Cohere Opens Compass Cloud Private Beta for Managed Enterprise Search

Cohere announced on September 25, 2026, that Compass, its retrieval platform for developers building AI applications on enterprise data, is entering private beta as a managed offering called Compass Cloud, with Cohere operating the full retrieval pipeline and model inference.

Compass surfaces relevant information from a company’s corpus for use in retrieval-augmented generation (RAG), search, and agentic workflows. Rather than assembling and operating the retrieval stack themselves, teams configure Compass through its APIs, an MCP server, or a Python SDK to shape the end-user experience around their data and workflow requirements, according to the Compass Cloud announcement.

From North Integration to a Managed Service

Until now, Compass principally powered retrieval for North, Cohere’s enterprise agent workspace, including its document libraries and MCP ecosystem. Cohere also shipped Compass into highly secure, self-hosted environments for partners in regulated industries whose workloads cannot be offloaded to SaaS.

Cohere said customer demand for a managed option had been clear and consistent: teams wanted what the company calls its best-in-class retrieval capabilities without the operational overhead of self-hosting. The company said Compass Cloud expands Compass to a broader market by having Cohere manage the full pipeline and model inference so customers can focus on building. Self-hosted deployments remain available for privacy-constrained projects, and Cohere is working with a limited number of enterprise teams as beta partners.

Three Changes in Enterprise Search Requirements

The announcement frames the launch around three developments Cohere said are changing enterprise search requirements as retrieval becomes supporting infrastructure for generative AI and agents.

The first is token economics: every irrelevant result passed to a model consumes tokens and occupies limited context space, so more precise retrieval creates smaller, higher-quality inputs that reduce inference costs and the time needed to complete a task. Cohere described retrieval as one of the most effective cost levers available to businesses.

The second is agentic access patterns. Agents may issue dozens of queries while completing one task, reformulating requests and traversing multiple sources; in these multi-hop loops, latency accumulates, relevance can drift, and permissions must be enforced at every step. Retrieval therefore has to perform reliably across sequences of machine-generated queries, not only single-shot searches.

The third is a fragmented retrieval stack: production pipelines often combine separate systems for ingestion, indexing, reranking, access control, and orchestration, leaving teams spending substantial effort on integration rather than retrieval quality. Cohere said Compass exists to address all three by providing relevant, governed context for generative workloads, supporting both agentic and conventional search applications, and consolidating the core retrieval stack into an integrated enterprise platform.

Seven-Stage Retrieval Architecture

Compass packages document processing and retrieval into one configurable service accessed through a single interface. The announcement describes seven stages.

Connect provides out-of-the-box connectors for file-sharing and cloud-storage workspaces such as SharePoint, OneDrive, and Google Drive, with data compatibility Cohere describes as multilingual, multimodal, and file format-agnostic. Parse converts multimodal enterprise files into AI-ready content, applying a parsing strategy to each document and using vision processing only where it adds value to reduce unnecessary model usage. Embed generates dense and sparse representations together, so search can match both semantic intent and domain-specific language across text and multimodal content.

Index keeps source files, parsed content, and embeddings as separate records, so a new embedding model can be adopted without crawling and uploading the same content again; at search time, the records sit in one index with their metadata. Retrieve combines semantic, sparse, and keyword search in a single request, with the strategies running independently or together and permissions enforced during retrieval rather than left to the application.

Rerank applies Cohere’s reranker to identify the most relevant passages from a broad candidate set, reducing irrelevant context and the tokens required for generation. Govern enforces multi-tenant access control and document-level permissions during retrieval, while retention policies expire content automatically and prevent deleted documents from being resynced.

Company-Reported Retrieval Benchmark

Cohere reported that Compass achieved a 14–16 point improvement over Azure Search on High Finance, an internal Cohere-built investment-banking benchmark, with retrieval accuracy measured as nDCG@10 rising from 64.8 to 81.1. High Finance is a Cohere-annotated set of questions asking models to retrieve relevant investment banking and hedge fund presentation material; the workload embeds a query, retrieves presentation materials from an index, and scores the top results.

Under the stated methodology, Azure Search documents were parsed with GPT-4.1 Mini, while Cohere Embed 4 used no separate parsing solution and embedded the input PDFs directly as images before searching. Cohere attributed the accuracy to its search and document-processing models working with Compass’ hybrid search of lexical, sparse, and dense retrieval followed by reranking, and said a gap of that size can separate an unsatisfactory answer from a great one for the end user.

Integration Paths, Beta Terms, and Next Steps

Compass Cloud offers two integration paths. The Compass API targets teams building custom RAG and search applications, connecting directly to the models and retrieval components that power Compass and giving developers control over how content is represented, retrieved, ranked, and incorporated into an application. A dedicated, independently versioned MCP server exposes retrieval as reusable tools through an open standard, so MCP-compatible clients can discover and call those tools directly and Compass stays portable across agent frameworks without a bespoke integration for each one.

Cohere describes agentic retrieval as an emerging approach at the intersection of large language models and information retrieval: instead of a single query-and-response step, an agent decides what to look for, searches for relevant information, evaluates the results, and refines its search before answering. With the Compass MCP server, the company said, agents can progressively narrow the retrieval corpus, reducing unnecessary context and making complex retrieval tasks faster and more token-efficient.

The private beta is open to teams building retrieval-heavy and agentic applications. Participants receive hands-on support from Cohere’s engineering team, early access to the cloud deployment, and direct influence on the product roadmap, with access requested through a form on the Compass page. Cohere is also running a live session on October 8, 2026, on X with its engineering and product leadership on where enterprise search goes next.