app.py

from fastapi import FastAPI
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModel
import torch

app = FastAPI(
    title="OpenAI-compatible Embedding API",
    version="1.0.0",
)

# Load model from Hugging Face Hub
MODEL_NAME = "BAAI/bge-small-en-v1.5"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModel.from_pretrained(MODEL_NAME)
model.eval()

class EmbeddingRequest(BaseModel):
    input: list[str]


@app.get("/")
def root():
    return {"message": "API is working"}


@app.post("/embeddings")
def create_embeddings(request: EmbeddingRequest):
    with torch.no_grad():
        tokens = tokenizer(request.input, return_tensors="pt", padding=True, truncation=True)
        output = model(**tokens)
        cls_embeddings = output.last_hidden_state[:, 0]
        norm_embeddings = torch.nn.functional.normalize(cls_embeddings, p=2, dim=1)

    data = [
        {
            "object": "embedding",
            "embedding": e.tolist(),
            "index": i
        }
        for i, e in enumerate(norm_embeddings)
    ]

    return {
        "object": "list",
        "data": data,
        "model": MODEL_NAME,
        "usage": {
            "prompt_tokens": sum(len(tokenizer.encode(x)) for x in request.input),
            "total_tokens": sum(len(tokenizer.encode(x)) for x in request.input),
        }
    }