How to Search Inside Your Own Videos: Find and Verify a Shot
Search inside your own videos with a clear query, then verify the source and time range. Follow real examples of a match, a wrong candidate, and no results.
To search inside your own videos, choose the right collection, confirm that AI understanding is complete, and describe a visible subject or action. Open each promising result and check the original footage and time range. If nothing useful appears, inspect the search mode, wording, and relevance setting before assuming the footage is missing.
This walkthrough uses an October 20, 2026 check in ShotAI for Mac, version 1.2.0. It shows a match, a plausible but wrong candidate, and an empty result. The small collection contains original synthetic graphics, not customer footage or a representative production library.
1. Prepare Searchable Footage
Start with a few videos you know well and are permitted to process. Keep originals in a stable location, with external drives connected. A saved search result cannot replace a missing source file.
For new material, follow Import and understand footage:
- Choose Import Video or Import Folder and the intended collection.
- Let local preparation and shot detection finish.
- Select the relevant shots and run AI understanding.
- Check completion and searchable-shot counts before searching.
Import and AI understanding are separate stages. Seeing a video in the library does not establish that all its shots are ready for semantic search.
What was tested here: The existing “Product Fit Test 2026-09-11” collection already had two videos and three searchable shots. The source files matched their original recorded checksums. This session did not reimport them or rerun understanding, so it does not measure setup or indexing time.
Review privacy and connectivity before processing restricted material. ShotAI keeps original video files locally, but AI understanding uses derived visual and possible audio data in cloud services. Semantic query text also uses online encoding.
2. Choose a Collection and Describe One Visible Target
Open the collection containing your intended footage. Begin with a concrete subject and condition, such as an object and its color, or a person performing an action. Avoid starting with unrelated archives or a request such as “the best shot.”
The first query in this check was:
A red coffee mug on a table.
At the session's starting relevance threshold of 50%, no result appeared. The threshold was temporarily changed to 45% in settings, and the same query returned one candidate from fixture_01.mp4, labeled potentially relevant at 48%.
This is an example of broadening candidate retrieval after checking a known target. It is not a recommendation to use 45% for every library. The starting 50% was this session's setting, not a claim about every installation's default. A relevance percentage is not the probability that the result is correct.
3. Open the Original and Check the Range
Open the candidate in the source player. Verify the requested object, action, and other essential conditions in the footage rather than accepting its thumbnail or description.
The result displayed a range of approximately 0:00–0:07. During playback and seeking, the source showed the red mug at about three seconds and a blue notebook at about five seconds. Both moments were still inside detected shot #1.
The candidate contained the requested object, but the entire returned range was not a red-mug shot. Record the part you need and finish precise trimming in your editor. A detected boundary is not necessarily your editorial boundary.
The review and export guide explains whole-shot selection and output options. If you export, check source links, picture, sound, frame rate, and duration in the destination. This search check did not perform a new export test.
4. Change One Condition and Reject Wrong Candidates
Keep the collection and settings unchanged while varying one part of the query. This makes it easier to see whether the result satisfies that condition.
The next query changed only the mug's color:
A blue coffee mug on a table.
At the same 45% threshold, ShotAI returned the same source shot as the only candidate, with 47% relevance. The source contained a red mug and, later, a blue notebook. It did not contain the requested blue mug. The candidate should therefore be rejected for this task.
This illustrates why ranking and correctness are different. The first or only result can still fail an essential requirement. The small difference between the displayed 48% and 47% does not provide a reliable acceptance rule.
5. Handle an Empty Result Without Inventing a Match
The third query was:
A purple bicycle in the snow.
The collection contains no bicycle or snow. At the 45% threshold, this query returned zero results. An empty result was appropriate for this known negative example.
For your own footage, check the following before drawing a conclusion:
| Check | Next action |
|---|---|
| Collection | Confirm the source belongs to the searched collection. |
| AI readiness | Check understanding and indexing completion, not only import status. |
| Search mode | Confirm semantic search is active; use exact search when you know a filename or tag. |
| Description | Start with the main subject and action; remove one unnecessary condition at a time. |
| Relevance setting | Broaden cautiously when checking a known target, then review every added candidate. |
| Source access | Keep the original path and drive available for playback or export. |
Do not keep weakening a request until an unrelated result looks acceptable. If the required content is absent, record the gap. The search guide explains scope and query refinement; troubleshooting covers processing and source-access problems.
The original 50% threshold was restored after this check. No customer material, source files, or existing AI processing was changed.
What Counts as a Useful Result?
A useful result is one you can open, verify against the request, and locate within its source. For an edit, you also need a suitable range and a checked handoff. A candidate list or relevance score alone is not completion.
These three queries demonstrate a review method, not general accuracy or speed. For a broader comparison, use the 30-query evaluation guide. To begin with your own material, follow Find your first useful shot.
FAQ
Why can a blue-mug query return a red mug?
Semantic retrieval finds related candidates; it does not certify every requested attribute. Inspect the actual object and reject results that fail a required condition.
Does no result prove the content is absent?
No. Collection scope, processing state, search mode, wording, and relevance settings can affect results. In this example, absence was known independently from the original synthetic material.
Should I use the 45% threshold from this example?
No universal threshold was established. It was a temporary setting for this small check. Test known matches and negatives in your own collection, and restore settings you changed only for diagnosis.
Can this workflow verify an exact spoken sentence?
This session tested visible graphics. It does not establish exact dialogue retrieval. For spoken content, verify the supported search mode and listen to the original passage.
Sources
Published by ShotAI. The October 20 check used version 1.2.0 and two original synthetic videos created on September 11, with no external footage or real customer information. Screenshots show the actual interface, cropped to exclude account details. The source-range image pairs two playback positions. This was not a new indexing run, customer case, or performance benchmark.
CLIP's original paper explains a foundation for matching images and text. It does not identify ShotAI's current model or validate this application's retrieval quality.