Skeletonboy/healthygamer-rag-assistant

mental-health-rag-assistant

Ask questions to a LLM with detailed up-to-date knowledge from the entire HealthyGamerGG Youtube Channel.

Currently hosted on HuggingFace Spaces at: https://huggingface.co/spaces/Skeletonboy/healthygamer-rag-assistant

This code repo contains:

• —yt-transcript.py - automatic video transcript scraper for all uploads of any Youtube channel
• —vector_db.py - FAISS vector store wrapper to use custom normalized embeddings
• —transcript_ops.py - transcript summarizer using third-party LLMs
• —app.py - deploy RAG question-answerer as Gradio app

Each video from the HealthyGamerGG Youtube channel is summarized in detail using GPT-4o, and embedded using stella_en_1.5B_v5 before being passed into a FAISS vector index for quick similarity-searching.

By passing in a mental-health related user query, the query is then embedded locally and used to retrieve top-K relevant summarized video transcripts.

All code written 100% completely from scratch.

Setup

To use, install required libraries: ''' pip install -r requirements.txt '''

Setup your API tokens in a .env file:

OPENAI_API_KEY = ...
YT_API_KEY = ...
HF_WRITE_TOKEN = ...

Run yt_transcript.py to pull transcripts from every youtube video available in the channel. Auto-generated transcripts will be used for videos without manually uploaded transcripts.

yt_transcript.py <YT_CHANNEL_ID> <transcript_savepath.json>

Run transcript_ops.py to clean, filter, and summarize transcripts. Embedding model is compatible on both CPU and GPU devices:

transcript_ops.py <transcript_savepath.json> <new_svepath.json>

Set file paths and run app.py to publish as Gradio app

gradio app.py