Ask My Files

The purpose of this project is to give a user the ability to upload files (PDFs, text files and audio files as well as URLs and YouTube videos with transcriptions) and then be able to search them using Open AI embeddings. Using Langchain's Node.js package, I have the user upload a file from my website which then goes to an S3 bucket. From there, it gets processed by a Lambda which then converts it to text, slices the text up into 500 character chunks, creates embeddings from it and then inserts it into a vector database. For the live version on the website, I am using Pinecone, although I created alternate backend repos that use MongoDB and PostgreSQL with the pgvector extension. Then, in the website I created, the user can then query the vector database that will show them which of the files they uploaded, and if it is a PDF or text file, what page it most closely answers their question. It also highlights the relevant words in the search so that the user knows which words matched. I also created an alternate URL where the user can do question and answers with the files using AWS Bedrock to host Claude 2. The PostgreSQL version also has the ability for the user to select different LLMs. The last feature I am going to add is to have the prompts be saved, similar to ChatGPT, where each prompt can provide context to the following prompts. This is currently in beta with a small set of users testing the functionality of the website.