[CAI-257] Chatbot/post queries with conversation (#1245)

* feat(chatbot): conversational chat

* fix: dict

* fix(chatbot): messages

apps/chatbot/config/prompts.yaml

@@ -60,9 +60,16 @@ refine_prompt_str: |
     Answer with either the original answer or with a refined answer to better answer the user query according to the `Chatbot Policy` listed above.
     Answer: 
 
+condense_prompt_str: |
+  Given a conversation (between Human and Assistant) and a follow up message from Human, rewrite the message to be a standalone question that captures all relevant context from the conversation.
+
+  The standalone question must be in Italian.
+
+  <Chat History>
+  {chat_history}
+
+  <Follow Up Message>
+  {question}
+
+  <Standalone question>
 
-# Reply to the user following these two steps:
-#     Step 1:
-#       Pay great attention in detail on the query's language and determine if it is formulated in Italian, English, Spanish, French, German, Greek, Croatian, or Slovenian ('yes' or 'no').
-#     Step 2:
-#       If Step 1 returns 'yes': reply according to the `Chatbot Policy` listed above. If `no`, reply you cannot speak that language and ask for a new query written in an accepted language.

apps/chatbot/src/app/main.py

@@ -6,9 +6,8 @@
 import uuid
 import boto3
 import datetime
-import time
 import jwt
-from typing import Annotated
+from typing import Annotated, List
 from boto3.dynamodb.conditions import Key
 from botocore.exceptions import BotoCoreError, ClientError
 from fastapi import FastAPI, HTTPException, Header
@@ -24,9 +23,15 @@
 chatbot = Chatbot(params, prompts)
 
 
+class QueryFromThePast(BaseModel):
+  id: str | None = None
+  question: str
+  answer: str | None = None
+
 class Query(BaseModel):
   question: str
   queriedAt: str | None = None
+  history: List[QueryFromThePast] | None = None
 
 class QueryFeedback(BaseModel):
   badAnswer: bool
@@ -75,7 +80,11 @@ async def query_creation (
   now = datetime.datetime.now(datetime.UTC)
   userId = current_user_id(authorization)
   session = find_or_create_session(userId, now=now)
-  answer = chatbot.generate(query.question)
+  answer = chatbot.chat_generate(
+    query_str = query.question,
+    messages = [item.dict() for item in query.history] if query.history else None
+  )
+
 
   if query.queriedAt is None:
     queriedAt = now.isoformat()

apps/chatbot/src/modules/chatbot.py

@@ -18,7 +18,7 @@
 from dotenv import load_dotenv
 
 load_dotenv()
-nest_asyncio.apply()
+#nest_asyncio.apply()
 
 
 USE_PRESIDIO = True if (os.getenv("CHB_USE_PRESIDIO", "True")).lower() == "true" else False
@@ -62,12 +62,13 @@ def __init__(
             chunk_sizes=params["vector_index"]["chunk_sizes"],
             chunk_overlap=params["vector_index"]["chunk_overlap"]
         )
-        self.qa_prompt_tmpl, self.ref_prompt_tmpl = self._get_prompt_templates()
+        self.qa_prompt_tmpl, self.ref_prompt_tmpl, self.condense_prompt_tmpl = self._get_prompt_templates()
         self.engine = get_automerging_engine(
             self.index,
             llm=self.model,
             text_qa_template=self.qa_prompt_tmpl,
             refine_template=self.ref_prompt_tmpl,
+            condense_template=self.condense_prompt_tmpl,
             verbose=self.params["engine"]["verbose"]
         )
 
@@ -83,16 +84,25 @@ def _get_prompt_templates(self) -> Tuple[PromptTemplate, PromptTemplate]:
             }
         )
 
-        ref_prompt_tmpl = PromptTemplate(
-            self.prompts["refine_prompt_str"],
-            prompt_type="refine",
-            template_var_mappings = {
-                "existing_answer": "existing_answer",
-                "query_str": "query_str"
+        ref_prompt_tmpl = None
+        # PromptTemplate(
+        #     self.prompts["refine_prompt_str"],
+        #     prompt_type="refine",
+        #     template_var_mappings = {
+        #         "existing_answer": "existing_answer",
+        #         "query_str": "query_str"
+        #     }
+        # )
+
+        condense_prompt_tmpl = PromptTemplate(
+            self.prompts["condense_prompt_str"],
+            template_var_mappings={
+                "chat_history": "chat_history",
+                "question": "question"
             }
         )
 
-        return qa_prompt_tmpl, ref_prompt_tmpl
+        return qa_prompt_tmpl, ref_prompt_tmpl, condense_prompt_tmpl
 
 
     def _get_response_str(self, engine_response: RESPONSE_TYPE) -> str:
@@ -169,11 +179,11 @@ def _messages_to_chathistory(self, messages: Optional[List[dict]] = None) -> Lis
                 chat_history += [
                     ChatMessage(
                         role = MessageRole.USER,
-                        content = message["query"]
+                        content = message["question"]
                     ),
                     ChatMessage(
                         role = MessageRole.ASSISTANT,
-                        content = message["response"]
+                        content = message["answer"]
                     )
                 ]
 

apps/chatbot/src/modules/engine.py

@@ -23,6 +23,7 @@ def get_automerging_engine(
         response_mode: str = "compact",
         text_qa_template: PromptTemplate | None = None,
         refine_template: PromptTemplate | None = None,
+        condense_template: PromptTemplate | None = None,
         verbose: bool = True,
         use_chat_engine: bool | None = None
     ) -> (RetrieverQueryEngine | CondenseQuestionChatEngine):
@@ -57,7 +58,8 @@ def get_automerging_engine(
 
