user name + facts during chat + QA prompt interactions

backend/database/auth.py

@@ -24,14 +24,13 @@ def get_user_from_uid(uid: str):
 
 
 def get_user_name(uid: str):
-    # TODO: "The User" or "User"?
-    if cached_name := get_cached_user_name(uid):
-        return cached_name
+    # if cached_name := get_cached_user_name(uid):
+    #     return cached_name
 
     user = get_user_from_uid(uid)
-    display_name = user.get('display_name', 'User').split(' ')[0] if user else 'User'
+    display_name = user.get('display_name', 'User').split(' ')[0] if user else 'The User'
     if display_name == 'AnonymousUser':
-        display_name = 'User'
+        display_name = 'The User'
 
     cache_user_name(uid, display_name)
     return display_name

backend/database/facts.py

@@ -1,15 +1,16 @@
 from typing import List
 
 from google.cloud import firestore
+from google.cloud.firestore_v1 import FieldFilter
 
 from ._client import db
 
 
 def get_facts(uid: str, limit: int = 100, offset: int = 0):
-    facts_ref = (
-        db.collection('users').document(uid).collection('facts')
-    )
-    facts_ref = facts_ref.order_by('created_at', direction=firestore.Query.DESCENDING)
+    # TODO: cache this
+    facts_ref = db.collection('users').document(uid).collection('facts')
+    facts_ref = facts_ref.order_by('created_at', direction=firestore.Query.DESCENDING).where(
+        filter=FieldFilter('deleted', '==', False))
     facts_ref = facts_ref.limit(limit).offset(offset)
     return [doc.to_dict() for doc in facts_ref.stream()]
 
@@ -31,3 +32,20 @@ def delete_facts(uid: str):
     for doc in facts_ref.stream():
         batch.delete(doc.reference)
     batch.commit()
+
+
+def delete_facts_for_memory(uid: str, memory_id: str):
+    batch = db.batch()
+    user_ref = db.collection('users').document(uid)
+    facts_ref = user_ref.collection('facts')
+    query = (
+        facts_ref.where(filter=FieldFilter('memory_id', '==', memory_id))
+        .where(filter=FieldFilter('deleted', '==', False))
+    )
+
+    removed_ids = []
+    for doc in query.stream():
+        batch.update(doc.reference, {'deleted': True})
+        removed_ids.append(doc.id)
+    batch.commit()
+    print('delete_facts_for_memory', memory_id, len(removed_ids))

backend/models/facts.py

@@ -4,6 +4,7 @@
 
 from pydantic import BaseModel, Field
 
+from database._client import document_id_from_seed
 from models.memory import CategoryEnum
 
 
@@ -21,6 +22,11 @@ class Fact(BaseModel):
     content: str = Field(description="The content of the fact")
     category: FactCategory = Field(description="The category of the fact", default=FactCategory.other)
 
+    @staticmethod
+    def get_facts_as_str(facts):
+        existing_facts = [f"{f.content} ({f.category.value})" for f in facts]
+        return '' if not existing_facts else '\n- ' + '\n- '.join(existing_facts)
+
 
 class FactDB(Fact):
     id: str
@@ -37,3 +43,16 @@ class FactDB(Fact):
     manually_added: bool = False
     edited: bool = False
     deleted: bool = False
+
+    @staticmethod
+    def from_fact(fact: Fact, uid: str, memory_id: str, memory_category: CategoryEnum) -> 'FactDB':
+        return FactDB(
+            id=document_id_from_seed(fact.content),
+            uid=uid,
+            content=fact.content,
+            category=fact.category,
+            created_at=datetime.utcnow(),
+            updated_at=datetime.utcnow(),
+            memory_id=memory_id,
+            memory_category=memory_category,
+        )

backend/routers/chat.py

@@ -5,9 +5,12 @@
 from fastapi import APIRouter, Depends
 
 import database.chat as chat_db
+from database.auth import get_user_name
+from database.facts import get_facts
 from models.chat import Message, SendMessageRequest, MessageSender
-from utils.other import endpoints as auth
+from models.facts import Fact
 from utils.llm import qa_rag, initial_chat_message
+from utils.other import endpoints as auth
 from utils.plugins import get_plugin_by_id
 from utils.retrieval.rag import retrieve_rag_context
 
@@ -39,7 +42,9 @@ def send_message(
     messages = filter_messages(messages, plugin_id)
 
     context_str, memories = retrieve_rag_context(uid, messages)
-    response: str = qa_rag(context_str, messages, plugin)
+    user_name = get_user_name(uid)
+    user_facts = [Fact(**fact) for fact in get_facts(uid)]
+    response: str = qa_rag(user_name, user_facts, context_str, messages, plugin)
 
     ai_message = Message(
         id=str(uuid.uuid4()),
@@ -59,7 +64,10 @@ def send_message(
 def initial_message_util(uid: str, plugin_id: Optional[str] = None):
     plugin = get_plugin_by_id(plugin_id)
 
