Slop TVNewsLatest
Explainers

Cohere's Compass searches video by speech and on-screen content

The enterprise search product returns a timestamped clip with a transcript and a visual description, and Cohere is recruiting design partners rather than shipping it to everyone.

Illustration: Cohere's Compass searches video by speech and on-screen content
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Cohere introduced Video Intelligence in Compass on November 16, 2026, in private preview, and is recruiting design partners rather than offering general availability.
  • A query returns a clip with the relevant start and end times, a link to the original footage, a timestamp, a transcript and a description of what appeared on screen.
  • Cohere says Compass led seven video embedding models across 11 datasets by an average of 4.4 points on nDCG@10; the company ran the evaluation itself.
  • The enterprise-knowledge group of that evaluation held 85 of the 11,133 queries, the smallest of the three domains Cohere measured.

Cohere's Compass can now find a specific moment inside a video, by what was said in it and by what appeared on screen. The company introduced the capability, called Video Intelligence, in a blog post on November 16, 2026, and it is in private preview: Cohere is recruiting design partners to test it, and the company has announced no general availability.

The problem it targets is the one anyone who keeps a footage library already knows. A company records its meetings, training sessions and media archives, then struggles to use much of it, because a video is harder for software to search than a document. Cohere's pitch is that the archive stops depending on someone's memory of where a thing was said or shown.

What Cohere's Video Intelligence does

Compass understands both what is said and what is shown on screen. Cohere's own example is a presenter who describes improving numbers while the chart behind them tells a different story; a search that reads only the transcript misses the chart.

Rather than hand back an hour-long recording, Compass returns the moment that answers the question. Each result carries a link to the original footage, a timestamp for the segment, a transcript of what was said and a description of what appeared. Cohere's stated reason is verification, so a reviewer can check a result against the source instead of trusting a summary.

How Compass picks the moment

The segment boundaries are chosen when the query arrives, not in advance. Cohere says Compass determines the relevant start and end times at search time, rather than serving results from a library of fixed clips. The company gives the reason in one line: a clip that begins two minutes early, or that cuts off before the answer, still leaves someone scrubbing through the footage.

Cohere also argues that some questions turn on the order of events, not a single frame. A frame can show a soccer player leaving the pitch; only the sequence shows that the player was sent off. The same gap runs through a presentation, where the useful evidence can be a change on screen rather than a sentence spoken.

Video and documents share one index

Video is indexed and retrieved inside the same Compass system that handles documents. A single search can return a recording segment next to the matching page, slide or spreadsheet sheet, and permission-aware retrieval limits each user to the content they are allowed to see.

That permission layer is the unglamorous part that decides whether an enterprise can use this at all. Compass applies to video the same access rules documents already live under, and organizations can configure fields, choose models and pick a deployment option.

An MCP server for agents

Cohere points to its Compass MCP server as the route for AI agents. Through that connection, an agent can reach the retrieval tools and reason over slides, charts and demonstrations as well as speech. The intended result is an agent that answers with a citation back to the exact moment, rather than a claim with no source.

The uses Cohere sketches run from meeting assistants that surface decisions and commitments, to training libraries that answer a procedure question, to sports and media teams searching by play, player, camera angle or visible brand, to legal and compliance teams locating statements and disclosures with the visual context attached.

The benchmark claim, and its caveat

Cohere evaluated Compass against seven video embedding models across 11 datasets, made up of six public benchmarks and five internal enterprise video datasets. Averaged across all 11, Cohere reports Compass in first place, 4.4 points ahead of the strongest alternative on nDCG@10, a measure of how well the top ten results are ranked. The company's table puts the average score at 0.69 across 11,133 queries.

The lead is not the same size everywhere, and the strongest rival changes with the material. Cohere reports 2.4 points over Amazon Nova Multimodal Embeddings across six public benchmarks, 5.1 points over Gemini Embedding 2 on soccer broadcasts, and 2.0 points over Qwen3-VL-Embedding-8B on business meetings, lectures and workplace-safety training.

That last figure is the one to read twice. Enterprise knowledge is the domain Cohere is selling into, and it is where the reported margin is narrowest. The evaluation's own query breakdown is uneven too: Superpower Daily reports that the enterprise-knowledge group held 85 of the 11,133 queries, the smallest of the three domains by a wide margin.

Two things follow from that. The evaluation is company-reported, and no outside lab has reproduced it. An 85-query slice is thin ground for a claim about how the system behaves on a company's own messy footage, which is part of why Cohere is asking design partners to run it on their archives rather than promising a number.

How it compares with ordinary video search

Most video search runs on metadata and a transcript. It matches a word in the title, the description or the captions, and it finds files. A transcript tool does the same job a little better, searching what people said and stopping there. Neither can answer where a chart changed or when a demonstration started.

Video Intelligence adds the visual track to the text one. It searches speech and on-screen content together, picks the boundaries at query time, and returns a timestamped, checkable result rather than a file to open. For someone cutting video, that is the difference between scrolling a timeline and jumping straight to the take, provided the tool can tell your take from the other 40.

What is in private preview, and what is not

Video Intelligence is not a product you can sign up for today. Access runs through a design-partner program: partners can deploy it in Cohere's managed cloud or their own environment, test it against their own footage and queries, and work with Cohere's engineering team. Cohere has published no price, no general-availability date and no model names for the retrieval stack.

The capability sits inside Compass, Cohere's enterprise search and discovery product, which already handled documents. Video Intelligence is the extension of that system to recorded meetings, training libraries, broadcasts and media archives. Until the design partners report back, the benchmark table is the strongest public evidence for it, and that table is Cohere's own.

Design partners register their interest through Cohere's Video Intelligence post; there is no general-availability date or public price yet.

Sources

  1. cohere.com - Cohere's launch post: the November 16, 2026 date, the capability, the design-partner terms and every evaluation figure
  2. x.com - Cohere's own X announcement, November 16, 2026
  3. cohere.com - what Compass is
  4. unite.ai - independent write-up of the launch
  5. superpowerdaily.com - independent write-up that reports the 85-query enterprise slice
  6. dev48.org - reproduction of the release text with the evaluation's query breakdown