Or try a sample:
How it works
- Transcript. The server fetches the video's captions (the samples ship with their transcripts, so they work even when YouTube blocks the server).
- Chunk & index. Captions are merged into ~30-second chunks that keep their exact start and end times, embedded with
gemini-embedding-001and stored in an in-memory FAISS index. One-off cost for a 5-minute video: about $0.0002. - Retrieve. Each question pulls the 5 most similar chunks.
- Answer.
gemini-2.5-flash-litesees each chunk as[mm:ss–mm:ss] textand may only cite those times. It's told to say so when the transcript doesn't cover the question. - Verify. The server checks every cited timestamp against the retrieved chunks and removes any that don't match, so a timestamp you can click is one the transcript backs up.
A typical answer uses about 1,000 input and 100 output tokens, roughly $0.00014 at list price ($0.10 / $0.40 per million tokens). The line under each answer shows the actual tokens, cost and server time.
Good to know
- Works on videos with captions. Private, age-restricted or very new videos may not load, and YouTube sometimes blocks the server; the samples always work.
- Answers come from the transcript only, not the visuals or audio.
- The server runs on a free tier and sleeps when idle; the first request can take up to a minute.
- Costs shown are list prices for the Gemini API paid tier (as of 29 Oct 2026). The demo may run on the free tier.
- Sample videos are Creative Commons Attribution (CC BY) licensed: Google, The Conversation and engineerguy (Bill Hammack).