app.py

import os
import torch
import numpy as np
from transformers import AutoTokenizer, AutoModel
from flask import Flask, request, jsonify
import logging

# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

app = Flask(__name__)

# Qwen3-Embedding-4B model for retrieval
MODEL_NAME = "Qwen/Qwen3-Embedding-4B"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
EMBEDDING_DIM = 2560  # Max dimension for Qwen3-Embedding-4B

class EmbeddingModel:
    def __init__(self):
        logger.info(f"Loading {MODEL_NAME} on {DEVICE}")
        self.tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
        self.model = AutoModel.from_pretrained(MODEL_NAME)
        self.model.to(DEVICE)
        self.model.eval()
        logger.info("✅ Model loaded successfully")
    
    def encode(self, texts, batch_size=16):
        """Encode texts to embeddings using Qwen3-Embedding-4B"""
        if isinstance(texts, str):
            texts = [texts]
        
        embeddings = []
        
        for i in range(0, len(texts), batch_size):
            batch_texts = texts[i:i + batch_size]
            
            # Qwen3 instruction format for retrieval
            batch_texts = [f"Instruct: Retrieve semantically similar text.\nQuery: {text}" for text in batch_texts]
            
            inputs = self.tokenizer(
                batch_texts,
                padding="longest",  
                truncation=True,
                max_length=32768,  # Qwen3 supports up to 32k context
                return_tensors="pt"
            ).to(DEVICE)
            
            with torch.no_grad():
                outputs = self.model(**inputs)
                # Use EOS token embedding for Qwen3
                eos_token_id = self.tokenizer.eos_token_id
                sequence_lengths = (inputs['input_ids'] == eos_token_id).long().argmax(-1) - 1
                
                batch_embeddings = []
                for j, seq_len in enumerate(sequence_lengths):
                    embedding = outputs.last_hidden_state[j, seq_len, :].cpu().numpy()
                    batch_embeddings.append(embedding)
                
                batch_embeddings = np.array(batch_embeddings)
                
                # Normalize embeddings
                batch_embeddings = batch_embeddings / np.linalg.norm(batch_embeddings, axis=1, keepdims=True)
                
                embeddings.extend(batch_embeddings)
        
        return embeddings

# Global model instance
embedding_model = None

def get_model():
    global embedding_model
    if embedding_model is None:
        embedding_model = EmbeddingModel()
    return embedding_model

@app.route("/", methods=["GET"])
def health_check():
    return jsonify({
        "status": "healthy",
        "model": MODEL_NAME,
        "device": DEVICE,
        "embedding_dim": EMBEDDING_DIM,
        "max_context": 32768
    })

@app.route("/embed", methods=["POST"])
def embed_texts():
    """Embed texts and return embeddings"""
    try:
        data = request.get_json()
        
        if not data or "texts" not in data:
            return jsonify({"error": "Missing 'texts' field"}), 400
        
        texts = data["texts"]
        if not isinstance(texts, list):
            texts = [texts]
        
        logger.info(f"Embedding {len(texts)} texts")
        
        model = get_model()
        embeddings = model.encode(texts)
        
        return jsonify({
            "embeddings": [embedding.tolist() for embedding in embeddings],
            "model": MODEL_NAME,
            "dimension": len(embeddings[0]) if embeddings else 0,
            "count": len(embeddings)
        })
        
    except Exception as e:
        logger.error(f"Embedding error: {str(e)}")
        return jsonify({"error": str(e)}), 500

@app.route("/embed_single", methods=["POST"])
def embed_single():
    """Embed single text (convenience endpoint)"""
    try:
        data = request.get_json()
        
        if not data or "text" not in data:
            return jsonify({"error": "Missing 'text' field"}), 400
        
        text = data["text"]
        logger.info(f"Embedding single text: {text[:100]}...")
        
        model = get_model()
        embeddings = model.encode([text])
        
        return jsonify({
            "embedding": embeddings[0].tolist(),
            "model": MODEL_NAME,
            "dimension": len(embeddings[0]),
            "text_length": len(text)
        })
        
    except Exception as e:
        logger.error(f"Single embedding error: {str(e)}")
        return jsonify({"error": str(e)}), 500

if __name__ == "__main__":
    # Initialize model on startup
    logger.info("🚀 Starting embedding service...")
    get_model()
    logger.info("✅ Service ready!")
    
    port = int(os.environ.get("PORT", 7860))
    app.run(host="0.0.0.0", port=port)