How to search inside a video for a moment

Scrubbing a timeline to find the six seconds you remember is the worst job in video. Analyze replaces it with a search box: the video is turned once into a timestamped index of captions and transcript, and after that you describe the moment you want and get playable cards back. The index costs real money to build and almost nothing to query, which is the shape of the whole feature.

How the index is built

A video is not searchable. An index of it is. Analyze makes one in a single pass: the source is cut into segments, the audio is transcribed by our own Whisper service into a word-aligned transcript, and each segment's frames are read by a vision model into a written caption. The result is a record where every claim carries a timecode.

That pass is the entire cost of the feature. Searching afterwards reads the cached index rather than the video, which is why the first question on a long file is slow and expensive and the tenth is neither.

Which plan: Starter and up. Analyze is billed 100% from the ✦ gold wallet — roughly ✦0.07 per analysed frame for the index, then ✦1 per search. Signing up grants no gold, so a free account cannot run this even where the section is visible.

What you can actually ask for

The search runs over three different scopes, and picking the wrong one is the usual reason an obvious moment does not come back.

The three search scopes
Scope	What it searches	Good for
Segments	The written captions and the transcript, on the clock	"the part where she explains the price"
Pictures	The sampled frames themselves	"a woman looks out of a window at people gathered outside"
Scenes	The shot structure of the video	"every scene shot outdoors", "the takes with no cut in them"

Phrase a picture search visually — describe what is literally on screen, the way a caption would. Phrase a segment search the way you would describe the content to a colleague. The two are searching different text and reward different wording.

Finding a moment, step by step

Analyse the video once. Open /video, pick Analyze, bring the video in and run the pass. The workspace quotes the frame count and the credit cost before it starts. Everything after this is cheap.
Pick the scope. Segments for what was said or explained, Pictures for what was on screen, Scenes for how it was shot.
Describe the moment. One sentence, concrete. The search is looking for a match against captions written by a model, so the closer your wording is to a plain description of the frame, the better it lands.
Read the confidence. Hits are scored out of ten against a threshold you control. If nothing clears the bar you are shown the closest matches instead of an empty result, labelled as such — lower the threshold to widen the net.
Narrow the search, if you already know roughly where. Tick a handful of segments first and the search runs only inside those. It costs the same and it is much more precise on a long source.
Keep the answer. Finished asks are saved with the video and appear as chips. Clicking one restores the question and its results with no model call and no charge.

Open Analyze — Opens the Analyze workspace. Nothing is charged until you run a pass or press search.

What a search costs, exactly

Two meters move when you press search. The credit charge is a flat ✦1, whatever the length of the video, because a search spends a handful of model calls over the cached index rather than re-reading the source. The daily allowance is charged a flat two-minute equivalent for the same reason.

Action	Gold credits	Charged against the daily minutes
Building the index	about ✦0.07 per analysed frame	the source's real length
One moment search	✦1	2 minutes, flat
Reopening a saved ask	none	none
Exporting the picked segments	none	none

Watch out: The daily minute allowance is 10 / 60 / 120 / 300 / 900 by plan and resets at UTC midnight. It is charged up front and is not refunded if a run fails.

When the search misses

The moment is shorter than a segment. A 30-second segment gets one caption. If the thing you want lasts two seconds inside it, re-run the pass with shorter segments — it costs more, because cost scales with frames.
You asked a picture question in the segment scope. Captions describe the span; frames describe the instant. Switch scopes rather than rewording.
Nothing cleared the confidence bar. You will see the closest matches with a warning rather than an empty page. Lower the threshold before assuming the moment is not there.
It was never said and never shown. The index only knows the transcript and the sampled frames. A detail that appears between two sampled frames and is never spoken is genuinely not in there.

The transcript is the cheap half of the index

Two different systems produced the thing you are searching, and they have very different resolutions. The transcript comes from our own Whisper service and is aligned word by word to the video's clock — it is exact, and anything actually said in the video is findable in it. The captions come from a vision model reading a handful of sampled frames per segment, and they are a description, not a record.

So a question about words lands harder than a question about pictures, and it lands at the timecode of the word rather than the span of the segment. If you can phrase what you are hunting for in terms of what somebody said near it, do — "the part just after he says the price" is a much stronger query than a description of the shot it happens in.

Narrowing a search, and coming back to it later

Two features make repeated work on the same video cheap, and both are easy to miss.

Scope the search to ticked segments. If you already know roughly where in the video the thing is, tick those segments first and the search runs only inside them. It costs the same ✦1 and it is dramatically more precise on a long source.
Reopen a saved ask. Every finished question is stored on the record as a chip. Clicking one restores the question and its results with no model call and no charge — which means the answers you found last month are still there, and still free.
Take the hits straight to an export. The segments a search returns are the same objects the export cuts from, so a found moment becomes a clip without a second search and without a charge.
Link straight to a search. The video page reads a q parameter from the URL, so /video?mode=analyze&q=the%20moment%20she%20turns opens the workspace with that query seeded. It fills the box; it does not press the button.

The pattern to take from all of this: analysing is a capital expense and searching is an operating one. Analyse the video once, well, and then interrogate it as often as you like.

Do I have to re-analyse the video for every search?

No. The pass runs once per video and every later search reads the cached index. Repeat queries on the same source skip the analysis entirely, which is why the first question is slow and the rest are not.

Can I search a video I analysed last month?

Yes. The record persists with its segments, transcript, frame notes and saved asks. Reopening one and clicking a saved ask restores the answer with no charge at all.

Does this work on the free plan?

No. Analyze needs Starter and is billed in gold credits, which a free account is granted none of. There is no green rung on this lane.

Can I search a YouTube link directly?

You can import from a link, and it becomes a normal source you analyse and then search. Twelve platforms are accepted and the same per-file cap applies as to an upload.

How precise are the timecodes?

Segment-level for caption matches, and word-aligned for anything found in the transcript. Picked segments export as exact cuts from the source, so what you saw in the card is what lands in the file.