Retrieval Engine
Movement Search at Scale
Query "every time a centre-back got beaten 1v1" and retrieve results across entire seasons — zero manual tagging, just raw data.
What it is
Movement search is retrieval, not tagging. Pick a window of movement — a recovery run, an awkward landing, a pressing burst — and retrieve the most similar windows from everything you have indexed, each result linking straight back to its source segment.
Nothing needs to be labeled in advance. Similarity is computed in signature space, so the system finds 'more like this' even for patterns nobody thought to define — the query language is movement itself.
How it works
Every indexed window carries its embedding; a search is a nearest-neighbor query over that index, filtered by whatever metadata you attach — session, athlete, context. Results return with scores and evidence links: window IDs, timestamps, source segments. An analyst can verify every hit, and the same query surface is what the knowledgebase API exposes to AI agents.
What you get
The practical effect: video review starts from candidates instead of from scratch. A pattern seen once becomes a lens over the whole corpus.
| Parameter | Value | Notes |
|---|---|---|
| Query | by example window | Similarity to a selected movement window. |
| Filters | tags, time range, cohort | Narrow results by session, athlete, or context. |
| Returns | windows + scores + evidence | Every hit links back to its source segment. |
| Coverage | what you index | Search spans your indexed corpus — nothing more is implied. |
Status & roadmap
Prototype. Retrieval over embeddings runs today in the internal evaluation workbench — retrieval galleries over a proof-of-concept corpus. The product query surface, filters, and evidence contract are specified in the API docs and labeled by status there.