PlaylistGuide-LLM

An LLM Chatbot which assists the user with answering questions about and finding locations of content in a Youtube Playlist

Problem Statement

Playlists can be quite long and finding the exact location of content can be tedious. This project aims to assist users with finding the locations (video and timestamp) of specific content and be able to answer questions about the content in a given playlist.

Project Overview

There are three main directories of interest in this project

• app - This is the directory where the transcript RAG/LLM Application is
• transcript_pulls - This directory contains research for several ways in which the transcripts were attempted to be pulled
• evaluation - This directory contains the generation of a ground truth data set as well as Hit Rate and MRR Evaluations

Walkthrough Video

Steps to reproduce

Optional - log into a GCP account and search for YouTube

You may need to Enable the API. Once enabled you can then click Manage

Click the credentials button on the left and then + Create Credentials

You do not have to pull the transcripts as an up to date pickle file should be present in the PlaylistGuide-LLM/app/data directory. If you do wish to pull fresh data, you will simply need to delete the processed.txt file (this keeps track of video IDs already pulled so they are not pulled on subsequent runs) and run pull_transcripts_metadata.py

Create OpenAI API Key and add to .env file

Setup an OpenAI API Account
Go to: https://platform.openai.com/settings/profile?tab=api-keys
Generate an API Key

and add it to your .env file

Create a .env File

PostgreSQL

POSTGRES_USER=postgres_user
POSTGRES_PASSWORD=postgres_password
POSTGRES_DB=user_feedback

Elasticsearch

ELASTIC_URL_LOCAL=http://localhost:9200
ELASTIC_URL=http://elasticsearch:9200
ELASTIC_PORT=9200

pgAdmin

PGADMIN_DEFAULT_EMAIL=[email protected]
PGADMIN_DEFAULT_PASSWORD=admin_password

OpenAI API Key

OPENAI_API_KEY=your_openai_api_key

YouTube API Key

YouTube_API_KEY=your_youtube_api_key

Clone the respository locally

git clone https://github.com/MichaelShoemaker/PlaylistGuide-LLM.git

change to the streamlit app directory
cd PlayListGuide-LLM/app

if you are on Linux you should be able to just run
make up
or if you are on another system the equivelant of
docker compose up --build

Note: It will take some time on the first run. Probably 15-30 minutes depending on your system.

Transcript Search

You should then be able to navigate to http://localhost:8501/ and search the transcripts

If you get a message with Video Title: N/A it most likely means Elasticsearch is not loaded yet. Give it several more minutes.
You can then try again.

Other directories of note:

• transcript_pulls - several different ways in which I tested pulling the transcripts
  • Within this directory under metadata_transcripts you can see the steps which were taken to generate ground_truth and evaluate Elasticsearch
• research - simply a "dumping ground" of other testing scripts and expiriments without structure or organization. This may be removed.