Clip length is set at search time
Cohere introduced a preview of Video Intelligence in Compass on November 16. The system generates transcripts and visual descriptions for recordings, so a moment can be found even when the speaker only says "look here" and the number appears on a slide or chart. Results are not predefined segments: start and end times are set at query time. Each result carries a timestamp, transcript and visual description and links back to the original footage, and matching document pages, slides and sheets can be returned alongside the clips. Cohere's example is a soccer match: a single frame shows a player leaving the pitch, while only the sequence shows that he was sent off.
Reading the self-reported scores
Cohere compared Compass with seven video embedding models on nDCG@10 across 11 datasets and 11,133 queries. Compass averaged 0.69, 4.4 points above the highest-scoring alternative. By group, it led Amazon Nova Multimodal Embeddings by 2.4 points on six public datasets, Gemini Embedding 2 by 5.1 points on soccer broadcast footage, and Qwen3-VL-Embedding-8B by 2.0 points on meetings, lectures and safety training. These are vendor-selected datasets, and the post's text and chart caption describe the internal datasets differently.
Who can use it now
The preview is limited to invited design partners, who get direct support from Cohere's search team. Deployment can be in Cohere's managed cloud or the customer's own environment, and retrieval respects user permissions. The post does not state supported video formats, length limits or the models used for indexing.