How to Use YouTube Transcripts for RAG: Fetch, Chunk and Cite Through an API

To use YouTube transcripts for RAG, fetch each transcript with timestamps, cut it into short windows and store every chunk with the video's metadata and a link to the second it starts. With Scavio's YouTube API that is three calls: /api/v1/youtube/channel/videos lists a channel's uploads 30 at a time, /api/v1/youtube/video returns the title, channel and publish date for each, and /api/v1/youtube/transcript with "format": "srt" returns the timed transcript in one synchronous call. A transcript costs 5 credits, the other two 1 credit each.

YouTube's own Data API only downloads captions for videos you have permission to edit, so it cannot feed a knowledge base built from other people's channels. The calls below were made on October 27, 2026, and every response shown is captured, with long fields trimmed.

Step 1: List the videos you want to index

curl -s https://api.scavio.dev/api/v1/youtube/channel/videos \
  -H "Authorization: Bearer $SCAVIO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"channel_id": "@IBMTechnology"}'

channel_id takes a channel id, an @handle or a channel URL. For @IBMTechnology this returned channel UCKWaEZ-_VweaEx1j62do_vQ, 30 uploads and has_more: true with a next_cursor for the next page. One row, as captured (thumbnail query string trimmed):

{
  "video_id": "615-yWyS1n0",
  "title": "What Is LLM Inference? How AI Models Generate Responses",
  "url": "https://www.youtube.com/watch?v=615-yWyS1n0",
  "thumbnail": "https://i.ytimg.com/vi/615-yWyS1n0/hq720.jpg",
  "duration_text": "7:46",
  "view_count": 96000,
  "published_time": "4 days ago",
  "is_live": false
}

To index a topic rather than a channel, /api/v1/youtube/search with {"search": "retrieval augmented generation explained", "type": "video"} returned 20 videos a page for 2 credits.

Step 2: Get the metadata you store with each chunk

curl -s https://api.scavio.dev/api/v1/youtube/video \
  -H "Authorization: Bearer $SCAVIO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"video_id": "T-D1OfcDW1M"}'

{
  "video_id": "T-D1OfcDW1M",
  "title": "What is Retrieval-Augmented Generation (RAG)?",
  "author": "IBM Technology",
  "channel_id": "UCKWaEZ-_VweaEx1j62do_vQ",
  "published_at": "2023-08-23T04:00:32-07:00",
  "length_seconds": 395,
  "view_count": 2075642,
  "keywords": ["IBM", "IBM Cloud", "AI", "Artificial Intelligence", "LLM"],
  "playability_status": "OK",
  "captions": [{ "language_code": "en", "language_name": "English" }]
}

published_at lets your retriever prefer recent videos, author and title make citations readable, and captions (each track also has a url; that and the 13-item keywords list are trimmed here) tells you which languages exist before you ask for one.

Step 3: Fetch the transcript with timestamps

curl -s https://api.scavio.dev/api/v1/youtube/transcript \
  -H "Authorization: Bearer $SCAVIO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"video_id": "T-D1OfcDW1M", "format": "srt"}'

The response holds video_id, language_code, language_name, format and content. For this 6-minute-36-second video, content was 92 SRT cues, starting:

Leave out format for plain text. Length does not change the call: the 2-hour-33-minute freeCodeCamp course "Learn RAG From Scratch" came back as 27,073 words in 2.6 seconds, and across 19 English videos the median call took 2.1 seconds.

Step 4: Handle videos without your language

Not every video has an English track. Asking for one on a Hindi-only lecture returned:

{
  "video_id": "Ty8gcCKuwNI",
  "language_code": "en",
  "language_name": "English",
  "format": null,
  "content": null,
  "available_languages": ["hi"],
  "message": "No English transcript for this video. Available languages: hi."
}

Retry with "language": "hi" and the transcript comes back (3,771 words here). The first call is still billed, so if you index many languages, read captions from step 2 first and request a language the video has. In our 20-video sample, 19 had English and 1 did not.

Step 5: Chunk by time and keep the timestamp

Sixty-second windows keep chunks small and give each one a start time you can link to:

def parse_srt(srt):
    cues = []
    for block in srt.strip().split("\n\n"):
        lines = block.splitlines()
        if len(lines) < 3:
            continue
        h, m, s = lines[1].split(" --> ")[0].replace(",", ".").split(":")
        cues.append((int(h) * 3600 + int(m) * 60 + float(s), " ".join(lines[2:])))
    return cues

def chunk(cues, video_id, window=60):
    chunks, cur, t0 = [], [], None
    for start, text in cues:
        if t0 is None:
            t0 = start
        if start - t0 >= window and cur:
            chunks.append({"start": int(t0), "text": " ".join(cur),
                           "url": f"https://www.youtube.com/watch?v={video_id}&t={int(t0)}s"})
            cur, t0 = [], start
        cur.append(text)
    if cur:
        chunks.append({"start": int(t0), "text": " ".join(cur),
                       "url": f"https://www.youtube.com/watch?v={video_id}&t={int(t0)}s"})
    return chunks

Run on the captured SRT above, this gave 7 chunks starting at 0, 60, 123, 185, 248, 310 and 372 seconds, of 199 to 229 tokens each plus a 49-token tail (o200k_base tokenizer). The second chunk starts "and the article said that it was Jupiter and 88 moons" and links to https://www.youtube.com/watch?v=T-D1OfcDW1M&t=60s.

Step 6: Embed, store and cite

Store each chunk with the metadata from step 2: video_id, title, author, published_at, start and url. Embed the text with the model you already use and write it to your vector store. When the retriever returns a chunk, the answer can cite "IBM Technology, What is Retrieval-Augmented Generation (RAG)?, at 1:00" with a link that opens the video at that second.

Plan the size before you embed. In our sample the median was 196 tokens per minute of video, so a 10-minute video is about 2,000 tokens and 7.2 hours of video was 83,000. For new uploads, re-run step 1 on a schedule and fetch transcripts only for video ids you have not stored.

All 17 YouTube endpoints

Indexing a 500-video channel is 17 list pages and 500 transcripts: 2,517 credits, about $5 on the Growth plan billed yearly. For a side-by-side of every option, see the best YouTube transcript APIs for RAG.

Getting started

1. Sign up at dashboard.scavio.dev/sign-up: 50 free credits, enough for 10 transcripts, no card required.
2. Copy your API key from the dashboard.
3. Send a POST to /api/v1/youtube/transcript with a video id and your key in the Authorization header.

Try the YouTube API free in the playground first, or give an agent the same calls through the MCP server.

Reference: YouTube API overview · transcript · video · channel videos · search · is there a stable way to pull YouTube transcripts?