Update app.py

app.py

@@ -29,7 +29,7 @@ PINECONE_API=os.getenv("PINECONE_API_KEY")
 pc = pinecone.Pinecone(
     api_key=PINECONE_API
 )
-index_name = "iocl"
+index_name = "iocl2"
 index = pc.Index(index_name)
 # MongoDB connection setup
 MONGO_URI = os.getenv("MONGO_URI")
@@ -142,21 +142,21 @@ def regenerate_response():
 
 
 
-# When generating a response, pass only the latest 15 messages to the LLM
 def generate_summary(chunks, query, chat_history):
     try:
-        # Limit the history sent to the LLM to the latest 15 question-answer pairs
-        limited_history = chat_history[-5:] if len(chat_history) > 5 else chat_history
+        # Limit the history sent to the LLM to the latest 3 question-answer pairs
+        limited_history = chat_history[-3:] if len(chat_history) > 3 else chat_history
 
         # Create conversation history for the LLM, only using the last 15 entries
         history_text = "\n".join([f"User: {q['improved_question']}\nLLM: {q['answer']}" for q in limited_history])
 
         # Define the system and user prompts including the limited history
         prompt = ChatPromptTemplate.from_messages([
-            ("system", """You are a chatbot specializing in answering queries related to Indian Oil Corporation Limited (IOCL). You will be provided with chunks of data from the IOCL website to answer user queries. Each chunk will include associated URLs, which you must reference in your responses to show the source of the information.
+            ("system", """You are a chatbot specializing in answering queries related to Indian Oil Corporation Limited (IOCL). You will be provided with chunks of data from the IOCL website to answer user queries. Each chunk will include associated URLs, You must give the url of the chunks which you are using to answer the query.
                             Key Guidelines:
-                        
-                            1.If the user query is not clear, or you think multiple answers are possbile, you can ask for clarification with proper reasoning. Do not mention chunk name in any of your replies.
+
+                            1.If the user query is not clear, or you think multiple answers are possbile, you should ask for clarification with proper reasoning.
+                            2.Do not mention chunk name in any of your replies.
                             2.Detailed and Clear: Provide thorough, clear, and concise responses without omitting relevant information from the data chunks.
                             3.Natural Summarization: When answering, you must not  directly quote chunk names,formats. Instead, summarize or interpret the data naturally and conversationally.
                             4.Use Conversation History: Refer back to the conversation history to maintain consistency and build on prior responses, if applicable.
@@ -164,22 +164,24 @@ def generate_summary(chunks, query, chat_history):
                             6.Graceful Handling of General Queries: If a user sends greetings, introduction, salutations, or unrelated questions, respond appropriately and conversationally.
                             7.Include Source URLs: Always include the URLs from the relevant chunks of data that you're using to answer the query.
                             8.Thoroughly looks for answer to the query in the provided chunks before replying, if you feel the query is irrelevant or answer is not present then you can ask user to clarify or tell that it cannot be answered.
-                            
+                            9.Sometimes chunks might contain very less data still use it if its relevant.
 
-                                          """),
+                            """),
 
             ("human", f'''
                             "Query":\n {query}\n
+                            Below are the pinecone chunks that should be used to answer the user query:
                             "Extracted Data": \n{chunks}\n
+                            Below is the previous conversation history:
                             "Previous Conversation History": \n{history_text}\n
-                            
+
                             '''
              )
         ])
 
         # Chain the prompt with LLM for response generation
         chain = prompt | llm
-        result = chain.invoke({"Query": query,"Extracted Data": chunks,"Previous Conversation History":history_text})
+        result = chain.invoke({"Query": query, "Extracted Data": chunks, "Previous Conversation History": history_text})
 
         # Return the generated response
         logging.info(f"LLM answer is :{result}")
@@ -189,13 +191,12 @@ def generate_summary(chunks, query, chat_history):
         st.error(f"Error answering your question: {e}")
         return None
 
-
-def get_context_from_messages(query,chat_history):
+def get_context_from_messages(query, chat_history):
     try:
 
         logging.info(f"Getting context from original query: {query}")
 
-        # Limit the history sent to the LLM to the latest 15 question-answer pairs
+        # Limit the history sent to the LLM to the latest 3 question-answer pairs
         limited_history = chat_history[-3:] if len(chat_history) > 3 else chat_history
 
         # Create conversation history for the LLM, only using the last 15 entries
@@ -203,12 +204,11 @@ def get_context_from_messages(query,chat_history):
 
         # Define the system and user prompts including the limited history
         prompt = ChatPromptTemplate.from_messages([
-            ("system", """"I will provide you with a user query and up to the last 3 messages from the chat history, including both questions and answers.Your task is to reformulate the user query nicely which makes complete sense and is completely self contained.
+            ("system", """"I will provide you with a user query and up to the last 3 messages from the chat history which includes both questions and answers.Your task is to understand  the user query nicely and restructure it if required such that  it  makes complete sense and is completely self contained.
                            The provided queries are related to Indian Oil Corporation limited (IOCL).
                            1. If the query is a follow-up, use the provided chat history to reconstruct a well-defined, contextually complete query that can stand alone."
                            2. if the query is self contained, if applicable try to improve it to make is coherent.
                            3. if the user query is salutations, greetings or not relevant in that case give the query back as it is.
-                           4. Even if the user query is just a one word query, you must formulate a well defined query.
                            I have provided an output format below, stricly follow it. Do not give anything else other than just the output.
                            expected_output_format: "query: String or None"
                            """),
@@ -221,7 +221,7 @@ def get_context_from_messages(query,chat_history):
 
         # Chain the prompt with LLM for response generation
         chain = prompt | llm
-        result = chain.invoke({"Query": query,"Previous Conversation History":history_text})
+        result = chain.invoke({"Query": query, "Previous Conversation History": history_text})
         logging.info(f"llm answer for query extraction is :{result}")
 
         # Return the generated response
@@ -231,6 +231,7 @@ def get_context_from_messages(query,chat_history):
         logging.error(f"exception occured in getting query from original query :{e}")
         return None
 
+
 def get_query_from_llm_answer(llm_output):
     match = re.search(r'query:\s*(.*)', llm_output)
     if match:
@@ -292,7 +293,7 @@ if user_question:
                 query=user_question
 
             query_embedding=embeddings.embed_query(query)
-            search_results = index.query(vector=query_embedding, top_k=3, include_metadata=True)
+            search_results = index.query(vector=query_embedding, top_k=15, include_metadata=True)
             matches=search_results['matches']
 
             content=""
@@ -306,7 +307,7 @@ if user_question:
             reply = generate_summary(content, query, st.session_state['chat_history'])
 
             if reply:
-                # Append the new question-answer pair t     o chat history
+                # Append the new question-answer pair to chat history
                 st.session_state['chat_history'].append({"question": user_question, "answer": reply,"improved_question":query})
 
                 # Update the current chat session in MongoDB