load_qa_chain.py

# Do the below in supabase
# create table public.document_embeds_openai(id bigint generated by default as identity not null,
# content text null,
# sources text null,
# embedding vector(1536) null,
# metadata jsonb null,
# constraint document_embeds_open_pkey primary key(id) ) tablespace pg_default;

# Also create the below function in supabase

# create or replace function match_documents_chain_embeds( query_embedding vector(768),
#                                                          match_thresholdfloat,match_count int)
# returns table(id bigint, sources text,content text, similarity float, metadata jsonb)
# language
# sql stable
# as $$
# select
# document_embeds_openai.id,
# document_embeds_openai.sources,
# document_embeds_openai.content,
# 1 - (document_embeds_openai.embedding <= > query_embedding) as similarity,
# metadata from document_embeds_openai where
# 1 - (document_embeds_openai.embedding <= > query_embedding) > match_threshold
# order by similarity desc limit match_count;
# $$;

import os
from typing import List

from dotenv import load_dotenv
from langchain import OpenAI
from langchain.schema import Document
from supabase import Client, create_client

from embeddings.openaiembeddings import OpenAIEmbeds
from llm_prompt_template.load_qa_prompt_template import LoadQATemplate

load_dotenv()

OPENAI_API_KEY = os.environ.get('OPENAI_API_KEY')
url = os.environ.get('SUPABASE_URL')
key = os.environ.get('SUPABASE_SECRET_KEY')


def ingest_documents(emb: OpenAIEmbeds, documents: List[Document], supabase: Client):
    document_list = []

    for document in documents:
        embedded_document = emb.generate_embeddings([document.page_content])
        row = {'content': document.page_content,
               'embedding': embedded_document,
               'sources': document.metadata['source'],
               'metadata': document.metadata
               }
        document_list.append(row)
        row = {}
    # data = supabase.table('document_embeds').insert(document_list).execute()
    data = supabase.table('document_embeds_openai').insert(document_list).execute()

def retrieve_top_k_docs(emb,supabase: Client, query) -> List[Document]:
    query_embedding = emb.generate_embeddings(query)
    supabase: Client = create_client(url, key)
    resp = supabase.rpc("match_documents_chain_embeds", params={'query_embedding': query_embedding,
                                                                'match_threshold': 0.7,
                                                                'match_count': 1}).execute()

    # print(resp.data)
    resp_len = len(resp.data)
    # for i in range(resp_len):
    #     print(resp.data[i]['id'])
    #     print(resp.data[i]['sources'])
    #     print(resp.data[i]['content'])
    #     print(resp.data[i]['similarity'])

    documents = [Document(
            page_content="source::" + resp.data[i]['sources'] + "content::" + resp.data[i]['content'],
            metadata=resp.data[i]['metadata']
    ) for i in range(len(resp.data))]

    return documents

if __name__=="__main__":
    supabase: Client = create_client(url, key)
    emb = OpenAIEmbeds()

    # Run the below commented lines once first time to load the vector table in supabase

    # Sample docuemnt to load
    # docs = [Document(page_content="Langchain is a LLM fraemwork. It abstracts the LLM model operations. "
    #                               "Thus it helps to adopt a polyglot architecture", metadata={"source": "my_website"})]
    #
    # ingest_documents(emb=emb,documents=docs,supabase=supabase)

    llm = OpenAI(model_name="text-davinci-003",
                     openai_api_key=OPENAI_API_KEY,
                     temperature=0,
                     max_tokens=1000)

    from langchain.chains.question_answering import load_qa_chain
    query = "What is langchain?"
    docs = retrieve_top_k_docs(emb=emb,supabase=supabase,query=query)
    prompt_template = LoadQATemplate()
    prompt = prompt_template.get_prompt()
    print(prompt)
    chain = load_qa_chain(llm, chain_type="stuff",prompt=prompt)
    resp = chain.run(input_documents=docs, question=query, tone="Sad")

    print(resp)