Dropbox Engineering VP Details Knowledge Graphs and DSPy in Dash AI
Josh Clemm, Vice President of Engineering at Dropbox, provides a technical deep-dive into the architecture powering Dropbox Dash, an AI-driven workspace assistant. Speaking on Jason Liu’s Maven course, Clemm explains how Dash addresses information fragmentation across multiple SaaS applications by aggregating proprietary user content. The system utilizes custom connectors to handle diverse API quirks and access controls, followed by robust content understanding that normalizes files into markdown and extracts metadata. For complex media like images, audio, and video, Dash employs multimodal models to generate contextual embeddings beyond simple transcription. A key component is the construction of knowledge graphs that link related entities such as meetings, documents, and people to enhance context. Retrieved data is stored in secure lexical (BM25) and vector indexes, enabling hybrid retrieval. Clemm also highlights the use of Model Context Protocol (MCP) for tool calling and DSPy for prompt optimization, alongside LLM-as-a-judge mechanisms. This approach allows Dash to deliver accurate, context-aware answers and agentic queries, effectively bridging the gap between general large language models and private enterprise data.
Wire timeline
Dropbox Engineering VP Details Knowledge Graphs and DSPy in Dash AI
Josh Clemm, Vice President of Engineering at Dropbox, provides a technical deep-dive into the architecture powering Dropbox Dash, an AI-driven workspace assistant. Speaking on Jason Liu’s Maven course, Clemm explains how Dash addresses information fragmentation across multiple SaaS applications by aggregating proprietary user content. The system utilizes custom connectors to handle diverse API quirks and access controls, followed by robust content understanding that normalizes files into markdown and extracts metadata. For complex media like images, audio, and video, Dash employs multimodal models to generate contextual embeddings beyond simple transcription. A key component is the construction of knowledge graphs that link related entities such as meetings, documents, and people to enhance context. Retrieved data is stored in secure lexical (BM25) and vector indexes, enabling hybrid retrieval. Clemm also highlights the use of Model Context Protocol (MCP) for tool calling and DSPy for prompt optimization, alongside LLM-as-a-judge mechanisms. This approach allows Dash to deliver accurate, context-aware answers and agentic queries, effectively bridging the gap between general large language models and private enterprise data.
Dropbox Tech Blog