Snowflake has partnered with TwelveLabs to bring video search and analysis into Cortex AI Functions as SQL calls.
How do you find one moment in 10,000 ad creatives without watching them? You stop tagging and start embedding.
Marengo is the perception engine, a multimodal encoder that turns visuals, audio and speech into comparable numbers in a single representation. It outputs 512-dimensional vectors per segment, according to Snowflake, with support for 36 languages on spoken content. A 2-minute video might produce 10 visual segments, 8 audio segments and 12 transcription segments, so counts scale with length and content.
The embed step is one function. SELECT AI_MULTI_EMBED with twelvelabs-marengo-embed-3-0 reads a file from a Snowflake stage, internal or external on S3, GCS or Azure, and writes vectors to a table. Cortex Search then indexes those vectors for plain-text queries like a ship burning at night. Challenger School built a footage-search tool on that path, said Yashwanth Karumanchi, Research and Development at Challenger School Foundation, whose team now gets timestamps, transcript context and Premiere review sequences linked to source files. Karumanchi said editors get a practical starting point instead of weeks of manual review.
Pegasus is the reasoning engine, a video-language model that watches a clip and writes text. SELECT AI_COMPLETE with twelvelabs-pegasus-1-5 asks for summaries, scene descriptions or structured metadata without a separate frame-extraction or transcription pipeline. Cortex Agents can then combine visual, audio and transcription signals with custom ranking through tool calling.
The deployment change is small and the budget change is not. Video stays on stages, but derived signals live beside campaign, catalog and spend data for joint queries. Governance helps, with row-level policies, role-based access and audit logs applying to video-derived data stored in Snowflake tables, according to Snowflake. That still leaves raw files and stage permissions to lock down, plus vector growth to monitor as libraries scale. Watch Marengo 3.0 and Pegasus 1.5, now generally available, on whether teams query archives enough to turn cold storage into an asset. That shift follows a landmark year for AI at Snowflake.





