qamatcher_server.py

import os
from flask import request, jsonify
from sentence_transformers import SentenceTransformer, util

# Define a writable directory for the model cache.
# This now respects the HF_HOME environment variable set in the Dockerfile.
cache_dir = os.environ.get("HF_HOME", "/tmp/huggingface")
os.makedirs(cache_dir, exist_ok=True)


print("Loading SentenceTransformer model (paraphrase-MiniLM-L6-v2)...")
matcher_model = SentenceTransformer('sentence-transformers/paraphrase-MiniLM-L6-v2', cache_folder=cache_dir)
print("SentenceTransformer model loaded.")

# Define a threshold for a "good" match
SIMILARITY_THRESHOLD = 0.6

def handle_match_question():
    data = request.get_json()
    if not data or 'user_question' not in data or 'documents' not in data:
        return jsonify({'error': 'Invalid request. "user_question" and "documents" are required.'}), 400

    user_question = data['user_question']
    documents = data['documents']

    if not documents:
        return jsonify({'answer': "There are no notes to search."})

    # Flatten the list of questions from all documents
    all_questions = []
    # Map each question to the original note text
    question_to_note_map = {}

    for doc in documents:
        note_text = doc.get('note_text', '')
        for q in doc.get('questions', []):
            all_questions.append(q)
            question_to_note_map[q] = note_text
    
    if not all_questions:
        return jsonify({'answer': "No questions have been generated for your notes yet."})

    try:
        # Encode the user's question and all stored questions
        user_embedding = matcher_model.encode(user_question, convert_to_tensor=True)
        stored_embeddings = matcher_model.encode(all_questions, convert_to_tensor=True)

        # Compute cosine similarity
        cosine_scores = util.pytorch_cos_sim(user_embedding, stored_embeddings)

        # Find the best match
        best_match_idx = cosine_scores.argmax()
        best_score = float(cosine_scores[0][best_match_idx])
        best_question = all_questions[best_match_idx]

        print(f"User Question: '{user_question}'")
        print(f"Best matched stored question: '{best_question}' with score: {best_score:.4f}")

        # Check if the match is good enough
        if best_score > SIMILARITY_THRESHOLD:
            # Return the note associated with the best-matched question
            answer = question_to_note_map[best_question]
        else:
            answer = "Sorry, I couldn't find a relevant note to answer your question."
        
        return jsonify({'answer': answer})

    except Exception as e:
        print(f"Error during question matching: {e}")
        return jsonify({'error': str(e)}), 500