Cohere opens Compass Cloud private beta

Cohere’s retrieval platform Compass is entering private beta as a managed offering called Compass Cloud, giving developers access to the search and retrieval engine that has mostly powered North, Cohere’s enterprise agent workspace. Teams reach it through APIs, an MCP server or a Python SDK instead of assembling and operating the stack themselves, and self-hosted deployments stay available for privacy-constrained projects.

What’s actually new

• The full retrieval pipeline in one service. Connect handles sources such as SharePoint, OneDrive and Google Drive, then Parse turns multimodal files into structured data, Embed generates dense and sparse representations together, Index keeps source files, parsed content and embeddings as separate records so a new embedding model does not require re-crawling, and Rerank and Govern finish the job with permission enforcement during retrieval rather than in the application.
• A specific accuracy claim. On High Finance, Cohere’s internal investment-banking retrieval benchmark, Compass scored 64.8 to 81.1 on nDCG@10, a 14 to 16 point improvement over Azure Search.
• An independently versioned MCP server. Retrieval ships as reusable tools over an open standard, so agents can call it without a bespoke integration per framework, and can narrow the corpus across multiple queries instead of one lookup.
• Limited beta, hands-on support. Cohere is working with a small number of enterprise teams, with a live session on X on October 8.

Why it matters

The pitch is token economics. Every irrelevant result a retriever returns is context an agent pays to read, so retrieval quality shows up directly in inference cost, and a 14 to 16 point nDCG gap is the kind of number that moves a buying decision. The more telling detail is that retrieval has become its own product category with its own managed SKU, which says how much of agent quality now lives below the model.

(Original source: Cohere)