How does natural-language video search work?
Traditional search needs labels. Someone has to tag footage, or the system relies on motion events and time. AI video search works differently. A model converts frames or short segments into a list of numbers called an embedding, which captures what the image contains. A second model, trained together with the first, converts text into numbers in the same space. When you type a description, the system finds the images whose numbers sit closest to the text.
This is why you can search for things nobody labeled in advance. The model learned broad associations between pictures and words during training, so a phrase like a white van by the loading dock can match footage that never had those words attached. Results are ranked by similarity, not by certainty, so the top result is the best guess, and you should expect to scan several. This flexibility is the main advantage over fixed rule libraries, which only know the situations someone configured in advance.
What is search by example image?
Sometimes words are the wrong tool. If you have a photograph of a vehicle, a bag or a person's clothing, you can supply it as the query. The system converts the example into an embedding and finds recorded moments that look similar. This works because the same shared space holds all images, so an image can be compared to another image as easily as to a text description. Searching with an image also avoids the ambiguity of language, since a picture of a specific object conveys details a short phrase cannot.
Example search is useful when a witness, a report or earlier footage gives you something visual to match. Quality matters: a clear example with the object filling the frame works better than a blurry crop. Expect similar-looking objects as well as the target, such as other dark sedans, so combine the example with time ranges and camera selection. As always, a visual match is a lead to verify, not proof of identity.
How do color and attribute filters narrow results?
Alongside meaning-based search, many systems record structured attributes when they analyze video. Object detection labels each item by class, such as person or vehicle, and a separate step may note color, for instance the dominant color of a vehicle or clothing. These become filters. Instead of ranking everything by similarity, you can say show me vehicles, red, between these times, and receive a clean list. Together these attributes give investigators a dependable first cut before any fuzzy ranking begins.
Filters are fast and predictable, while language search is flexible and fuzzy. Used together they are powerful: filter by class and time to reduce thousands of moments to dozens, then use a text description to rank them. Be aware that color depends on lighting. Under sodium lamps, infrared night mode, or strong shadow, colors shift or disappear, so a color filter may miss true matches at night. Test your own site at night before trusting any color-based result from that period.
Can a saved search become an alert?
Searching the past is useful, but the same capability can watch the future. Some platforms let you save a search and run it continuously against incoming video, notifying you when something matches. This catches situations the standard rules never anticipated. A rule library covers common cases such as line crossing or loitering; a saved description can cover the unusual one, like a ladder leaning against a wall. That makes it a strong way to cover unusual risks that are specific to one location, such as a particular tool or vehicle type.
The caution is noise. A vague description produces many weak matches, and an alert that fires constantly is quickly ignored. Start narrow, limit it to relevant cameras and hours, and review what it catches for a few days before relying on it. Treat saved searches as a supplement to detection rules and to human review, not as a guarantee that a particular situation will always be noticed. Remember that the person receiving the alert still has to judge it, so keep volume low enough to stay credible.
What are the limits and how should you search well?
AI search can miss, and it can return false matches. Small or distant objects, poor light, rain, motion blur and unusual angles all reduce quality. Long, complicated descriptions tend to work worse than a few clear concepts. If the first query is weak, rephrase it, add a camera and a time window, and try an example image. Use filters first to shrink the search space, then rank. Keep notes of which phrasings work best on your cameras, because patterns tend to repeat across a site.
Search is a tool for finding footage faster, not for establishing facts. Always watch the actual clip and check the time stamp before drawing conclusions or sharing it. Searching video for people can raise privacy and legal questions, particularly around face or identity features, so set clear policies on who may search and why, keep an audit trail, and confirm requirements with counsel or the local authority. Retaining a record of who searched for what also protects the organization if the use of search is ever questioned.
Key terms
- Embedding
- A list of numbers a model produces to represent the content of an image or text so similar items sit close together.
- Shared space
- A common mathematical space where paired image and text models place their embeddings, letting words be compared directly with pictures.
- Search by example
- Using a sample image as the query so the system finds recorded moments that look similar to it.
- Attribute filter
- A structured narrowing option such as object class or color, applied to detections recorded when video was analyzed.
- Saved search alert
- A stored search run continuously on incoming video that notifies you when a new moment matches the description.
- Similarity ranking
- Ordering results by how close their embeddings are to the query, so the best guess comes first but is not certain.
Search matters most after the fact: a trailer may record a long night of activity, and you need the moment a vehicle arrived. AI VisionStream supports natural-language and visual AI search of video, so a description can find a clip without scrubbing through hours of recording. Edge recording on the unit and two weeks of cloud recording are standard, which sets how far back searches can reach.
Common questions
Can I search video by typing what I am looking for?
Yes, with AI video search. You type a short description, such as a person in a red jacket near the gate, and the system returns recorded moments ranked by how closely they match. You still review the clips yourself.
Is AI video search always accurate?
No. Small, distant or poorly lit subjects can be missed, and similar-looking objects can be returned. Use time and camera filters, rephrase queries and always watch the clip. Treat results as leads to verify, not as proof.
What is the difference between search and a filter?
A filter uses recorded attributes such as object class or color and returns exact matches. Language search ranks moments by meaning and is more flexible but fuzzier. Using both together narrows results quickly and ranks the best matches first.