-    text = initial_chat_message(plugin)
+    user_name = get_user_name(uid)
+    user_facts = [Fact(**fact) for fact in get_facts(uid)]
+
+    text = initial_chat_message(user_name, user_facts, plugin)
     ai_message = Message(
         id=str(uuid.uuid4()),
         text=text,

backend/scripts/rag/app.py

@@ -23,6 +23,9 @@
 from models.transcript_segment import TranscriptSegment
 from utils.llm import qa_rag
 from utils.retrieval.rag import retrieve_rag_context
+from database.auth import get_user_name
+from database.facts import get_facts
+from models.facts import Fact
 
 # File to store the state
 STATE_FILE = 'chat_state.json'
@@ -120,7 +123,9 @@ def send_message(text: str):
     else:
         context_str, memories, topics, dates_range = data
 
-    response: str = qa_rag(context_str, get_messages(), None)
+    user_name = get_user_name(uid)
+    user_facts = [Fact(**fact) for fact in get_facts(uid)]
+    response: str = qa_rag(user_name, user_facts, context_str, get_messages(), None)
 
     # Generate visualization
     ai_message_id = str(uuid.uuid4())

backend/scripts/rag/facts.py

@@ -1,5 +1,4 @@
 import threading
-from datetime import datetime
 from typing import Tuple
 
 import firebase_admin
@@ -13,48 +12,35 @@
 from utils.llm import new_facts_extractor
 
 
-def get_preferences_from_memory(memories: List[dict], uid: str) -> List[Tuple[str, List[Fact]]]:
+def get_facts_from_memory(memories: List[dict], uid: str) -> List[Tuple[str, List[Fact]]]:
     all_facts: List[Tuple[str, List[Fact]]] = []
     only_facts: List[Fact] = []
     user_name = get_user_name(uid)
     print('User:', user_name)
     for i, memory in enumerate(memories):
         data = Memory(**memory)
-        try:
-            new_facts = new_facts_extractor(data.transcript_segments, user_name, only_facts)
-        except Exception as e:
-            # LLM failed to parse output, we can skip 1 or 2, every 200.
+        new_facts = new_facts_extractor(data.transcript_segments, user_name, only_facts)
+        if not new_facts:
             continue
         all_facts.append([memory['id'], new_facts])
         only_facts.extend(new_facts)
 
         print(uid, 'Memory #', i + 1, 'retrieved', len(new_facts), 'facts')
 
-    # for fact in only_facts:
-    #     print(fact.category.value.upper(), '~', fact.content)
     return all_facts
 
 
 def execute_for_user(uid: str):
     facts_db.delete_facts(uid)
 
     memories = memories_db.get_memories(uid, limit=2000)
-    data: List[Tuple[str, List[Fact]]] = get_preferences_from_memory(memories, uid)
+    data: List[Tuple[str, List[Fact]]] = get_facts_from_memory(memories, uid)
     parsed_facts = []
     for item in data:
         memory_id, facts = item
         memory = next((m for m in memories if m['id'] == memory_id), None)
         for fact in facts:
-            parsed_facts.append(FactDB(
-                id=document_id_from_seed(fact.content),
-                uid=uid,
-                content=fact.content,
-                category=fact.category,
-                created_at=datetime.utcnow(),
-                updated_at=datetime.utcnow(),
-                memory_id=memory_id,
-                memory_category=memory['structured']['category'],
-            ))
+            parsed_facts.append(FactDB.from_fact(fact, uid, memory['id'], memory['structured']['category']))
         facts_db.save_facts(uid, [fact.dict() for fact in parsed_facts])
 
 

backend/utils/llm.py

@@ -1,10 +1,9 @@
-import tiktoken
 import json
 from datetime import datetime
-from typing import List, Tuple, Optional
+from typing import List, Optional
 
 from langchain_core.output_parsers import PydanticOutputParser
-from langchain_core.prompts import ChatPromptTemplate, PromptTemplate
+from langchain_core.prompts import ChatPromptTemplate
 from langchain_openai import ChatOpenAI, OpenAIEmbeddings
 from pydantic import BaseModel, Field
 
@@ -237,13 +236,15 @@ def generate_embedding(content: str) -> List[float]:
 # ****************************************
 # ************* CHAT BASICS **************
 # ****************************************
-def initial_chat_message(plugin: Optional[Plugin] = None) -> str:
+def initial_chat_message(user_name: str, user_facts: List[Fact], plugin: Optional[Plugin] = None) -> str:
     if plugin is None:
-        prompt = '''
+        prompt = f'''
         You are an AI with the following characteristics:
         Name: Friend, 
         Personality/Description: A friendly and helpful AI assistant that aims to make your life easier and more enjoyable.
         Task: Provide assistance, answer questions, and engage in meaningful conversations.
+        
+        You are made for {user_name}, you already know the following facts about {user_name}: {Fact.get_facts_as_str(user_facts)}.
 
