yt-summarizer

Production-style RAG system for YouTube transcript summarization and question answering.

Features

• Fetches English transcripts from YouTube
• Summarizes the video content
• Supports question answering over the transcript
• Provides a local web interface with Gradio
• Exposes a FastAPI backend for programmatic access
• Uses centralized runtime configuration through .env
• Emits structured logs for pipeline steps, latency, warnings, and errors

Libraries Used

• gradio for the web UI
• fastapi and uvicorn for the API backend
• youtube-transcript-api for retrieving video transcripts
• langchain for prompt and chain orchestration
• langchain-ollama for the LLM integration
• langchain-huggingface and sentence-transformers for embeddings
• faiss-cpu for vector search over transcript chunks

Installation

Install dependencies:

uv sync

The default LLM backend uses Ollama with phi3:mini:

ollama pull phi3:mini

Configuration

Create a local .env file from the example:

cp .env.example .env

Main runtime parameters:

YT_LLM_MODEL=phi3:mini
YT_LLM_TEMPERATURE=0.5
YT_LLM_TIMEOUT_SECONDS=60
YT_LLM_RETRY_ATTEMPTS=2
YT_EMBEDDING_MODEL=sentence-transformers/all-MiniLM-L6-v2
YT_CHUNK_SIZE=1000
YT_CHUNK_OVERLAP=200
YT_RETRIEVAL_TOP_K=7
YT_DATA_DIR=data
YT_LOG_LEVEL=INFO
YT_LOG_JSON=true
YT_API_PORT=8000
YT_GRADIO_PORT=7865

Run the Gradio UI

Start the FastAPI backend first:

uv run yt-summarizer-api

Then run the Gradio frontend:

uv run yt-summarizer-ui

The UI launches locally on port 7865 by default and calls the API backend.

Run the FastAPI Backend

uv run yt-summarizer-api

The API launches locally on port 8000.

Useful URLs:

http://localhost:8000/health
http://localhost:8000/docs

Example request:

curl -X POST http://localhost:8000/ingest_video \
  -H "Content-Type: application/json" \
  -d '{"video_url": "https://www.youtube.com/watch?v=VIDEO_ID"}'

Then query the online endpoints using the returned video_id:

curl -X POST http://localhost:8000/summarize \
  -H "Content-Type: application/json" \
  -d '{"video_id": "VIDEO_ID"}'

curl -X POST http://localhost:8000/ask \
  -H "Content-Type: application/json" \
  -d '{"video_id": "VIDEO_ID", "question": "What is the main topic?"}'

Alternative Local Entrypoint

uv run src/main.py