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

Engineering VP Josh Clemm recently spoke about the innovative technologies and approaches that Dropbox Dash is using to revolutionize the way users access and interact with their work-related content. In a detailed deep-dive, Clemm explored the role of knowledge graphs, MCP (Multiturn Conversational Prompting), and DSPy (DeepSpeed Prompt Optimizer) in enhancing the capabilities of Dash.

The challenge of managing a vast array of proprietary content across multiple applications and platforms is a significant hurdle for businesses. With numerous tabs and accounts open, finding specific information becomes increasingly difficult. While large language models (LLMs) are advancing rapidly, they often lack access to proprietary data stored within walled gardens. This is where Dropbox Dash steps in, offering a unified solution to connect and search across various third-party apps and content sources.

At the core of Dash's functionality is its context engine, which relies on custom crawlers and connectors to gather data from diverse applications. Building these connectors is a complex task, as each platform has unique API quirks, rate limits, and permission systems. However, overcoming these challenges is crucial for aggregating all content into a single, searchable space.

Once the content is collected, Dash focuses on understanding and enriching it. This involves normalizing different file formats and structures to ensure consistency. By doing so, Dash can provide a cohesive and accurate representation of the data, enabling more effective search and retrieval.

Knowledge graphs play a pivotal role in Dash's architecture. These graphs allow for the organization and interlinking of information, facilitating efficient querying and understanding of complex relationships between data points. By leveraging knowledge graphs, Dash can offer more accurate and context-aware responses to user queries.

MCP, or Multiturn Conversational Prompting, is another critical component of Dash's technology stack. This approach enables the system to handle multi-step conversations, allowing users to ask follow-up questions and receive contextual answers. By using MCP, Dash can maintain a coherent line of inquiry, providing more comprehensive and helpful responses over time.

To further optimize the performance of these systems, Dash employs DSPy, a powerful tool for prompt optimization. DSPy helps refine and fine-tune the prompts used by the LLMs, ensuring that the models generate the most relevant and accurate results. This continuous improvement process enhances the overall effectiveness of Dash in understanding and responding to user queries.

In addition to these technologies, Dash also incorporates a unique approach to LLM evaluation. The system acts as a judge, assessing the quality and relevance of the LLM's responses. This ensures that Dash delivers the most useful and accurate information to its users.

Overall, Dropbox Dash is a groundbreaking solution that addresses the challenges of managing proprietary content across multiple platforms. By utilizing knowledge graphs, MCP, and DSPy, Dash provides a unified and intelligent approach to accessing and interacting with work-related information. As the technology continues to evolve, Dash remains committed to pushing the boundaries of what is possible in the realm of enterprise content management.