Update app.py

app.py

@@ -36,18 +36,72 @@ def preprocess_text(text):
     # Print the length of cleaned_chunks
   len(cleaned_chunks)
 
+# step 4 from sematic search(embed chuncks)
+#Load the pre-trained embedding model that converts text to vectors
+model = SentenceTransformer('all-MiniLM-L6-v2')
+
+def create_embeddings(text_chunks):
+  # Convert each text chunk into a vector embedding and store as a tensor
+  chunk_embeddings = model.encode(cleaned_chunks, convert_to_tensor=True) # Replace ... with the text_chunks list
+
+  # Print the chunk embeddings
+  print(chunk_embeddings)
+
+  # Print the shape of chunk_embeddings
+  print(chunk_embeddings.shape)
+
+  # Return the chunk_embeddings
+  return chunk_embeddings
+
+# Call the create_embeddings function and store the result in a new chunk_embeddings variable
+chunk_embeddings = create_embeddings(cleaned_chunks) # Complete this line
+    
     # Return the cleaned_chunks
   return cleaned_chunks
 
 # Call the preprocess_text function and store the result in a cleaned_chunks variable
 cleaned_chunks =  preprocess_text(safe_restaurant)
 
+# step 5 from sematic searc (find and print top chunk)
+# Define a function to find the most relevant text chunks for a given query, chunk_embeddings, and text_chunks
+def get_top_chunks(query, chunk_embeddings, text_chunks):
+  # Convert the query text into a vector embedding
+  query_embedding = model.encode(query, convert_to_tensor = True) # Complete this line
+
+  # Normalize the query embedding to unit length for accurate similarity comparison
+  query_embedding_normalized = query_embedding / query_embedding.norm()
+
+  # Normalize all chunk embeddings to unit length for consistent comparison
+  chunk_embeddings_normalized = chunk_embeddings / chunk_embeddings.norm(dim=1, keepdim=True)
+
+  # Calculate cosine similarity between query and all chunks using matrix multiplication
+  similarities = torch.matmul(chunk_embeddings_normalized, query_embedding_normalized) # Complete this line
+
+  # Print the similarities
+  print(similarities)
+  # Find the indices of the 3 chunks with highest similarity scores
+  top_indices = torch.topk(similarities, k=3).indices
+
+  # Print the top indices
+  print(top_indices)
+  # Create an empty list to store the most relevant chunks
+  top_chunks = []
+
+  # Loop through the top indices and retrieve the corresponding text chunks
+  for x in top_indices:
+    chunks = text_chunks[x]
+    top_chunks.append(chunks)
+  # Return the list of most relevant chunks
+  return top_chunks
+    
 def respond(message, history):
     #responses = ["Yes", "No"]
     #return random.choice(responses)
 
-    messages= [{"role": "system", "content": "Nice chat bot."}]
+    messages= [{"role": "system", "content": "You are a friendly chat bot that help people find restaurants that fit their needs/dietary restriction and location range!"}]
 
+    best_restaurant_chunks = get_top_chunks(message, chunk_embeddings, cleaned_chunks)
+    print(best_restaurant_chunks)
 
     if history:
         messages.extend(history)