app.py

from fastapi import FastAPI, HTTPException
from onnxruntime import InferenceSession
from transformers import AutoTokenizer
import numpy as np
import os

app = FastAPI()

# Initialize tokenizer (doesn't require PyTorch/TensorFlow)
tokenizer = AutoTokenizer.from_pretrained(
    "Xenova/multi-qa-mpnet-base-dot-v1",
    use_fast=True,  # Uses Rust implementation
    legacy=False
)

# Load ONNX model
session = InferenceSession("model.onnx")

@app.post("/api/predict")
async def predict(text: str):
    try:
        # Tokenize without framework dependencies
        inputs = tokenizer(
            text,
            return_tensors="np",  # Get NumPy arrays directly
            padding=True,
            truncation=True,
            max_length=32  # Match your model's expected input size
        )
        
        # Prepare ONNX inputs
        onnx_inputs = {
            "input_ids": inputs["input_ids"].astype(np.int64),
            "attention_mask": inputs["attention_mask"].astype(np.int64)
        }
        
        # Run inference
        outputs = session.run(None, onnx_inputs)
        
        # Convert to native Python types
        return {
            "embedding": outputs[0].astype(np.float32).tolist(),
            "tokens": tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
        }
        
    except Exception as e:
        raise HTTPException(status_code=400, detail=str(e))