systex_phase2/semantic_search.py

### Python = 3.9
import os
from dotenv import load_dotenv
load_dotenv('.env')

import openai 
openai_api_key = os.getenv("OPENAI_API_KEY")
openai.api_key = openai_api_key

from langchain_openai import OpenAIEmbeddings
embeddings_model = OpenAIEmbeddings()

from langchain_community.document_loaders.csv_loader import CSVLoader
from langchain_community.vectorstores import Chroma

import pandas as pd
import re

from langchain_community.embeddings.openai import OpenAIEmbeddings
from langchain_community.vectorstores import SupabaseVectorStore
from supabase.client import create_client

def grandson_vectordb(vectordb_directory = "./chroma_grandson"):
    questions = ['我的跨損啊', "我要看孫子"]
    
    vectorstore = Chroma.from_texts(
        texts=questions,
        embedding=embeddings_model,
        persist_directory=vectordb_directory
        )
    return vectorstore

def grandson_semantic_cache(q, SIMILARITY_THRESHOLD=0.83, k=1, vectordb_directory = "./chroma_grandson"):
    vectordb_directory = "./chroma_grandson"
    if os.path.isdir(vectordb_directory):
        vectorstore = Chroma(persist_directory=vectordb_directory, embedding_function=embeddings_model)
    else:
        print("create new vector db ...")
        vectorstore = grandson_vectordb(vectordb_directory)

    docs_and_scores = vectorstore.similarity_search_with_relevance_scores(q, k=1)
    doc, score = docs_and_scores[0]

    
    if score >= SIMILARITY_THRESHOLD:
        cache_question = doc.page_content
        answer = "你有三個孫子，男生在國小念書，你要看他照片嗎"
        return cache_question, answer
    else:
        return None, None
    
def create_qa_vectordb(supabase, vectordb_directory="./chroma_db_carbon_questions"):

    if os.path.isdir(vectordb_directory):
        vectorstore = Chroma(persist_directory=vectordb_directory, embedding_function=embeddings_model)
        vectorstore.delete_collection()

    response = supabase.table("QA_database").select("Question, Answer").execute()
    questions = [row["Question"] for row in response.data]

    ######### generate embedding ###########
    # embedding = embeddings_model.embed_documents(questions)

    ########## Write embedding to the supabase table  #######
    # for id, new_embedding in zip(ids, embedding):
    #     supabase.table("video_cache_rows_duplicate").insert({"embedding": embedding.tolist()}).eq("id", id).execute()

    ########### Vector Store #############
    # Put pre-compute embeddings to vector store. ## save to disk
    vectorstore = Chroma.from_texts(
        texts=questions,
        embedding=embeddings_model,
        persist_directory=vectordb_directory
        )
    
    return vectorstore


# vectorstore = Chroma(persist_directory="./chroma_db_carbon_questions", embedding_function=embeddings_model)
def semantic_cache(supabase, q, SIMILARITY_THRESHOLD=0.83, k=1, vectordb_directory="./chroma_db_carbon_questions"):

    if os.path.isdir(vectordb_directory):
        vectorstore = Chroma(persist_directory=vectordb_directory, embedding_function=embeddings_model)
    else:
        print("create new vector db ...")
        vectorstore = create_qa_vectordb(supabase, vectordb_directory)

    docs_and_scores = vectorstore.similarity_search_with_relevance_scores(q, k=1)
    doc, score = docs_and_scores[0]

    
    if score >= SIMILARITY_THRESHOLD:
        cache_question = doc.page_content
        response = supabase.table("QA_database").select("Question, Answer").eq("Question", cache_question).execute()
        # qa_df = pd.DataFrame(response.data)
        # print(response.data[0])
        answer = response.data[0]["Answer"]
        # video_cache = response.data[0]["video_cache"]
        return cache_question, answer
    else:
        return None, None