Cohere's Compass Beats Google and Amazon Searching Video by Exact Moment

Cohere adds timestamped video search to Compass, outperforming video-native embeddings from Amazon, Google, and Qwen on an 11-dataset retrieval suite.

·
·
·
Read4 min
TypeNews
  • Cohere launched Video Intelligence in Compass, in private preview for design partners.
  • Returns timestamped clips with transcripts and visual descriptions, not whole recordings.
  • Jointly indexes speech and visuals; agents can query via an MCP server endpoint.
  • Ranks first on average across 11 datasets, 4.4 nDCG@10 points ahead of the next model.
  • Beats Amazon Nova, Gemini Embedding 2, and Qwen3-VL-Embedding-8B in their respective domains.
  • Video sits in the same permission-aware index as documents for mixed-modality RAG.

Cohere has added video retrieval to Compass, its enterprise search platform. The private-preview feature, called Video Intelligence, searches speech and on-screen content, then returns the relevant segment with timestamps, a transcript, and a visual description. Cohere is recruiting design partners to test it on their own footage.

The addition gives developers a way to search recorded meetings, training libraries, and media archives at the moment level. It also places video in the same permission-aware index as documents, allowing search applications and AI agents to retrieve evidence across several formats with one query.

Search lands on the relevant moment

Enterprise video questions often depend on events unfolding over time. A sequence can establish that a player was sent off, while an isolated frame may show only the player leaving the pitch. A presenter may describe improving KPIs while the displayed chart shows a decline. Transcript search misses visual evidence, and frame search misses spoken context and motion.

Compass jointly indexes speech and visuals. At query time, it identifies the relevant start and end points and returns that bounded segment, avoiding fixed-length chunks and full-recording results. Each result links to the source clip, giving reviewers a way to verify the surrounding context for legal, compliance, and research work.

One index spans clips and documents

Video shares a retrieval index with files such as presentations, PDFs, and spreadsheets. A query can therefore return a slide, a table row, and a 45-second clip in the same result set. The integration includes:

  • Permission-aware retrieval: Search results inherit access controls so users see only authorized footage.
  • Model Context Protocol access: An MCP server endpoint lets AI applications retrieve what appears on screen, including slides, charts, and demonstrations, alongside spoken content.
  • Configurable retrieval: Teams can select fields and models for their application.
  • Deployment options: Customers can use Cohere's managed cloud or deploy in their own environment.
Diagram showing Compass searching video and documents through one index
Compass combines video and document retrieval in one permission-aware index.

Cohere reports a benchmark lead

Cohere evaluated its system against seven video embedding models across 11 datasets, combining public benchmarks with internal enterprise corpora. The company reports that Compass ranked first on average, leading the strongest alternative by 4.4 nDCG@10 points. Different competitors produced the closest result in each domain.

Domain Evaluation material Closest model Reported lead
Overall 11 datasets across three domains Varied by domain 4.4 points
Public benchmarks Six open-source moment and video retrieval benchmarks Amazon Nova Multimodal Embeddings 2.4 points
Sports Soccer broadcasts with player, team, and play queries Gemini Embedding 2 5.1 points
Enterprise knowledge Meetings, lectures, and workplace-safety training Qwen3-VL-Embedding-8B 2.0 points

nDCG@10, or normalized discounted cumulative gain at 10, measures the relevance and ordering of the first 10 retrieved results, with higher-ranked results receiving more weight. The metric evaluates retrieval quality; generation accuracy, indexing speed, query latency, and operating cost remain outside its scope.

Because Cohere ran the evaluation and used some internal corpora, the complete comparison cannot yet be reproduced independently from public datasets alone. Customer trials can provide a stronger test by using each organization's footage, access rules, terminology, and query patterns.

Table comparing Cohere video retrieval with seven embedding models across 11 datasets
Cohere's evaluation covers public, sports, and enterprise video retrieval.

Four workflows drive the pitch

  • Meeting intelligence: Retrieve decisions, commitments, and supporting slides from call recordings.
  • Education and training: Locate a specific explanation or demonstration inside a lecture library.
  • Sports and media: Find plays, camera angles, or visible brand logos for editing and sponsor reporting.
  • Legal and compliance review: Search large collections of footage while preserving links to the source evidence.

These workflows currently depend heavily on transcript search and manual review. Joint speech-and-visual retrieval can surface evidence that appears only on screen while reducing the amount of footage a person must inspect.

Cohere is selling the retrieval layer

Video embedding models already convert clips and text queries into numerical representations that can be compared for similarity. Cohere's differentiator is the surrounding enterprise system: dynamic time boundaries, source-linked results, access controls, mixed document and video retrieval, agent integration, and customer-managed deployment.

The reported benchmark scores indicate that its underlying video representation competes with dedicated models from Google, Amazon, and the Qwen team. Production evaluations will also need to measure timestamp precision, permission enforcement, latency, and performance on each customer's footage.

Production questions remain

  • Supported video formats, codecs, file sizes, languages, and maximum recording lengths
  • Indexing time, query latency, throughput limits, and pricing
  • Accuracy on charts, small on-screen text, overlapping speakers, and rapid scene changes
  • Access-control synchronization, permission revocation, audit logging, and data retention
  • MCP authentication, API response schemas, error handling, and production support commitments

The feature remains in private preview, and Cohere has yet to publish general-availability timing or service-level commitments. Organizations with substantial video collections can request access for deployment in Cohere's cloud or their own environment.

Trending
  • No trending articles

Comments

avatar

Next Reads