quivr/backend/llm/qa.py

import os
from typing import Any, List

from langchain.chains import ConversationalRetrievalChain
from langchain.chat_models import ChatOpenAI
from langchain.chat_models.anthropic import ChatAnthropic
from langchain.docstore.document import Document
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.memory import ConversationBufferMemory
from langchain.vectorstores import SupabaseVectorStore
from llm import LANGUAGE_PROMPT
from supabase import Client, create_client
from utils import ChatMessage


class CustomSupabaseVectorStore(SupabaseVectorStore):
    '''A custom vector store that uses the match_vectors table instead of the vectors table.'''
    user_id: str
    def __init__(self, client: Client, embedding: OpenAIEmbeddings, table_name: str, user_id: str = "none"):
        super().__init__(client, embedding, table_name)
        self.user_id = user_id
    
    def similarity_search(
        self, 
        query: str, 
        user_id: str = "none",
        table: str = "match_vectors", 
        k: int = 4, 
        threshold: float = 0.5, 
        **kwargs: Any
    ) -> List[Document]:
        vectors = self._embedding.embed_documents([query])
        query_embedding = vectors[0]
        res = self._client.rpc(
            table,
            {
                "query_embedding": query_embedding,
                "match_count": k,
                "p_user_id": self.user_id,
            },
        ).execute()

        match_result = [
            (
                Document(
                    metadata=search.get("metadata", {}),  # type: ignore
                    page_content=search.get("content", ""),
                ),
                search.get("similarity", 0.0),
            )
            for search in res.data
            if search.get("content")
        ]

        documents = [doc for doc, _ in match_result]

        return documents

def get_environment_variables():
    '''Get the environment variables.'''
    openai_api_key = os.getenv("OPENAI_API_KEY")
    anthropic_api_key = os.getenv("ANTHROPIC_API_KEY")
    supabase_url = os.getenv("SUPABASE_URL")
    supabase_key = os.getenv("SUPABASE_SERVICE_KEY")
    
    return openai_api_key, anthropic_api_key, supabase_url, supabase_key

def create_clients_and_embeddings(openai_api_key, supabase_url, supabase_key):
    '''Create the clients and embeddings.'''
    embeddings = OpenAIEmbeddings(openai_api_key=openai_api_key)
    supabase_client = create_client(supabase_url, supabase_key)
    
    return supabase_client, embeddings

def get_qa_llm(chat_message: ChatMessage, user_id: str):
    '''Get the question answering language model.'''
    openai_api_key, anthropic_api_key, supabase_url, supabase_key = get_environment_variables()
    supabase_client, embeddings = create_clients_and_embeddings(openai_api_key, supabase_url, supabase_key)
    
    vector_store = CustomSupabaseVectorStore(
        supabase_client, embeddings, table_name="vectors", user_id=user_id)
    memory = ConversationBufferMemory(
        memory_key="chat_history", return_messages=True)
    
    ConversationalRetrievalChain.prompts = LANGUAGE_PROMPT
    
    qa = None
    # this overwrites the built-in prompt of the ConversationalRetrievalChain
    ConversationalRetrievalChain.prompts = LANGUAGE_PROMPT
    if chat_message.model.startswith("gpt"):
        qa = ConversationalRetrievalChain.from_llm(
            ChatOpenAI(
                model_name=chat_message.model, openai_api_key=openai_api_key, 
                temperature=chat_message.temperature, max_tokens=chat_message.max_tokens), 
                vector_store.as_retriever(), memory=memory, verbose=True, 
                max_tokens_limit=1024)
    elif anthropic_api_key and chat_message.model.startswith("claude"):
        qa = ConversationalRetrievalChain.from_llm(
            ChatAnthropic(
                model=chat_message.model, anthropic_api_key=anthropic_api_key, temperature=chat_message.temperature, max_tokens_to_sample=chat_message.max_tokens), vector_store.as_retriever(), memory=memory, verbose=False, max_tokens_limit=102400)
    return qa