     if use_chat_engine:
         automerging_engine = CondenseQuestionChatEngine.from_defaults(
-            query_engine = automerging_engine
+            query_engine = automerging_engine,
+            condense_question_prompt = condense_template
         )
 
     return automerging_engine

apps/chatbot/src/modules/test_chatbot.py

@@ -35,9 +35,9 @@ def test_messages_to_chathistory():
     ###########################3
 
     messages = [
-        {"query": "aaaa", "response": "bbbb"},
-        {"query": "cccc", "response": "dddd"},
-        {"query": "eeee", "response": "ffff"},
+        {"question": "aaaa", "answer": "bbbb"},
+        {"question": "cccc", "answer": "dddd"},
+        {"question": "eeee", "answer": "ffff"},
     ]
     chat_history = CHATBOT._messages_to_chathistory(messages)
 

apps/chatbot/src/modules/vector_database.py

@@ -3,10 +3,10 @@
 import time
 import json
 import tqdm
-import logging
 import hashlib
 import html2text
 import pytz
+from logging import getLogger
 from datetime import datetime
 from bs4 import BeautifulSoup
 from selenium import webdriver
@@ -35,7 +35,10 @@
 
 from dotenv import load_dotenv
 
+
 load_dotenv()
+logger = getLogger(__name__)
+
 
 PROVIDER = os.getenv("CHB_PROVIDER")
 assert PROVIDER in ["google", "aws"]
@@ -154,7 +157,7 @@ def create_documentation(
     driver_options.add_argument('--disable-gpu')
     driver_options.add_argument('--no-sandbox')
     driver_options.add_argument('--disable-dev-shm-usage')
-    
+
     for file in tqdm.tqdm(html_files, total=len(html_files), desc="Extracting HTML"):
 
         if file in dynamic_htmls or "/webinars/" in file or "/api/" in file:
@@ -198,8 +201,8 @@ def create_documentation(
                 }
             ))
 
-    logging.info(f"[vector_database.py - create_documentation] Number of documents with content: {len(documents)}")
-    logging.info(f"[vector_database.py - create_documentation] Number of empty pages in the documentation: {len(empty_pages)}. These are left out.")
+    logger.info(f"Number of documents with content: {len(documents)}")
+    logger.info(f"Number of empty pages in the documentation: {len(empty_pages)}. These are left out.")
     with open("empty_htmls.json", "w") as f:
         json.dump(empty_pages, f, indent=4)
 
@@ -214,7 +217,7 @@ def build_automerging_index_redis(
         chunk_overlap: int
     ) -> VectorStoreIndex:
 
-    logging.info("[vector_database.py - build_automerging_index_redis] Storing vector index and hash table on Redis..")
+    logger.info("Storing vector index and hash table on Redis..")
 
     Settings.llm = llm
     Settings.embed_model = embed_model
@@ -230,9 +233,9 @@ def build_automerging_index_redis(
             key=key,
             val=value
         )
-    logging.info(f"[vector_database.py - build_automerging_index_redis] hash_table_{NEW_INDEX_ID} is now on Redis.")
+    logger.info(f"[vector_database.py - build_automerging_index_redis] hash_table_{NEW_INDEX_ID} is now on Redis.")
 
-    logging.info(f"[vector_database.py - build_automerging_index_redis] Creating index {NEW_INDEX_ID} ...")
+    logger.info(f"Creating index {NEW_INDEX_ID} ...")
     nodes = Settings.node_parser.get_nodes_from_documents(documents)
     leaf_nodes = get_leaf_nodes(nodes)
 
@@ -267,7 +270,7 @@ def build_automerging_index_redis(
     )
     automerging_index.set_index_id(NEW_INDEX_ID)
     put_ssm_parameter(os.getenv("CHB_LLAMAINDEX_INDEX_ID"), NEW_INDEX_ID)
-    logging.info("[vector_database.py - build_automerging_index_redis] Created vector index successfully and stored on Redis.")
+    logger.info("Created vector index successfully and stored on Redis.")
 
     delete_old_index()
 
@@ -296,7 +299,7 @@ def load_automerging_index_redis(
             schema=REDIS_SCHEMA
         )
 
-        logging.info("[vector_database.py - load_automerging_index_redis] Loading vector index from Redis...")
+        logger.info("Loading vector index from Redis...")
         storage_context = StorageContext.from_defaults(
             vector_store=redis_vector_store,
             docstore=REDIS_DOCSTORE,
@@ -310,7 +313,7 @@ def load_automerging_index_redis(
 
         return automerging_index
     else:
-        logging.error("[vector_database.py - load_automerging_index_redis] No index_id provided.")
+        logger.error("No index_id provided.")
 
 
 def delete_old_index():
@@ -320,4 +323,4 @@ def delete_old_index():
             if f"{INDEX_ID}/vector" in str(key) or f"hash_table_{INDEX_ID}" == str(key):
                 REDIS_CLIENT.delete(key)
 
-        logging.info(f"[vector_database.py - delete_old_index] Deleted index with ID: {INDEX_ID} and its hash table from Redis.")
+        logger.info(f"Deleted index with ID: {INDEX_ID} and its hash table from Redis.")