Latest in AI
Cohere Video Intelligence: How to Evaluate Timestamped Video Search
Compass can retrieve moments inside recordings, but a useful pilot needs to measure more than whether the right video appears.
Cohere introduced Video Intelligence in Compass on November 16, letting applications retrieve timestamped video segments with transcripts and visual descriptions alongside documents. The company is recruiting design partners. For a team evaluating it, the useful question is whether the returned clip contains enough evidence to answer the question without watching the recording again.
The announcement describes clip boundaries chosen at query time and retrieval that respects user permissions. Cohere reports an average 4.4-point nDCG@10 lead over the strongest alternative across 11 datasets. That is a company-run ranking evaluation, not a measurement of how accurately an agent will answer your team's questions.
Check access before planning an integration
Cohere invites partners to evaluate their own footage in its managed cloud or their own environment. The launch page does not publish video-specific pricing, ingestion limits or a general-availability date. Request those details before estimating a production rollout; a normal API account should not be treated as proof of access.
The broader Compass product page describes document-level security, role-based access and integration through Compass MCP. Those platform capabilities help frame a pilot, but they do not settle how a particular video connector propagates access changes. Ask which identity a search uses and how a revoked permission reaches the index and any cached results.
Build one recording that can expose several failures
Here is a proposed pilot, not a test I have run against Compass. Record a short fictional product meeting with information that changes during the conversation. At minute two, someone proposes releasing on Friday. At minute six, the team rejects Friday and agrees on Monday. Put the final date on a slide without reading every word aloud.
Write the answer key before uploading the recording. Include the agreed date, who approved it, the relevant time interval and which details are spoken versus visible. This avoids letting the system's first answer become the definition of success.
Ask, “When did we agree to release?” A result containing only the early proposal should fail, even if its transcript contains the right project name. Then ask, “What date is shown on the final slide?” That separates visual retrieval from a system that succeeds only when the answer appears in speech. Use synthetic names and harmless content so the first evaluation does not require exposing a real internal meeting.
Score the clip and the answer separately
Record whether a relevant segment appears near the top of the results, whether its boundaries include the complete decision, and whether the final answer accurately reflects that segment. These are separate failures with different remedies. A generator cannot reliably repair missing evidence, while an excellent retrieved clip can still be misread.
For the fictional meeting, a useful result should include enough conversation to distinguish a proposal from an approval. A clip starting after “we rejected Friday” might leave a misleading fragment. Save the returned interval and inspect the original footage rather than scoring only the generated description.
Also ask a question the recording cannot answer, such as the release budget when no budget was discussed. Count an unsupported confident answer as a failure. Add a conflicting document with an older date and ask the system to identify the disagreement, rather than quietly choosing whichever source it retrieved first.
Test permission changes with harmless fixtures
Create two test users with different access to the fictional recording. Repeat the same query as both users, then revoke the authorized user's access and repeat it after the documented propagation interval. Check the search result, transcript, generated answer and original-file link. A hidden video with a visible transcript is still an exposure.
Ask the provider how deletion affects derived text, embeddings and cached clips, and include its stated behavior in the acceptance criteria. Do not infer deletion guarantees from a general security description. For a production decision, measure ingestion time and cost on representative recording lengths as well as query latency.
A pilot should end with a small failure log: wrong segment, incomplete interval, unsupported answer, stale permission or unacceptable cost. That gives the engineering team something specific to improve and tells you whether video search saves review time on the recordings you actually need.
I'm Paras. I write about AI systems and the engineering decisions behind them.
I help technical teams turn complex product capabilities into clear explanations and useful evaluation guides.