Collections: one embedding model
Create Collection now takes an optionalembedding_model.
It defaults to omni_retriever — OmniRetriever, the 3072-dim multilingual vector space — and is
currently the only accepted value, so you can keep omitting the field entirely.- Per-collection model selection is retired: every new collection indexes into the same vector space, and live streams no longer need a collection created a particular way.
- Collections created before this change keep the model they were created with, so
Move Video across two different models still returns
409.
Video Datalake
First preview of the Video Datalake — 41 endpoints:- Collections with fixed embedding (OmniRetriever), optional safety detection and face recognition.
- Ingestion — upload by URL, file, or resumable session; attach live streams.
- Async operations — uniform
202 + Operationmodel with polling and HMAC-signed webhooks. - Moments & derived content — captions, transcripts, frames, clips, titles, summaries, speakers, entities.
- Search — semantic, keyword, hybrid, and image, returning moment lists.
- Face library and safety-detection events.
Agents & tools
Operate the Datalake without writing glue code:- MCP Server — connect Claude, Cursor, or any MCP client and drive the lake in natural language.
- CLI and a Claude Code / Codex plugin for terminal and agent workflows.