RAG over YouTube videos

Let users ask questions about a channel, a course or a conference, and answer with a link to the exact second where it was said. The hard part is getting clean, timestamped transcripts at scale; the rest is a standard RAG pipeline.

The pipeline

1. List the videos of a channel or playlist with /v1/videos.
2. Fetch transcripts 50 at a time with /v1/batch.
3. Chunk the timestamped segments into passages of a few hundred words, keeping each passage's start time.
4. Embed the passages and store them with video_id and start.
5. Retrieve and answer, citing https://youtu.be/VIDEO_ID?t=START.

1-2. Get the transcripts

import os, requests

API = "https://api.transkrio.com/v1"
H = {"x-api-key": os.environ["TRANSKRIO_API_KEY"]}

videos = requests.get(f"{API}/videos", headers=H,
                      params={"url": "@lexfridman", "limit": 100}).json()["videos"]
urls = [v["url"] for v in videos]

transcripts = []
for i in range(0, len(urls), 50):
    res = requests.post(f"{API}/batch", headers=H, timeout=300,
                        json={"urls": urls[i:i + 50]}).json()
    transcripts += [t for t in res["results"] if t["ok"]]

Videos without captions come back with ok: false and cost nothing.

3. Chunk with timestamps

def chunks(t, max_words=400):
    buf, start = [], None
    for s in t["segments"]:
        start = s["start"] if start is None else start
        buf.append(s["text"])
        if sum(len(x.split()) for x in buf) >= max_words:
            yield {"video_id": t["video_id"], "start": int(start), "text": " ".join(buf)}
            buf, start = [], None
    if buf:
        yield {"video_id": t["video_id"], "start": int(start), "text": " ".join(buf)}

passages = [c for t in transcripts for c in chunks(t)]

4-5. Embed, retrieve, cite

Use any embedding model and vector store (pgvector, Qdrant, Chroma, a NumPy array for a prototype). At query time, retrieve the top passages and put them in the prompt with their link:

context = "\n\n".join(
    f"[https://youtu.be/{p['video_id']}?t={p['start']}]\n{p['text']}" for p in top_passages
)
prompt = f"Answer using only these excerpts and cite the links.\n\n{context}\n\nQuestion: {question}"

Tips

• Prefer manual captions (is_generated: false) when both exist; Transkrio picks them by default.
• Auto-generated captions have no punctuation. Embeddings handle it fine; for display, ask the LLM to clean the quoted passage.
• Store the transcript once. Re-fetching is cheap (cached), but you don't need to.
• Add podcasts to the same index: /v1/videos and /v1/batch accept podcast feeds too.

No-code version: the n8n node plus n8n's vector store nodes. Agent version: give your model the MCP server. More code in YouTube transcripts in Python.