One box. Type what you remember: something someone said, what was in frame, when it was shot, a camera, a codec, a filename. Media Nexus works out which kind of question it is and answers with the exact file or the exact timecode, and says why each result matched.
An illustration of how the box reads a request, using the same published rules. It searches nothing here; your own archive answers it.
People remember footage by what was said, what it looked like, when it happened, or what it is called. A search that makes them pick a mode first has already lost half of them.
Answered from transcripts, with the line and its timecode.
Answered from what the vision model saw in each frame.
Answered from the date the camera recorded, not the day it was copied.
Answered from filenames, folders, tags and technical metadata.
Each layer answers the part of a request it can prove. The ranking combines them, and the literal layers never wait on a model.
Every filename, folder, tag, camera, lens, project, location, codec, resolution and transcript line is in a full-text index that updates in the same write as the catalogue, so a new clip is findable the moment it is indexed. Words match as prefixes, and a misspelled word is corrected to the closest real term in your own archive.
camera: codec: project: folder: ext: type:-word to excludeA local embedding model turns each clip’s description and every transcript passage into a vector, and turns the request into one too. Requests and passages are embedded with the retrieval instructions the model was trained on, so “the weather pushed the shoot back” finds “we lost two days to the storm” without sharing a word.
Whisper transcribes every clip on your own machine. Lines are joined into short, playable moments, and a spoken request returns the moment itself: the words, the clip, and the start and end timecode, ready to open at the line.
said:"exact words" for a verbatim lineA local vision model labels what it sees in each frame grab, so a clip is findable by its picture even when its name says nothing. For a visual request, the best candidates are shown back to the vision model and checked against the request, including what must not be there. A frame it cannot confirm is never presented as a verified shot.
“Shot in March”, “last week”, “week 12” and after:2025-06 become exact date ranges before any model runs. The date the camera recorded comes first, then the file’s own date, then the day it was indexed, and every result labels which one it used. Leave out the year and it picks the most recent year that actually has footage in that range, and says so.
Keyword scores and similarity scores are on different scales, so adding them rewards whichever one happens to be bigger. Media Nexus fuses the two by position instead.
The literal layer and the meaning layer each rank the archive on their own terms.
A clip near the top of both lists outranks a clip at the top of only one, the standard method for combining search signals without tuning weights.
Name a file that exists and it is the first result, with nothing similar-sounding in the way.
A clip found by meaning alone says “Similar meaning”. Only a real transcript line is quoted, and only a checked frame is called a verified shot.
If a model is offline, the literal results still return and the response says which layer was unavailable.
A request about what was said and what was seen only returns clips that contain both.
Each result carries the reasons it matched, in plain words, so nobody has to open a file to find out whether it is the right one.
“We lost two full days to that storm before the boats could go out.”
Aerial pass along the breakwater at first light.
Day 03, A cam, ProRes 422 HQ.
Example results from a synthetic test library. The reasons are the engine’s own vocabulary.
Every model the search uses runs on hardware you control. There is no per-search bill, no usage meter and no clip uploaded so a cloud service can look at it.
| Part | What does it | Where it runs | Cost per search |
|---|---|---|---|
| Exact words, operators, dates | SQLite full-text index and deterministic rules | Your Media Nexus server | $0 |
| Meaning | Local embedding model | Your Media Nexus server | $0 |
| Transcripts | Whisper, run once per clip | Your Media Nexus server | $0 |
| Frame labels and verification | Local vision model | Your Media Nexus server | $0 |
| Reading an ambiguous request | Small local language model, only when the rules cannot tell | Your Media Nexus server | $0 |
Media Nexus indexes approved storage read-only, and never renames, moves or deletes originals. Transcription and tagging run on hardware the client controls.
We keep a relevance-judged test library: synthetic footage, transcripts and tags, with requests phrased the way people actually ask. A ranking change ships only if it finds more of the right moments.
Measured 18 Oct 2026 on a synthetic test library: 260 judged clips and 90 requests among 1,500 distractors, local models only. Not client media, and not a promise about any particular archive.
No. There is one box. Filenames and short keyword lookups are answered literally without waiting on any model, and the meaning layer is added on top when it helps.
No. Transcription, frame labelling, embeddings and query reading all run on hardware you control. Search costs nothing per query.
The literal layer keeps answering. The response says which layer was unavailable instead of quietly returning a weaker page of results.
Yes, once the clip has been through AI tagging. The vision model labels what it sees in each frame grab, and those labels are searchable even when the filename and folder say nothing.
Meaning-only matches have to clear a similarity floor set for the embedding model in use, and a visual shot has to be confirmed by the vision model. What remains is labelled by how it matched, so a loose match reads as one rather than as an answer.
Tell us about your team and your media. If it is a fit, we build the archive first, on your storage, and you search it before you decide.