         Send an initial message to start the conversation, make sure this message reflects your personality, \
         humor, and characteristics.
@@ -256,6 +257,8 @@ def initial_chat_message(plugin: Optional[Plugin] = None) -> str:
         Name: {plugin.name}, 
         Personality/Description: {plugin.chat_prompt},
         Task: {plugin.memory_prompt}
+        
+        You are made for {user_name}, you already know the following facts about {user_name}: {Fact.get_facts_as_str(user_facts)}.
 
         Send an initial message to start the conversation, make sure this message reflects your personality, \
         humor, and characteristics.
@@ -401,8 +404,6 @@ def new_facts_extractor(
     # TODO: later, focus a lot on user said things, rn is hard because of speech profile accuracy
     # TODO: include negative facts too? Things the user doesn't like?
 
-    existing_facts = [f"{f.content} ({f.category.value})" for f in existing_facts]
-    facts = '' if not existing_facts else '\n- ' + '\n- '.join(existing_facts)
     prompt = f'''
     You are an experienced detective, whose job is to create detailed profile personas based on conversations.
     
@@ -418,7 +419,7 @@ def new_facts_extractor(
     This way we can create a more accurate profile. 
     Include from 0 up to 3 valuable facts, If you don't find any new facts, or ones worth storing, output an empty list of facts. 
 
-    Existing Facts: {facts}
+    Existing Facts: {Fact.get_facts_as_str(existing_facts)}
 
     Conversation:
     ```
@@ -427,12 +428,18 @@ def new_facts_extractor(
     '''.replace('    ', '').strip()
     # print(prompt)
 
-    with_parser = llm.with_structured_output(UserFacts)
-    response: UserFacts = with_parser.invoke(prompt)
-    return response.facts
+    try:
+        with_parser = llm.with_structured_output(UserFacts)
+        response: UserFacts = with_parser.invoke(prompt)
+        return response.facts
+    except Exception as e:
+        print(f'Error extracting new facts: {e}')
+        return []
 
 
-def qa_rag(context: str, messages: List[Message], plugin: Optional[Plugin] = None) -> str:
+def qa_rag(
+        user_name: str, user_facts: List[Fact], context: str, messages: List[Message], plugin: Optional[Plugin] = None
+) -> str:
     conversation_history = Message.get_messages_as_string(
         messages, use_user_name_if_available=True, use_plugin_name_if_available=True
     )
@@ -442,13 +449,16 @@ def qa_rag(context: str, messages: List[Message], plugin: Optional[Plugin] = Non
         plugin_info = f"Your name is: {plugin.name}, and your personality/description is '{plugin.description}'.\nMake sure to reflect your personality in your response.\n"
 
     prompt = f"""
-    You are an assistant for question-answering tasks. Use the following pieces of retrieved context and the conversation history to continue the conversation.
-    If you don't know the answer, just say that you didn't find any related information or you that don't know. Use three sentences maximum and keep the answer concise.
+    You are an assistant for question-answering tasks. 
+    You are made for {user_name}, you already know the following facts about {user_name}: {Fact.get_facts_as_str(user_facts)}.
+    
+    Use what you know about {user_name}, the following pieces of retrieved context and the chat history to continue the chat.
+    If you don't know the answer, just say that there's no available information about it. Use three sentences maximum and keep the answer concise.
     If the message doesn't require context, it will be empty, so follow-up the conversation casually.
     If there's not enough information to provide a valuable answer, ask the user for clarification questions.
     {plugin_info}
     
-    Conversation History:
+    Chat History:
     {conversation_history}
 
     Context:
@@ -461,11 +471,13 @@ def qa_rag(context: str, messages: List[Message], plugin: Optional[Plugin] = Non
     return llm.invoke(prompt).content
 
 
-def qa_emotional_rag(context: str, memories: List[Memory], emotion: str) -> str:
+def qa_emotional_rag(user_name: str, user_facts: List[Fact], context: str, memories: List[Memory], emotion: str) -> str:
     conversation_history = Memory.memories_to_string(memories)
 
     prompt = f"""
-    You are a thoughtful friend. Use the following pieces of retrieved context, the conversation history and user's emotions to share your thoughts and give the user positive advice.
+    You are a thoughtful friend of {user_name}, you already know the following facts about {user_name}: {Fact.get_facts_as_str(user_facts)}.
+    
+    Use the following pieces of retrieved context, the conversation history and user's emotions to share your thoughts and give the user positive advice.
     Thoughts and positive advice should be like a chat message. Keep it short.
     User's emotions:
     {emotion}