCreateParams):
    client = AsyncOpenAI(api_key="ollama", base_url="http://localhost:11434/v1")
    response = await client.chat.completions.create(**params.model_dump())
    if params.stream:

        async def _stream():
            async for chunk in response:
                for choice in chunk.choices:
                    choice.delta.content += " [ollama]"
                yield f"data: {chunk.model_dump_json()}\n\n"
            else:
                yield "data: [DONE]\n\n"

        return StreamingResponse(_stream())
    return response

# Order matters, you should firstly define your own entrypoint functions, then
# include the existing router in order to override the original process with
# yours
app.include_router(router)

And that's it! You can now run the server using uvicorn:

uvicorn examples.main:app --reload

Architecture

graph TD
    client["`*openai client* (Python/Node.js/RESTful)`"]-->fastoai[FastOAI server]
    fastoai-->ChatCompletion[OpenAI Chat Completion API]
    fastoai-->openai[OpenAI API]
    fastoai-->others[Any other API]