Cohere taught Compass to find the right minute in a video: top-10 results 4.4 points more accurate
Recorded calls sit unused: finding when the team agreed to push back a release means scrubbing through an hour of video. Cohere has taught Compass to understand speech and the screen, and to return the exact minute with a timestamp. Across 11 datasets, it outperformed seven competing models, placing relevant clips 4.4 points higher in the top 10 results (nDCG@10). For RAG and agents, video is now searched like any other document.
AI-processed from Cohere Blog; edited by Hamidun News
Imagine you need to find out when the team agreed to move the release to the third quarter. The document turns up in seconds, but the recording of the call means scrubbing through an hour of video. That is why almost nobody goes back over archives of meetings, lectures, and webinars.
Cohere is tackling this problem in Compass, its enterprise search product. The new Video Intelligence feature returns not the whole recording but the exact minute where the answer was said, with a timestamp, a speech transcript, and a description of the visuals. The headline's "top-10 results 4.
4 points more accurate" means the following. Cohere measured nDCG@10, a metric that shows how high the genuinely relevant moments rank within the top ten results. On average, Compass came out 4.
4 points ahead of the strongest of the seven alternative models.
What Compass can do
The core idea is to search for moments, not files. Instead of a one-hour recording, the system returns a short segment that answers the question.
A transcript alone is not enough for this. A speaker may say that metrics are rising while the slide shows a chart going down. A single frame of a footballer leaving the pitch does not tell you he was sent off; that only becomes clear from the sequence of events. Compass takes into account words, images, and how events unfold over time.
There is also a technical detail. Many systems split video into fixed-length chunks in advance. A clip can then start two minutes before the relevant part or end before the answer, and the person still has to scrub through it. In Compass, the start and end of a segment are calculated at query time, for the specific question asked.
Video alongside documents
Video does not live in a separate store. A single query returns both the fragment of the recording and the relevant pages, slides, and tables from linked documents. All of it comes from one unified index that respects access rights: an employee sees only what they are cleared to see.
For developers, two practical points stand out:
- Video can be plugged into an existing RAG pipeline and agentic workflows without building a separate system;
- An agent can connect through Compass's MCP server and reason about what is shown on screen (slides, charts, demonstrations), not only about what was said.
Every result links back to the source recording, so the answer can be verified. For lawyers and compliance teams this matters: instead of trusting a summary, you can open the minute where something was said. Cohere also names other use cases: assistants for meeting recordings, presentations, earnings calls, and analyst briefings; finding answers in training libraries; and locating players, episodes, and brands in sports and media archives.
What the benchmarks showed
Cohere compared Compass with seven video embedding models across 11 datasets and 11,133 queries. The set included six open benchmarks, three enterprise datasets (meetings, lectures, and safety training), and two datasets of football broadcasts. The leading competitor changes from domain to domain, but Compass comes out ahead in each:
- Open benchmarks: a 2.4-point lead over Amazon Nova Multimodal Embeddings;
- Sports: a 5.1-point lead over Gemini Embedding 2;
- Enterprise knowledge: a 2.0-point lead over Qwen3-VL-Embedding-8B.
Limitations
The figures come from Cohere itself, some of the datasets are internal, and there has been no independent verification yet. A 2–5 point gap is noticeable, but it is not a chasm. On your own data the result may differ, especially if the recordings are noisy and the screen carries little information.
The product is not generally available yet: the company is recruiting design partners. They will be able to test Compass on their own recordings and queries, deploy it in Cohere's managed cloud or in their own environment, and influence the roadmap. No pricing or general launch timeline has been announced.
The practical takeaway: if your company holds hundreds of hours of recordings that nobody revisits, the most useful step right now is to assemble 50–100 real questions for them. With that set, you can honestly compare Compass against any alternative instead of relying on presentations.
- Nov 21, 2026A viral GTA 6 Claude clip looks fake: half an hour of control burned through the 5-hour Pro limit
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
- Nov 21, 2026Waymo pulls robotaxis off highways after new entries into work zones
- Nov 21, 2026Google Opens Pre-Orders for Gemini Smart Glasses from $395, Shipping December 10
Key Questions Answered
What Compass can do?
The core idea is to search for moments, not files. Instead of a one-hour recording, the system returns a short segment that answers the question. A transcript alone is not enough for this. A speaker may say that metrics are rising while the slide shows a chart going down. A single frame of a footballer leaving the pitch does not tell you he was sent off; that only becomes clear from the sequence of events. Compass takes into account words, images, and how events unfold over time. There is also a technical detail.
What are the key details on Video alongside documents?
Video does not live in a separate store. A single query returns both the fragment of the recording and the relevant pages, slides, and tables from linked documents. All of it comes from one unified index that respects access rights: an employee sees only what they are cleared to see.
What the benchmarks showed?
Cohere compared Compass with seven video embedding models across 11 datasets and 11,133 queries. The set included six open benchmarks, three enterprise datasets (meetings, lectures, and safety training), and two datasets of football broadcasts.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.