app.py

import os
import streamlit as st
from pinecone import Pinecone
from sentence_transformers import SentenceTransformer
import torch
from splade.models.transformer_rep import Splade
from transformers import AutoTokenizer

# Title of the Streamlit App
st.title("Medical Hybrid Search")

# Initialize Pinecone globally
index = None

# Function to initialize Pinecone
def initialize_pinecone():
    api_key = os.getenv('PINECONE_API_KEY')  # Get Pinecone API key from environment variable
    if api_key:
        pc = Pinecone(api_key=api_key)
        return pc
    else:
        st.error("Pinecone API key not found! Please set the PINECONE_API_KEY environment variable.")
        return None

# Function to connect to the 'pubmed-splade' index
def connect_to_index(pc):
    index_name = 'pubmed-splade'  # Hardcoded index name
    if index_name in pc.list_indexes().names():
        index = pc.Index(index_name)
        return index
    else:
        st.error(f"Index '{index_name}' not found!")
        return None

# Function to encode query using sentence transformers model
def encode_query(model, query_text):
    return model.encode(query_text).tolist()

# Function to create hybrid scaled vectors
def hybrid_scale(dense, sparse, alpha):
    if alpha < 0 or alpha > 1:
        raise ValueError("Alpha must be between 0 and 1")
    hsparse = {
        'indices': sparse['indices'],
        'values': [v * (1 - alpha) for v in sparse['values']]
    }
    hdense = [v * alpha for v in dense]
    return hdense, hsparse

# Initialize Pinecone
pc = initialize_pinecone()

# If Pinecone initialized successfully, proceed with index management
if pc:
    # Connect directly to 'pubmed-splade' index
    index = connect_to_index(pc)

    # Model for query encoding
    model = SentenceTransformer('msmarco-bert-base-dot-v5')
    
    # Initialize sparse model and tokenizer
    sparse_model_id = 'naver/splade-cocondenser-ensembledistil'
    sparse_model = Splade(sparse_model_id, agg='max')
    sparse_model.eval()  # Set the model to evaluation mode
    tokenizer = AutoTokenizer.from_pretrained(sparse_model_id)

    # Query input
    query_text = st.text_input("Enter a Query to Search", "Can clinicians use the PHQ-9 to assess depression?")
    
    # Alpha input
    alpha = st.slider("Set Alpha (for dense and sparse vector balancing)", 0.0, 1.0, 0.5)

    # Button to encode query and search the Pinecone index
    if st.button("Search Query"):
        if query_text and index:
            # Encode query to get dense and sparse vectors
            dense_vector = encode_query(model, query_text)
            input_ids = tokenizer(query_text, return_tensors='pt')
            with torch.no_grad():
                sparse_vector = sparse_model(d_kwargs=input_ids.to('cpu'))['d_rep'].squeeze()

            # Prepare sparse vector format for Pinecone
            indices = sparse_vector.nonzero().squeeze().cpu().tolist()
            values = sparse_vector[indices].cpu().tolist()
            sparse_dict = {"indices": indices, "values": values}

            # Scale dense and sparse vectors
            hdense, hsparse = hybrid_scale(dense_vector, sparse_dict, alpha)

            # Search the index
            results = index.query(
                vector=hdense,
                sparse_vector=hsparse,
                top_k=3,
                include_metadata=True
            )
            
            st.write("Results:")
            for match in results.matches:
                st.markdown(f"Score: {match.score:.4f}")
                st.write(f"Answer: {match.metadata.get('context', 'No context available.')}")
                st.write("---")
        else:
            st.error("Please enter a query and ensure the index is initialized.")