app.py

import faiss
import numpy as np
from FlagEmbedding import FlagModel
from flask import Flask, request, jsonify
from datasets import load_dataset
import gradio as gr
import os
import time
from functools import lru_cache

# Initialize components
app = Flask(__name__)
model = None
index = None
corpus = None

def initialize_components():
    global model, index, corpus
    
    # Load model with safety checks
    if model is None:
        model = FlagModel(
            "BAAI/bge-large-en-v1.5",
            query_instruction_for_retrieval="Represent this sentence for searching relevant passages:",
            use_fp16=True
        )
    
    # Load corpus from Hugging Face dataset
    if corpus is None:
        dataset = load_dataset("awinml/medrag_corpus_sampled", split='train')
        corpus = [f"{row['id']}\t{row['contents']}" for row in dataset]
    
    # Create FAISS index in memory
    if index is None:
        embeddings = model.encode([doc.split('\t', 1)[1] for doc in corpus])
        dimension = embeddings.shape[1]
        index = faiss.IndexFlatIP(dimension)
        index.add(embeddings.astype('float32'))

@app.route("/retrieve", methods=["POST"])
def retrieve():
    start_time = time.time()
    
    # Validate request
    data = request.json
    if not data or "queries" not in data:
        return jsonify({"error": "Missing 'queries' parameter"}), 400
    
    # Initialize components if needed
    initialize_components()
    
    # Process queries
    queries = data["queries"]
    topk = data.get("topk", 3)
    return_scores = data.get("return_scores", False)
    
    # Batch processing
    query_embeddings = model.encode_queries(queries)
    scores, indices = index.search(query_embeddings.astype('float32'), topk)
    
    # Format results
    results = []
    for i, query in enumerate(queries):
        query_results = []
        for j in range(topk):
            doc_idx = indices[i][j]
            doc = corpus[doc_idx]
            doc_id, content = doc.split('\t', 1)
            result = {
                "document": {
                    "id": doc_id,
                    "contents": content
                },
                "score": float(scores[i][j])
            }
            query_results.append(result)
        results.append(query_results)
    
    return jsonify({
        "result": results,
        "time": f"{time.time() - start_time:.2f}s"
    })

# Gradio UI for testing
def gradio_interface(query, topk):
    response = requests.post(
        "http://localhost:7860/retrieve",
        json={"queries": [query], "topk": topk}
    )
    return response.json()["result"][0]

# Start server
if __name__ == "__main__":
    # First-time initialization
    initialize_components()
    
    # Create Gradio interface
    iface = gr.Interface(
        fn=gradio_interface,
        inputs=[
            gr.Textbox(label="Medical Query", placeholder="Enter your medical question..."),
            gr.Slider(1, 10, value=3, label="Top Results")
        ],
        outputs=gr.JSON(label="Retrieval Results"),
        title="Medical Retrieval System",
        description="Search across medical literature using AI-powered semantic search"
    )
    
    # Run both Flask and Gradio
    iface.launch(server_name="0.0.0.0", server_port=7860, share=True)