Chat with YouTube channels instead of watching them

What it does

YT Navigator is a Django application that ingests up to 100 videos from a YouTube channel, extracts transcripts, and stores them in PostgreSQL with PGVector embeddings. Users query the archive through natural language search or chat with a ReAct agent that retrieves specific transcript segments and exact timestamps. It is essentially a full-stack RAG interface built on top of YouTube content.

The interesting bit

The search layer is deliberately over-engineered in a useful way: it runs semantic and BM25 keyword searches in parallel, deduplicates and reranks with a cross-encoder, then groups results by video. The chat layer routes queries through llama-3.1-8b-instant to decide whether to answer directly, deflect irrelevant questions, or hand off to a qwen-qwq-32b ReAct agent that can run semantic search and raw SQL against the database. That orchestration is more sophisticated than the usual vector-DB-demo fare.

Key highlights

• Hybrid retrieval: vector similarity plus BM25, fused and reranked before returning timestamped transcript chunks.
• ReAct chat agent with a routing layer: three response paths (direct, static rejection, or tool-assisted) to reduce hallucinations and keep smaller models useful.
• Uses scrapetube and youtube-transcript-api for ingestion, bge-small-en-v1.5 for embeddings, and Groq-hosted models for inference.
• Full Django monolith with server-rendered templates; no separate frontend framework.
• Supports per-user sessions and independent channel scans.

Caveats

• The roadmap still lists Docker support and tests as unfinished, even though the README already includes installation instructions for both.
• Ingestion is capped at 100 videos per channel, and the roadmap explicitly flags the mobile experience as needing improvement.

Verdict

Researchers and developers who treat YouTube channels as primary sources will find this a practical, if unpolished, research assistant. If you need production-grade ingestion beyond 100 videos or a polished mobile UI, treat it as a proof of concept.