From transcript to highlight reel: chapters, key moments & query reels

Once your media is transcribed, embedded, and labeled, you're holding a remarkably rich data structure: a list of timestamped segments, each with text, a speaker, and a vector. Search is the obvious thing to build on it. But the same foundation powers a second family of features that turn passive footage into something you can actually edit: chapters, highlights, and reels.

All three reuse the artifact from the transcription pipeline — no new models, no cloud round-trip. Here's how a flat timeline becomes navigable, then publishable.

Chapters: grouping segments by topic, not by clock

The naïve way to chapter a video is to chop it every five minutes. That's useless — it cuts mid-thought and ignores where the conversation actually turns. MediaFind chapters on meaning instead. Because every segment already has an embedding, we can measure how the topic drifts from one segment to the next and place a boundary where the semantic distance spikes.

Each chapter gets a short title drawn from the text it contains — its most distinctive keywords — so the sidebar reads like a table of contents you didn't have to write. Click one and the player jumps straight there.

Key moments: ranking, not just listing

A highlight is a moment worth keeping. MediaFind ranks segments by how central each one is to the recording's overall meaning — it embeds every segment, takes the centre of mass of those vectors, and scores each segment by how close it sits to that centroid. A diversity penalty (maximal marginal relevance) then keeps the picks from all saying the same thing, so the top spans cover the recording rather than crowding around one moment. You're not scrubbing a timeline hunting for the good part; the good parts are proposed for you, each a real timestamped span you can accept, trim, or ignore.

Query reels: a search that returns a video

Here's where it gets fun. A query reel takes a search — "every time we mention the competitor" — and instead of a list of links, returns a single stitched clip of all the matching moments, in order. Under the hood it's the search pipeline feeding the clip exporter.

It's all the same artifact

Chapters, key moments, and query reels feel like three features, but they're three views of one structure: timestamped, embedded segments. That's the recurring theme of MediaFind's design — do the expensive work once, on-device, and let everything else be a cheap transformation on top.

Next: the plumbing that makes all of this possible at library scale — how a folder of files becomes an index, and how that index stays correct as the folder changes.