Enterprise AI efforts often stall at the point where a model needs access to internal information without weakening data controls. Advanced Micro Devices, Inc. (NASDAQ: AMD) and Cohere Compass are positioning a combined retrieval platform and compute configuration for organizations that want to search documents and repositories within infrastructure they control.
The companies are targeting a familiar enterprise problem: business information may sit across documents, presentations, technical manuals and internal repositories, each with separate systems and access rules. Connecting that material to employees, applications and AI agents requires a retrieval layer as well as a model. AMD said the configuration is designed for organizations that need to preserve control over sensitive data while building and operating that foundation.
Compass is Cohere’s platform for developers building AI applications with enterprise data. According to AMD, it combines document processing, indexing and retrieval with Cohere Embed, Rerank and Parse models. Those functions can provide the context used by knowledge assistants, retrieval-augmented generation applications and AI agents, rather than leaving each application to assemble its own search pipeline.
The deployment model is a key part of the proposition. AMD said Compass on supported AMD infrastructure can support on-premises deployment, role-based access controls and document-level security. That may matter to organizations that need AI systems to use internal material while retaining their own boundaries for where data is stored and which users or applications may retrieve it.
AMD’s role is the infrastructure layer. The company said AMD Instinct GPUs accelerate supported AI model workloads, AMD EPYC processors provide host computing resources and ROCm supplies the GPU software foundation. The stated goal is to give customers a defined basis for sizing retrieval services around model requirements, document volumes and application demand, while keeping those services close to enterprise data.
The companies describe several expected use cases: helping employees and agents retrieve material from technical documentation, operating procedures and policies; grounding AI applications in business-specific context; and supporting knowledge assistants and agentic workflows. AMD also cited support teams locating product documentation, engineers retrieving internal technical knowledge and business teams navigating policies and procedures. These are proposed uses, not published measures of search quality, deployment time or operating cost.
For CIOs and platform leaders, the practical question is whether the configuration reduces the integration burden without imposing a new control problem. A retrieval service still needs accurate document preparation, well-defined permissions and capacity planning, and the companies have not published performance results for a particular enterprise deployment. Organizations will also need to assess how the configuration fits their existing identity systems, repositories and application architecture.
The more durable significance is that private search is becoming a shared infrastructure concern rather than a feature embedded separately in each AI project. Legacy enterprise search often focused on locating documents for people, while early generative-AI deployments frequently relied on isolated prompts or loosely connected data stores. AMD and Cohere are proposing a middle layer that gives applications and agents controlled access to organizational knowledge, making governance and retrieval architecture central to how enterprises scale AI workflows.
