Engineering VP Josh Clemm on how we use knowledge graphs, MCP, and DSPy in Dash

Josh Clemm, VP of Engineering at Dropbox, explains how Dash uses knowledge graphs, MCP, and DSPy to build a universal work search and AI assistant.

• •Dash connects to third-party SaaS apps via custom crawlers, normalizes content into formats like markdown, and builds a unified index using BM25 and dense vector embeddings.
• •Index-based retrieval was chosen over federated retrieval for its speed, company-wide access, and ability to pre-process enriched datasets, despite higher complexity and cost.
• •Knowledge graphs model cross-app relationships (e.g., calendar invites linked to attendees, documents, Jira projects) using canonical IDs to improve context quality.
• •MCP tool definitions cause context window bloat and latency (up to 45s per query); Dropbox mitigates this with a "super tool" wrapping the index, local storage of tool results, and sub-agents with narrow toolsets.
• •Token usage is reduced by serving only the most relevant graph data per query rather than returning raw, unfilte