Project Architecture

Overview

This document outlines the overall architecture of the project and the workflow from file upload to response delivery.

Workflow

User sends file to Backend
Storing files on server
This makes the process slow. To optimize your chat app, you can remove the writing text file and saving file on the server (both PDF files and text files will be saved in the public/temp folder).

Sending File to Cloudinary
Cloudinary is free and great for testing. It functions similarly to AWS S3 buckets.

Queue Management
A queue is created in the controller file to sync the process using BULLMQ.

File Processing
Processing is done in worker.js file:

• Created raw text from the file path.
• Chunked the text into 1000-character segments.
• Sent the segments to generate vector embeddings using Gemini.
• Saved the vector embeddings into a PostgreSQL database.
  (For this, you can pull the pgvector Docker image and connect it.)

First Part Completed
User project got created and saved.

Prompt Handling

User sends a prompt
Convert the prompt into vector embeddings using the same process and query the database to get the nearest neighbor. I used cosine similarity; you can use L1 or L2 based on your needs.

Database Query
The database returns all relevant data. Extract the text from the returned data.

Text Cleaning
Clean the extracted text. I wrote a simple algorithm to clean the text. Optimization can be done in two places:

1. Don't ask for all the data, only request the text to send to the LLM.
2. Use a library to clean the data.

Sending to LLM
Send the cleaned text to the LLM (Large Language Model).

Returning Response
Receive the response from the LLM and send it back to the user.

This is how Vector Embeddings are Stored in DB