Skip to main content
AI Search finds useful moments by meaning rather than by filename alone. Describe a scene, subject, action, spoken idea, or sound; the editor searches media available in the active workspace and returns candidates you can preview, add at the playhead, or drag to a precise destination. For the walkthrough below, search for:
AI Search panel in the video editor showing six landscape and tea-estate result cards

AI Search showing six landscape and tea-estate matches

​
Find and add a moment

1

Open AI Search

Select AI Search in the left rail. The search field receives focus so you can type without first clicking into it.
2

Describe the desired content

Enter misty landscapes, tea estates. Use visible or audible details that could identify the source moment; you do not need to know its filename.
3

Choose the search intent

Keep All for a broad first pass, or select Visual, Dialogue, or Audio before submitting. The filter changes the request sent to search, not merely which completed cards are hidden.
4

Run the query

Press Enter or select Search. The panel shows a searching state, then reports the result count and active filter above the grid.
5

Inspect a promising card

Select a thumbnail or focus the card and press Enter or Space to open its full preview. Hovering or focusing a card also reveals Add to timeline and Open preview controls.
6

Place the result

Use Add range or Add asset in the preview, press A, select the card’s add control, or drag the card to the canvas or timeline. Confirm the new layer at its intended time before continuing.

​
Panel keyboard controls

​
Choose the right filter

Filters participate in retrieval and ranking. If a query has already completed, changing the filter reruns the same query with the new intent. Begin with All to discover the available result mix. Switch to Visual when you want the strongest emphasis on terrain, mist, plantation rows, workers, or time of day.

​
Read the result grid

The compact two-column grid is designed for fast scanning. Each card contains an aspect-ratio-aware thumbnail and a title of up to two lines. The panel intentionally avoids dense badges so the imagery and title remain the primary comparison signals. The header above the grid communicates two independent facts:
  • The left side shows Searching while a request is pending, then the number of results.
  • The right side shows the active filter, such as All or Visual.
The grid can show up to 24 candidates at a time. When the search service reports a larger total, the header can represent all matches rather than only the cards currently visible.

​
Card interactions

Video results can expose more than a static poster. Moving across an eligible thumbnail scrubs representative frames. Holding the pointer still briefly can start a muted hover preview that continues until the pointer leaves. Reduced-motion preferences prevent the delayed playback transition.
Hover to compare composition and action, then open the full preview for timing decisions. A thumbnail is useful for triage, but the preview is the authoritative place to judge a source range.

​
Preview and refine a result

The full-screen preview does not add anything by itself. It gives the selected result enough space for playback, range inspection, navigation, and an intentional add action. The preview header shows:
  • The result’s position within the visible result set, such as 2 of 6.
  • The source moment range when a valid range exists.
  • The result title.
  • Add range for a timed segment or Add asset when there is no editable result range.
  • Info for the result inspector and Close for returning to the grid.
Use the left and right controls to compare adjacent results without closing the preview. For timed video or audio, the trim strip at the bottom represents the source interval. Move its handles to keep only the useful portion; the preview and final insertion use the selected interval. The lightbox enforces a valid positive range, clamps it to known media duration, and prevents an inverted selection.

​
Preview keyboard controls

​
Add versus drag

All successful placement paths confirm that the selected result maps to supported workspace media before adding a layer. Choose the interaction based on how precisely you already know the destination.

​
Source time and timeline time are different

A search result can identify a useful interval inside a longer source. That interval is the source range. The location where you add or drop it is the composition placement. For example, a result may represent source seconds 12.0–18.0. Dropping it at 00:15.00 on the timeline keeps the 12–18 second source trim but places that trimmed segment at the 15-second composition position. The source timestamps never become timeline coordinates.

​
Search state model

Use the panel copy—not the presence or absence of thumbnails alone—to determine what happened.
A visible result is not automatically insertable. Preview can work even when the underlying workspace media is temporarily unavailable or incompatible with the editor. If preview works but insertion fails, follow the diagnostic guide rather than rerunning the search repeatedly.

​
Write stronger queries

Semantic search works best when the description names evidence that could appear in a frame, transcript, or audio segment.

​
Build from observable concepts

Combine two or more of these elements:
  • Subject: tea workers, winding road, tea leaves, mountain ridge.
  • Setting: plantation, valley, hillside, mountain region.
  • Condition: misty, fog-covered, rain-soaked, sunrise.
  • Action: walking between rows, harvesting, fog moving across hills.
  • Framing: aerial landscape, macro detail, wide establishing shot.
The walkthrough query—misty landscapes, tea estates—works because it combines atmosphere, scene scale, and a recognizable environment without depending on a filename.

​
Refine one axis at a time

If the first pass is too broad:
  1. Keep the same subject and add an action or framing detail.
  2. Switch from All to Visual without rewriting the query, so you can compare the effect of the filter.
  3. Add a time-of-day or weather cue.
  4. Remove adjectives that are not observable in the media.
If the first pass returns nothing, reverse that process: remove the least essential qualifier, return to All, and retry once. This makes each change interpretable and avoids turning a useful semantic query into an overconstrained sentence.

​
Verify the insertion

After an add or drop completes, check all four layers of correctness:
  1. Identity: the new layer depicts or contains the intended result.
  2. Placement: its start frame and track match the add or drop gesture.
  3. Source range: playback begins and ends on the expected portion of the original resource.
  4. Presentation: visual dimensions, audio capability, and duration are sensible for the edit.
If the grid never appears, the result count disagrees with the panel state, or a card previews but cannot be inserted, continue with AI Search diagnostics.