Grab YT Comments

This helps you read YouTube comments without drowning in the noise. It extracts comments from any video URL and classifies them locally with LFM2.5-350M, giving you a clear dashboard of what people are asking, praising, criticizing, joking about, or pushing back on.

Project Structure

grab-yt-comments/
|-- backend/
|   |-- api.py              # FastAPI API for the React frontend
|   |-- app.py              # Existing Gradio UI entrypoint
|   |-- scraper.py          # YouTube Data API comment scraping
|   |-- classifier.py       # LLM classification and SQLite persistence
|   |-- pyproject.toml      # Backend dependencies for uv
|   |-- requirements.txt    # Backend dependencies for pip-style deploys
|   |-- uv.lock
|   `-- .gitignore
|-- frontend/
|   |-- App.tsx
|   |-- services/
|   |-- components/
|   |-- package.json
|   `-- .gitignore
|-- .gitignore
|-- LICENSE
`-- README.md

Prerequisite:

Local LLM must be running

Classification depends on a local OpenAI-compatible LLM server. Before running this project, install and start LFM2.5-350M with llama.cpp.

1. Install llama.cpp:

winget install llama.cpp

2. Start the model server:

llama-server -hf LiquidAI/LFM2.5-350M-GGUF:Q4_K_M

Once started, the server is available at http://localhost:8080.

```

Backend Setup

From the repo root:

cd backend
uv sync

Create backend/.env:

YOUTUBE_API_KEY=your_youtube_api_key_here

Local LLM settings:

LOCAL_LLM_URL=http://127.0.0.1:8080/v1/chat/completions
LOCAL_LLM_MODEL=LiquidAI/LFM2.5-350M-GGUF:Q4_K_M
CLASSIFICATION_REQUEST_TIMEOUT=30

Run The App

Start the FastAPI backend:

cd backend
uv run python -m uvicorn api:app --host 127.0.0.1 --port 8000 --reload

In a second terminal, start the React frontend:

cd frontend
npm install
npm run dev

Open the frontend:

http://127.0.0.1:3000

The frontend calls the backend at:

http://localhost:8000/api/v1/comments

You can override that by setting VITE_API_URL in frontend/.env.

Existing Gradio App

The original Gradio workflow is still available:

cd backend
uv run python app.py

Classification Labels

When classification is enabled, comments are classified with the local OpenAI-compatible LLM endpoint into:

• appreciation
• humor
• questions
• criticism
• personal experience
• feedback
• spam

The backend stores scraped comments first, classifies each saved comment, writes labels back to SQLite, and returns the final result from the database.