Cohere launches Compass Cloud private beta for managed enterprise retrieval

Cohere has introduced Compass Cloud, a managed offering for its Compass retrieval platform, now entering private beta. Previously available primarily through self-hosted deployments or powering Cohere's North workspace, Compass Cloud allows Cohere to manage the full pipeline and model inference, enabling developers to focus on building applications without infrastructure concerns.

Addressing Enterprise Search Challenges

The platform targets three key shifts in enterprise retrieval requirements:

• Token economics: Precise retrieval reduces irrelevant context, lowering inference costs and task completion time.
• Agentic access patterns: Support for multi-hop queries where agents reformulate requests and traverse multiple sources, requiring reliable performance across sequences of machine-generated queries.
• Fragmented stacks: Consolidation of ingestion, indexing, reranking, and access control into a single integrated service to reduce integration overhead.

Full-Stack Retrieval Architecture

Compass Cloud packages document processing and retrieval into a configurable service with the following components:

• Connect: Out-of-the-box connectors for SharePoint, OneDrive, and Google Drive, supporting multilingual and multimodal data.
• Parse: Transforms complex documents into structured data, using vision processing only when necessary to optimize model usage.
• Embed: Generates dense and sparse representations to match both semantic intent and exact terminology.
• Index: Stores source files, parsed content, and embeddings as separate records, allowing new embedding models to be adopted without re-crawling content.
• Retrieve: Combines semantic, sparse, and keyword search strategies while enforcing permissions during retrieval.
• Rerank: Uses Cohere’s reranker to identify the most relevant passages, reducing irrelevant context for generation.
• Govern: Enforces multi-tenant access control and document-level permissions, with automatic expiration of content based on retention policies.

Performance and Integration

On the High Finance benchmark, an internal Cohere-annotated set of investment banking questions, Compass achieved a 14-16 point improvement in retrieval accuracy (nDCG@10) over Azure Search (rising from 64.8 to 81.1). This gain is attributed to hybrid search capabilities and Cohere’s best-in-class search and document-processing models.

Developers can integrate Compass Cloud via two primary methods:

• Compass API: For custom RAG and search applications, offering control over content representation and ranking.
• Model Context Protocol (MCP): A dedicated, versioned MCP server that exposes retrieval as reusable tools, allowing agents to progressively narrow retrieval corpora for complex tasks.

The private beta is open to limited enterprise teams, with a live session scheduled for October 8th featuring Cohere’s engineering and product leadership.