George Avatar Chatbot (Dynavap RAG)

Overview

• Web chatbot that feels like chatting with “George,” the Dynavap expert from YouTube.
• Uses RAG over transcripts from George’s YouTube videos. Transcripts are generated using OpenAI speech-to-text (Whisper via gpt-4o-mini-transcribe).
• Starts with animated avatar + text responses; optional in-browser TTS.

Project Structure

• server/ — FastAPI backend with RAG endpoint and transcript ingestion.
• web/ — Minimal static chat UI with animated avatar, optional TTS, and source links under responses.
• data/ — Transcripts and vector index (created after ingest).

Quick Start

1. Prereqs

• Python 3.10+
• An OpenAI API key in OPENAI_API_KEY
• Optional: YouTube Data API key in YOUTUBE_API_KEY (only for listing videos by channel; transcription is handled by OpenAI, not the YouTube API)

2. Install deps

python -m venv .venv
source .venv/bin/activate
pip install -r server/requirements.txt

3. Ingest transcripts

Option A — Provide a list of video URLs/IDs in data/videos.txt (one per line):

python server/ingest.py --videos-file data/videos.txt --max-videos 50

Option B — Use a YouTube channel ID (needs YOUTUBE_API_KEY):

export YOUTUBE_API_KEY=YOUR_KEY
python server/ingest.py --channel-id UCxxxxxxxxxxxx --max-videos 50

This downloads audio via yt-dlp, transcribes it with OpenAI, then creates data/transcripts.jsonl and a FAISS index in data/index/.

4. Run the server

export OPENAI_API_KEY=YOUR_OPENAI_KEY
uvicorn server.app:app --reload --port 8000

Open http://localhost:8000 to use the chatbot.

Notes

• Persona: The assistant responds in George’s friendly, knowledgeable style focused on Dynavap. You can customize the prompt in server/app.py.
• TTS: The web client uses the browser’s speechSynthesis as a fallback. A server TTS provider can be added later.
• Assets: The avatar is a lightweight CSS/SVG animation; you can replace it with a Lottie or video later.
• Avatar image: Place your George picture as goerge.png (or george.png) in the project root. The web UI loads it from /avatar and displays it as-is. Replace anytime.
• Transcription model: Set OPENAI_TRANSCRIBE_MODEL to change the model (default gpt-4o-mini-transcribe).

Troubleshooting downloads/transcription

• Some videos require cookies or get rate-limited. Provide a cookies file (Netscape format, exported from your browser) and/or a proxy. These are passed to yt-dlp for the audio download:

python server/ingest.py --videos-file data/videos.txt \
  --cookies /path/to/youtube_cookies.txt \
  --proxy http://127.0.0.1:8080

• You can also set env vars: YT_COOKIES and YT_PROXY.
• If transcription returns empty, the script will error for visibility. Share the console output if you need help.