Agentic RAG pipeline for YouTube and podcast transcripts: ingest videos, then ask questions and get answers with clickable timestamp links. Built with LangChain, Supabase (pgvector), Voyage (embeddings), and Claude (Anthropic).

1. Ingest — Give a YouTube URL; the pipeline downloads the transcript, merges short segments into paragraphs, runs semantic chunking, embeds with Voyage, and stores in Supabase.
2. Ask — Ask any question; a Claude agent uses retrieval tools (search transcripts, list episodes, get context) and answers with citations in the form [Episode Title | MM:SS | link].
3. Supports English and Mandarin transcripts (YouTube provides them; no extra config).

git clone <repo>
cd podwise
uv sync   # or: pip install -e ".[dev]"

cp .env.example .env
# Edit .env: set SUPABASE_URL, SUPABASE_SERVICE_KEY, VOYAGE_API_KEY, ANTHROPIC_API_KEY.

Ingest a video (transcript → chunks → Supabase):

uv run python -m src.index "https://www.youtube.com/watch?v=VIDEO_ID"

Ask a question over all ingested content:

uv run python main.py ask "What does Saining Xie say about world models?"
uv run python main.py ask "Compare what was said about sleep across episodes"

Answers are printed in the terminal with Markdown formatting and citation links.

List indexed episodes (optional):

uv run python -m src.retrieval.tools list

podwise/
├── main.py                 # CLI: ask
├── src/
│   ├── config.py           # Env settings
│   ├── index.py            # Ingest: YouTube → clean → chunk → embed → Supabase
│   ├── ingestion/          # youtube, cleaner, chunker (semantic)
│   ├── embedding/         # Voyage embedder
│   ├── storage/            # Supabase (pgvector) wrapper
│   ├── retrieval/         # search_transcripts, get_episode_list, get_episode_context (+ LangChain tools)
│   └── agent/              # Claude + tools (ReAct-style)
└── .env.example