Search your library

Semantic search finds shots related to a description of their content. Start with what should be visible: the subject, action, setting, or camera view. Use exact search when you already know a filename or tag. Both approaches work best when the collection and processing state are clear.

Start with a visible description

1. Open the intended collection or check the collection filter.
2. Confirm that its shots have completed AI understanding.
3. Enter a concrete description in the search field.
4. Open promising results and inspect the source footage.

A useful first query is “someone watering plants in a garden.” You can refine it to “a close-up of hands watering plants” if the first results are too broad. Chinese descriptions are also supported in the verified workflow; you do not need to translate every query into the language used by the generated tags.

Describe requirements you can check on screen. “Approved campaign footage” contains a business decision that visual similarity cannot establish. Search for the visible content first, then check approval in the system where that decision is recorded.

Narrow or broaden the results

The Filters interface includes collection, duration, and keyword controls. Use the collection to preserve project scope and duration to reduce candidates that are clearly unsuitable for the intended use. Check the displayed labels rather than assuming a date filter refers to the original filming date.

The search relevance setting affects how loosely related results can be shown. A stricter setting can reduce irrelevant candidates; a broader setting can reveal alternatives. Change one constraint at a time so that you can tell why the result list changed.

If you ask for several conditions in one sentence, inspect each condition separately in the returned shot. A high-ranked result may match the subject while missing the requested action or camera movement.

Read relevance as a ranking signal

Relevance indicates similarity to the request. It is not a probability that every statement in the query is true. Open the shot, look at the actual picture, and confirm the source and range before adding it to your selection.

Returned ranges follow detected shot boundaries. A matching object may appear in only part of that range. AI descriptions can also omit details or get an attribute wrong; they are aids to discovery, not substitutes for review.

When there are no useful matches

Confirm AI completion and the collection first. Then shorten the query to its main subject and action, remove unnecessary details, or broaden relevance. If you know the filename or an existing tag, switch to exact search.

Semantic search requires the online services used for query processing. When unavailable, the app may fall back to exact search; check the displayed search state before interpreting the results.

For alternate candidates based on a known shot, see Find similar shots. Once you have useful candidates, follow Review and export to prepare an output and check it in the destination application.