Create and retrieve embeddings using Langchain and Supabase

retrievalQA.js

////////////////////////////////////
/* "dependencies": {
    "@supabase/supabase-js": "^2.13.1",
    "langchain": "^0.0.44"
} */
////////////////////////////////////


import { OpenAI } from "langchain/llms";
import {
    RetrievalQAChain,
    ConversationalRetrievalQAChain,
} from "langchain/chains";
import { SupabaseVectorStore } from "langchain/vectorstores";
import { OpenAIEmbeddings } from "langchain/embeddings";
import { RecursiveCharacterTextSplitter } from "langchain/text_splitter";
import { ChatOpenAI } from "langchain/chat_models";
import * as fs from "fs";
import * as dotenv from "dotenv";
import { createClient } from "@supabase/supabase-js";

dotenv.config();

const openAIKey = process.env.OPENAI_API_KEY;
const client = createClient(
    process.env.SUPABASE_URL || "",
    process.env.SUPABASE_PRIVATE_KEY || ""
);

export const createEmbeddings = async () => {
    // Initialize the LLM of choice to answer the question.
    const model = new OpenAI({ apiKey: openAIKey });
    const text = fs.readFileSync("pizzareview.txt", "utf8");

    // Split and create docs
    const textSplitter = new RecursiveCharacterTextSplitter({
        chunkSize: 200,
        chunkOverlap: 50,
    });
    const docs = await textSplitter.createDocuments([text]);

    // Create a vector store from the documents.
    const vectorStore = await SupabaseVectorStore.fromDocuments(
        docs,
        new OpenAIEmbeddings(),
        {
            client,
            tableName: "documents",
            queryName: "match_documents",
        }
    );

    // Create a chain that uses the OpenAI LLM and Supabase vector store.
    const chain = RetrievalQAChain.fromLLM(model, vectorStore.asRetriever());
    const res = await chain.call({
        query: "What did he rate the pizza?",
    });
    console.log({ res });
};

// createEmbeddings()

export const query = async () => {
    const chat = new ChatOpenAI({
        modelName: "gpt-3.5-turbo",
        apiKey: openAIKey,
    });

    // console.log(client)
    const vectorStore = await SupabaseVectorStore.fromExistingIndex(
        new OpenAIEmbeddings(),
        {
            client,
            tableName: "documents",
            queryName: "match_documents",
        }
    );

    let chain = ConversationalRetrievalQAChain.fromLLM(
        chat,
        vectorStore.asRetriever(),
        { returnSourceDocuments: true }
    );

    console.log(chain.questionGeneratorChain.prompt.template);

    const query = "What did he rate the pizza?";
    const res = await chain.call({ question: query, chat_history: [] });

    // /* Ask it a follow up question */
    // const chatHistory = query + res.text;
    // const followUpRes = await chain.call({
    //     question: "How was the char on the pizza?",
    //     chat_history: chatHistory,
    // });
};

// query();

This worked perfectly! thank you!

When you call the new OpenAIEmbeddings on line 71, are you paying for that each query? I currently have a similar setup working using the MemoryVectorStore, but it regenerates my document embeddings each query. I'm looking for a way to reduce costs by using supabase vector DB to store all my docs and then only have to generate embedding for the question itself.

Or am I misunderstanding something?

@tehfailsafe hey, im dealing with embedding costs too, just like you were. Did you find a smart way to handle it?

@Jakisundays Have you figured it out? facing the same issue.

Thanks Man! Really appreciated from this.