Notion Strengthens Workspace Search with Cohere Rerank

Notion directly integrated Cohere Rerank into its workspace search, creating a faster and more accurate search experience even in environments with large volumes of documents and databases. Since Notion AI leverages knowledge from connected tools like Slack and Google Drive, search quality was central to product performance.

The problem it faced was clear. As organizations grew, the volume of information surged, making it harder to quickly find what was needed, and the following requirements had to be met simultaneously.

• Improve search accuracy on large-scale datasets
• Fast search across environments handling thousands of documents and databases
• Maintain high relevance without relying on embeddings and vector search
• Reliably support multilingual data and non-English-speaking users

The solution was to put Cohere Rerank into the workspace search pipeline. First, it reranked search results to surface highly relevant documents before the generative model saw them, and for smaller workspaces—typically under 1,000 documents—it skipped embedding and vector search to reduce complexity and cost.

The infrastructure was built on Amazon SageMaker with auto-scaling applied. This allowed compute resources to flex between high-traffic and quiet periods, delivering stable performance to users around the world.

Multilingual support was also important. Cohere's models natively support multiple languages, providing the same level of search quality to users handling multilingual data, especially in EMEA and APAC. Notion explained that by leaving precision to Rerank, the team could focus on other areas like recall.

The effects can be summarized in three points.

• Improved search performance: delivering more relevant results quickly
• Cost reduction: reducing the burden of embedding generation, storage, and vector search
• Scalability: operating search functionality flexibly in response to traffic changes

Notion also uses Cohere Rerank in the process of merging search results from multiple sources such as Slack and GitHub, maintaining high accuracy while improving cost efficiency. Currently, improving user experience every month is treated as a core KPI, and the number of users who have tried Notion AI has reached millions, with related revenue also growing rapidly.