Cohere's Compass Adds Timestamped Video Retrieval for AI Agents
Cohere's Compass now retrieves timestamped video segments with transcripts, ranking first across 11 datasets with an average nDCG@10 score of 0.69 in its own evaluation.

Priya Raman
· 3 min read
Cohere launched Video Intelligence in its Compass enterprise search platform on November 16, 2026, claiming a 4.4-point average lead over the strongest alternative in its video-ranking evaluation. The capability lets AI agents retrieve timestamped video segments, transcripts and visual descriptions alongside documents — a change that touches how organizations search meeting recordings, training footage and compliance evidence.
Compass ranked first overall across 11 datasets with an average score of 0.69 on nDCG@10, a metric that measures how accurately the top 10 search results are ranked. The evaluation ran 11,133 queries against seven competing video embedding models, using six public benchmarks and five internal enterprise video datasets.
What does the new capability actually do?
Compass searches both spoken content and visible on-screen activity. A query returns a bounded segment chosen at search time, with start and end points based on the request rather than a library of pre-cut clips. Each result carries a link to the original footage, a timestamp, a transcript and a visual description, so a reviewer can verify the evidence directly.
Video sits in the same index as documents. A single result set can pair a recording segment with a related page, slide or spreadsheet, and permission-aware retrieval limits each user to content they're authorized to see. Organizations can configure fields, choose models and select deployment options, Cohere said.
A Compass MCP server connects agents to the system, letting them analyze slides, charts and demonstrations in addition to spoken material. The design keeps a route back to the source footage — the feature Cohere positions as its central proposition: a shared evidence trail across video and documents.
Which HR and workplace workloads does it touch?
The intended applications connect retrieval to review work, and several map directly onto people-process decisions:
- Meeting assistants could surface decisions, forecasts and commitments from recorded meetings, earnings calls or analyst briefings.
- Employees could locate a particular demonstration or procedure inside training material.
- Legal and compliance teams could find statements and disclosures with supporting visual context.
- Sports and media teams could search by play, player, camera angle or visible brand for highlights, licensing and sponsor reporting.
For HR teams sitting on growing libraries of recorded onboarding sessions, safety training and all-hands meetings, the capability question is whether timestamped retrieval actually shortens the path from question to verified footage — not the benchmark score itself.
How big are the reported gains?
The strongest alternative varied by domain, which buyers should weigh against their own content mix. Cohere reported:
- A 4.8-point lead over Amazon Nova 2, averaged across five open-source public benchmarks.
- A 5.1-point lead over Gemini Embedding 2 on soccer broadcasts.
- A 5.0-point lead over the same model on MultiVENT 2.0 news retrieval.
- A 2.0-point lead over Qwen3-VL Embedding 8B on business meetings, lectures and workplace-safety training — the domain closest to typical corporate video collections, and the narrowest margin in the set.
That 2.0-point lead on workplace content matters most for HR-adjacent buyers, since it's the slimmest reported advantage and the domain they'd actually search.
Who gets access, and on what terms?
Cohere is recruiting design partners rather than shipping general availability. Partners can deploy Video Intelligence in Cohere's managed cloud or their own environment, evaluate it against their own footage, queries and use cases, and work directly with Cohere's AI search team. Participants can also influence future capabilities based on their organizational needs.
The vendor-announcement math is straightforward: the benchmark numbers come from Cohere's own evaluation, including five internal datasets no outsider can inspect. The design-partner program effectively makes customer-owned footage and permissioned workloads the real test of whether the reported ranking gains translate into retrieval people will use — and whether the 2.0-point margin on meetings and training holds up outside Cohere's own data.
For people-operations leaders, the near-term decision isn't purchase but participation: whether joining the design-partner cohort is worth the evaluation effort to shape how agents will cite video evidence in meeting summaries, training lookup and compliance reviews going forward.
Source: unite.ai
More from Priya Raman
Show full bio
Staff writer covering companies, deals and strategy in Recruiting and talent acquisition, learning and development at HR Briefing.
273 articles