improved prompt and reduced chunks being fetched and chat history being passed

app.py

@@ -146,38 +146,37 @@ def regenerate_response():
 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[-10:] if len(chat_history) > 10 else chat_history
+        limited_history = chat_history[-5:] if len(chat_history) > 5 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 who specialises in answering user queries related to Indan Oil Corporation Limitied(IOCL).
-                            this is the extracted chunks of data from the Indian Oil Corporation Limited (IOCL) website. You will be provided with a query, and you must use these chunks of data to answer it comprehensively.
-                            Each chunk of data has website source page urls associated to it, you must include these relevant urls in your answer telling the source.
-                            Additionally, the conversation history is also provided which may contain relevant context or prior queries and responses. Use this history to ensure your answer is accurate and coherent, building on previous information if necessary.
-
+            ("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.
                             Key Guidelines:
-                            1.Accuracy is paramount: If the extracted data or conversation history does not contain the information required to answer the query, clearly state, "The answer is not available in the context." Do not attempt to provide a speculative or incorrect response.
-                            2.Be detailed: Provide clear, concise, and thorough answers without omitting any relevant information from the extracted data.
-                            3.Avoid quoting field names: When responding, avoid directly quoting or referencing field names or formats from the extracted data. Instead, present the information naturally, as if summarizing or interpreting the data. Try to give the answer in points.
-                            4.Use the conversation history: When applicable, refer to earlier parts of the conversation to ensure consistency and accuracy in your response.
-                            6.Answer the queries in conversational style.
-                            7.It is must to include urls present in the chunk which you are using to formulate your answer.
-                            8. Its possible that sometimes a user query will be greetings, your introduction, salutations, exclamations or completely unrelated to iocl handle them gracefully.
+                            1.Accurate Responses: If the answer to a query isn't found in the provided chunks or conversation history, respond with: "The answer is not available in the context." Avoid speculation or providing incorrect answers.
+                            2.Detailed and Clear: Provide thorough, clear, and concise responses without omitting relevant information from the data chunks.
+                            3.Natural Summarization: When answering, avoid directly quoting field names or 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.
+                            5.Ignore Unanswered Queries: If the conversation history contains previous responses like "The answer is not available in the context," disregard them when formulating your current response.
+                            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.
+
                                           """),
+
             ("human", f'''
-                            Previous Conversation History: \n{history_text}\n
-                            "Extracted Data": \n{chunks}\n
                             "Query":\n {query}\n
+                            "Extracted Data": \n{chunks}\n
+                            "Previous Conversation History": \n{history_text}\n
+                            
                             '''
              )
         ])
 
         # Chain the prompt with LLM for response generation
         chain = prompt | llm
-        result = chain.invoke({"Previous Concversation History":history_text,"Extracted Data": chunks, "Query": query})
+        result = chain.invoke({"Query": query,"Extracted Data": chunks,"Previous Conversation History":history_text})
 
         # Return the generated response
         logging.info(f"LLM answer is :{result}")
@@ -210,7 +209,7 @@ def get_context_from_messages(query,chat_history):
                            """),
             ("human", f'''  
                             "Query":\n {query}\n
-                            Previous Conversation History: \n{history_text}\n
+                            "Previous Conversation History": \n{history_text}\n
                             '''
              )
         ])
@@ -288,7 +287,7 @@ if user_question:
                 query=user_question
 
             query_embedding=embeddings.embed_query(query)
-            search_results = index.query(vector=query_embedding, top_k=5, include_metadata=True)
+            search_results = index.query(vector=query_embedding, top_k=3, include_metadata=True)
             matches=search_results['matches']
 
             content=""
@@ -298,7 +297,7 @@ if user_question:
                 content += f"chunk{i}: {chunk}\n" + f"url{i}: {url}\n"
 
 
-
+            print(f"content being passed is {content}")
             reply = generate_summary(content, query, st.session_state['chat_history'])
 
             if reply: