Enterprises have vast amounts of valuable information locked inside recorded meetings, training videos and media archives. But unlike documents, video remains difficult for AI applications to search and use.
Today, we're introducing Video Intelligence in Compass, enabling AI applications and agents to search across video collections, find relevant moments and retrieve timestamped evidence alongside documents.
Compass understands both what's said and what's shown on screen. Instead of returning an hour-long recording, it retrieves the exact moment that answers a question, with a timestamped clip, transcript and visual description.
We're now recruiting design partners. Register your interest.
Search for moments, not just files
Consider the questions people actually ask of video:
- When did the team agree to move the launch to Q3?
- Where does the presenter explain revenue trends on the whiteboard?
- When did a particular player get sent off during the game?
Answering them takes more than matching keywords in a transcript or scanning individual frames. It takes an understanding of what's said, what's shown and how events unfold over time. A single frame can show a player leaving the pitch, but only the sequence shows they were sent off. A presenter can describe improving KPIs while the chart on screen tells a different story.
Timing matters too. A clip that begins two minutes early, or cuts off before the answer, still leaves someone scrubbing through the footage.
Compass makes video searchable by meaning, context and precise moments in time.
From raw footage to discoverable intelligence
A single search returns the exact moment in a recording alongside the matching page, slide and sheet from related documents, all drawn from one index that respects what each user is allowed to see.
Finding the right moment in a video is only part of the challenge. AI applications also need to retrieve that information, verify its source and use it alongside other enterprise knowledge.
Compass brings video into the same retrieval system as documents, with the precision and controls needed for enterprise AI applications.
- Precise moments, not predefined clips: Rather than relying on fixed video segments, Compass determines the relevant start and end times at search time, returning the moment that answers the question.
- Video and documents in one search: Applications and AI agents can retrieve relevant footage alongside documents, bringing video into enterprise search, RAG and agentic workflows.
- Evidence you can verify: Every result links back to the original footage, with timestamps, transcripts and visual descriptions so users can check the source.
- Built for enterprise requirements: Permission-aware retrieval ensures users only access authorized content, while configurable fields, model choices and deployment options let organizations adapt the solution to their needs.
Connect your agent to the Compass MCP server so it can reason over what's shown on screen, such as slides, charts and demonstrations, not just what's said.
Unlock new applications for your video data
Compass makes video accessible to AI applications, including retrieval-augmented generation (RAG) and agentic search. By making spoken and visual content searchable alongside documents, Compass gives your applications and AI agents eyes and ears to reason over new modalities and use information that was previously difficult to access.
Meetings and enterprise knowledge: Build AI assistants that surface decisions, forecasts and commitments from recorded meetings, research presentations, earnings calls and analyst briefings. Find exactly when something was discussed and retrieve the relevant clip, rather than searching through hours of recordings.
Education and training: Help employees and learners find answers across recorded lectures, onboarding sessions and training libraries. Retrieve the exact explanation, demonstration or procedure relevant to a question, making educational content easier to discover, reference and reuse.
Sports and media archives: Search large video libraries for specific plays, players, events, camera angles or visible brands. Accelerate highlight creation, footage discovery, content licensing and sponsor reporting without manually reviewing recordings.
Legal and compliance: Find statements, disclosures and other relevant moments across recorded footage. Retrieve timestamped clips with supporting transcripts and visual context, enabling reviewers to locate and verify information against the original source.
Video Intelligence turns recorded meetings into a searchable record of decisions, returning just the minute where something was agreed so anyone can check it against the source.
Strong retrieval performance across enterprise video
Good moment retrieval means finding the most relevant moments and identifying precisely where they start and end. We evaluated Compass against seven video embedding models across 11 datasets, including six public benchmarks and five internal enterprise video datasets.
Averaged across all 11 datasets, Compass ranks first, with an nDCG@10 4.4 points ahead of the strongest alternative. The lead holds up across all three domains, even as the strongest competitor rotates from one domain to the next:
- Public benchmarks: 2.4 points ahead of Amazon Nova Multimodal Embeddings, averaged across six widely used open-source moment and video retrieval benchmarks. These include the English subset of MultiVENT 2.0, which tests retrieval of real-world events by the people, places, and dates involved.
- Sports: 5.1 points ahead of Gemini Embedding 2 on soccer broadcast footage, where queries look for specific players, teams, and plays.
- Enterprise knowledge: 2.0 points ahead of Qwen3-VL-Embedding-8B on business meetings, lectures, and workplace-safety training.
Results are measured using nDCG@10, which evaluates how accurately the top ten search results are ranked.
Cohere Video Intelligence ranks first across our evaluation suite. Scores are nDCG@10 over 11,133 queries across 11 datasets, including six open-source public benchmarks (10,591 queries), enterprise knowledge (meetings, lectures and safety training; 85 queries) and sports (two soccer broadcast datasets; 247 queries).
Build with Video Intelligence
We're inviting design partners to test and shape Video Intelligence using real-world workloads.
Partners can deploy in our managed cloud or their own environment, evaluate performance against their own footage and queries, and collaborate directly with our engineering team to inform future development. Register your interest below.












