Use past user messages, inferred questions as context to extract questions

- Keep inferred questions in logs
- Improve prompt to GPT to try use past questions as context
- Pass past user message and inferred questions as context to help GPT
  extract complete questions
- This should improve search results quality

- Example Expected Inferred Questions from User Message using History:
  1. "What is the name of Arun's daughter?"
    => "What is the name of Arun's daughter"
  2. "Where does she study?" =>
    => "Where does Arun's daughter study?" OR
    => "Where does Arun's daughter, Reena study?"

src/khoj/processor/conversation/gpt.py

@@ -78,40 +78,63 @@ Summarize the notes in second person perspective:"""
     return str(story).replace("\n\n", "")
 
 
-def extract_questions(text, model="text-davinci-003", api_key=None, temperature=0, max_tokens=100):
+def extract_questions(
+    text, model="text-davinci-003", conversation_log={}, api_key=None, temperature=0, max_tokens=100
+):
     """
     Infer search queries to retrieve relevant notes to answer user query
     """
     # Initialize Variables
     openai.api_key = api_key or os.getenv("OPENAI_API_KEY")
 
+    # Extract Past User Message and Inferred Questions from Conversation Log
+    chat_history = "".join(
+        [
+            f'Q: {chat["intent"]["query"]}\n\n{chat["intent"].get("inferred-queries") or list([chat["intent"]["query"]])}\n\n'
+            for chat in conversation_log.get("chat", [])[-4:]
+            if chat["by"] == "khoj"
+        ]
+    )
+
     # Get dates relative to today for prompt creation
     today = datetime.today()
     current_new_year = today.replace(month=1, day=1)
     last_new_year = current_new_year.replace(year=today.year - 1)
 
     prompt = f"""
-You are Khoj, a chat assistant with the ability to search the users notes
-What searches, if any, will you need to perform to answer the users question below? Provide search queries as a JSON list of  strings
+You are Khoj, a chat assistant with the ability to search the users notes and continue the existing conversation.
+What searches, if any, will you need to perform to answer the users question below?
+Provide search queries as a JSON list of strings
 Current Date: {today.strftime("%HH:%MM %A, %Y-%m-%d")}
 
 Q: How was my trip to Cambodia?
 
-["My Cambodia trip experience"]
+["How was my trip to Cambodia?"]
 
-Q: How are you feeling?
+Q: When did i go there?
 
-[]
+["When did I go to Cambodia?"]
 
 Q: What national parks did I go to last year?
 
 ["National park I visited in {last_new_year.strftime("%Y")} dt>=\\"{last_new_year.strftime("%Y-%m-%d")}\\" dt<\\"{current_new_year.strftime("%Y-%m-%d")}\\""]
 
+Q: How are you feeling today?
+
+[]
+
 Q: Is Bob older than Tom?
 
 ["When was Bob born?", "What is Tom's age?"]
 
-Q: {text}"""
+Q: What is their age difference?
+
+["What is Bob's age?", "What is Tom's age?"]
+
+{chat_history}
+Q: {text}
+
+"""
 
     # Get Response from GPT
     response = openai.Completion.create(prompt=prompt, model=model, temperature=temperature, max_tokens=max_tokens)

src/khoj/routers/api.py

@@ -204,15 +204,14 @@ def chat(q: Optional[str] = None):
         else:
             return {"status": "ok", "response": []}
 
-    # Extract search queries from user message
-    queries = extract_questions(q, model=model, api_key=api_key)
+    # Infer search queries from user message
+    inferred_queries = extract_questions(q, model=model, api_key=api_key, conversation_log=meta_log)
 
     # Collate search results as context for GPT
     result_list = []
-    for query in queries:
+    for query in inferred_queries:
         result_list.extend(search(query, n=2, r=True, score_threshold=0, dedupe=False))
     collated_result = "\n\n".join({f"# {item.additional['compiled']}" for item in result_list})
-    logger.debug(f"Reference Context:\n{collated_result}")
 
     try:
         gpt_response = converse(collated_result, q, meta_log, api_key=api_key)
@@ -224,7 +223,10 @@ def chat(q: Optional[str] = None):
     # Update Conversation History
     state.processor_config.conversation.chat_session = message_to_prompt(q, chat_session, gpt_message=gpt_response)
     state.processor_config.conversation.meta_log["chat"] = message_to_log(
-        q, gpt_response, khoj_message_metadata={"context": collated_result}, conversation_log=meta_log.get("chat", [])
+        q,
+        gpt_response,
+        khoj_message_metadata={"context": collated_result, "intent": {"inferred-queries": inferred_queries}},
+        conversation_log=meta_log.get("chat", []),
     )
 
     return {"status": status, "response": gpt_response, "context": collated_result}