Update app.py

app.py

@@ -1,37 +1,37 @@
+# app.py (Updated for a model with pre-normalized embeddings)
+
 import sys
 import subprocess
 from flask import Flask, render_template, request, flash, redirect, url_for, jsonify
 import torch
-import torch.nn.functional as F # Import the functional module
+import torch.nn.functional as F
 from transformers import AutoTokenizer, AutoModel
 import os
 import chromadb
 from huggingface_hub import snapshot_download
 
-# (App setup and load_resources function are unchanged)
 app = Flask(__name__)
 app.secret_key = os.urandom(24)
 
 CHROMA_PATH = "chroma_db"
 COLLECTION_NAME = "bible_verses"
-# *** CHANGE 1: UPDATE THE MODEL NAME ***
-MODEL_NAME = "sentence-transformers/all-mpnet-base-v2"
-# *** CHANGE 2: UPDATE THE DATASET REPO NAME ***
-DATASET_REPO = "broadfield-dev/bible-chromadb-mpnet"
+# *** CHANGE 1: USE A MODEL WITH NORMALIZED EMBEDDINGS ***
+MODEL_NAME = "sentence-transformers/multi-qa-mpnet-base-dot-v1"
+# *** CHANGE 2: USE THE NEW REPO FOR THE NEW DATABASE ***
+DATASET_REPO = "broadfield-dev/bible-chromadb-multi-qa-mpnet"
 STATUS_FILE = "build_status.log"
 
 chroma_collection = None
 tokenizer = None
 embedding_model = None
 
-# Mean Pooling Function - Take attention mask into account for correct averaging
+# Mean Pooling Function - Crucial for sentence-transformer models
 def mean_pooling(model_output, attention_mask):
-    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
+    token_embeddings = model_output[0]
     input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
     return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
 
 def load_resources():
-    # (This function is unchanged)
     global chroma_collection, tokenizer, embedding_model
     if chroma_collection and embedding_model: return True
     print("Attempting to load resources...")
@@ -95,13 +95,14 @@ def search():
     if not user_query:
         return render_template('index.html', results=[])
 
-    # *** CHANGE 3: USE THE CORRECT POOLING STRATEGY FOR SBERT MODELS ***
     encoded_input = tokenizer([user_query], padding=True, truncation=True, return_tensors='pt')
     with torch.no_grad():
         model_output = embedding_model(**encoded_input)
     
-    sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
-    query_embedding = F.normalize(sentence_embeddings, p=2, dim=1)
+    query_embedding = mean_pooling(model_output, encoded_input['attention_mask'])
+
+    # *** REMOVED: NO LONGER NEED TO NORMALIZE THE QUERY EMBEDDING ***
+    # query_embedding = F.normalize(query_embedding, p=2, dim=1)
 
     search_results = chroma_collection.query(
         query_embeddings=query_embedding.cpu().tolist(),