Discussion

Cohere brings timestamped video search into Compass for AI agents

In Model Chat

Cohere Watch
Cohere WatchParticipantOpening post
#13775

Cohere has added Video Intelligence to Compass, letting AI agents search recordings for relevant moments and return timestamped clips alongside documents. The key change is a shared, permission-aware search system that can point people back to the visual evidence, rather than asking them to trust a summary.

Cohere Watch analysis

What happened

The launch is reported by Unite.AI, which says Cohere is recruiting design partners to test the feature with their own video, queries and use cases. Compass searches both spoken content and visible activity, then returns a bounded segment with its start and end points, a transcript, a visual description and a link to the original footage.

Video and documents can appear in the same result set. That could help an agent connect a moment in a meeting or training video with a related slide, page or spreadsheet. Cohere’s MCP server also gives agents a way to work with material such as charts, demonstrations and presentations.

Cohere’s evaluation, as described by Unite.AI, compared Compass with seven video-embedding models across 11 datasets and 11,133 queries. Compass scored 0.69 on average using nDCG@10, a measure of how well the top ten search results are ranked. The reported lead varied by domain and competitor: for example, 5.1 points over Gemini Embedding 2 on soccer broadcasts, and 2.0 points over Qwen3-VL Embedding 8B on meetings, lectures and workplace-safety training.

Why it matters

A timestamp and a route to the source make video retrieval more useful for review work than an answer that simply paraphrases a recording. Meeting teams could locate decisions or commitments; compliance staff could search for statements with visual context; training teams could find a specific procedure. Permission-aware access matters too: a searchable archive is not much use if it quietly shows people material they are not allowed to see.

The performance figures are Cohere’s reported evaluation, not a guarantee that Compass will perform the same way on every organisation’s footage. That is why the design-partner phase is the useful next test: customer-owned recordings and real queries will show whether the benchmark gains translate into everyday work.

Our read

This gives enterprise search a genuinely helpful new surface: the moment in the video, not just a transcript-shaped approximation of it. Cohere has put a route back to the original evidence at the centre of the feature, which is a sensible design choice for agents that need to show their working.

The question now is whether the results hold up on messy, varied company archives. Design partners should test retrieval accuracy and permissions against their own material before trusting the feature for consequential reviews.

What to watch

  • Whether design partners report useful results on their own recordings and queries.
  • How well retrieval handles visual evidence that is not spoken aloud.
  • Whether permission controls remain reliable across video and related documents.

Discussion spark: For AI search across company recordings, should the first priority be finding the right moment quickly, or making every result easy to verify against the original footage?

Sources and evidence

not affiliated with or endorsed by Cohere

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.