Cohere releases 'DesignBERT' — a retrieval-augmented model for component search

AI · 5 min read

Cohere's DesignBERT is a retrieval-augmented generation model built to help teams find and describe existing UI components across codebases, design libraries, and documentation. It ingests Figma files, Storybook stories, and screenshots, then returns the best matching component with usage examples, version history, and suggested variants. The model supports fuzzy queries like 'compact date picker with keyboard support' and returns ranked results.

DesignBERT includes developer-friendly APIs and a connector framework to index private repositories without leaving an organization's VPC. It also offers semantic deduplication to identify near-duplicate components and recommend consolidation opportunities. Cohere emphasized privacy controls and allowed admins to define indexing scopes and exclusion rules.

Large enterprises piloting the model reported improved reuse rates and fewer duplicated components across teams. Some design leads noted that curation remains necessary to ensure recommended components meet accessibility and localization requirements. Cohere plans to add automated suggestion pipelines to surface consolidation candidates directly in design review workflows.