TubeChat Pro (YouTube RAG Assistant)

Streamlit app that lets you chat with a YouTube video using a local RAG stack. The app downloads a video's audio with yt-dlp, transcribes it with Whisper, indexes the transcript with Chroma + Ollama embeddings, and answers your questions with an Ollama-served LLM.

Features

• Paste a YouTube URL and auto-download the audio (mp3).
• Local transcription via Whisper (base model).
• RAG pipeline: chunking with RecursiveCharacterTextSplitter, embeddings with nomic-embed-text, retrieval via Chroma, responses with llama3.2.
• Chat-style UI built in Streamlit; transcript is viewable in-app.
• Audio file is cleaned up after processing; transcript is written to transcription.txt for reuse.
• CLI sample (main.py) that demonstrates the same RAG flow against a web page.

Prerequisites

• Python 3.10+ and pip.
• FFmpeg installed and available on PATH (required by yt-dlp and Whisper).
• Ollama running locally with the models:
  • ollama pull nomic-embed-text
  • ollama pull llama3.2

Installation

pip install \
  streamlit \
  yt-dlp \
  openai-whisper \
  langchain \
  langchain-community \
  langchain-text-splitters \
  langchain-ollama \
  chromadb

Running the Streamlit app

streamlit run app.py

1. Open the provided local URL.
2. Paste a YouTube link and click ANALYZE ⚡.
3. After processing, expand Show Transcript if you want to inspect the raw text.
4. Ask questions in the chat box; answers are grounded in the transcript.

Notes:

• Audio is removed after analysis; the transcript persists at transcription.txt (also listed in .gitignore).
• If the page fails to load models, confirm Ollama is running and the two models above are available.

Using the CLI example

python main.py

The script loads a sample Turkish Wikipedia page, builds the same RAG chain, and answers questions in a loop (q to quit).

Troubleshooting

• ffmpeg not found: Install FFmpeg and ensure it is on PATH.
• Ollama errors / model not found: Pull the required models and restart Ollama.
• Long processing times: Whisper transcription and embedding creation are CPU-heavy; GPU-capable environments speed this up.
• Permission issues writing files: Ensure the repo directory is writable; audio and transcript files are created in the project root.

👨‍💻 Author

Batuhan Küçükaydın
Software Engineer | Computer Engineer | iOS Developer
📫 LinkedIn • GitHub • Medium

⭐️ Support

If you like this project, please consider giving it a star 🌟
It really helps me keep building and improving!