Mixedbread · The retrieval layer an agent actually runs on — they train the embedder, the reranker and the search agent together, which makes them the clearest public test of whether a co-designed stack beats an assembled one.

They take the files a company already has — PDFs, decks, videos, audio, code, Slack, Drive — and turn them into something an AI agent can search and quote from, with the page and passage it came from, so the agent cites what it acted on. They build the whole chain themselves: the embedding model (Wholembed V3), the reranker (mxbai-rerank listwise), the search agent (Toast 1) and the database underneath (Silo). Sold as an API, a CLI, an MCP server and an agent skill rather than a framework you build inside (https://www.mixedbread.com/, read 2026-09-15).

Sector

Memory & skills

Stage

Seed

On the roster since

15 Sept 2026