Dropbox Dash is a product designed to help knowledge workers organize, share, and secure content across various applications.
The product utilizes AI-powered features like advanced search and content access control to improve productivity and collaboration.
Developed using retrieval-augmented generation (RAG) and AI agents, Dash offers powerful search capabilities and granular access controls.
Challenges in building the product included data diversity, data fragmentation, and data modalities unique to business environments.
RAG, an industry-standard approach, was chosen for its ability to combine information retrieval with generative models for accurate responses.
Choosing the right retrieval system and model was crucial for Dash, balancing latency, quality, and data freshness considerations.
AI agents play a key role in handling multi-step tasks autonomously, breaking down complex queries into actionable steps.
The agents follow a structured planning and execution process, leveraging LLMs and DSLs for efficient task handling.
Lessons learned include the importance of selecting appropriate tools, optimizing prompts for different LLMs, and understanding trade-offs between model size, latency, and accuracy.
Future directions include enabling AI agents for multi-turn conversations, self-evaluation, fine-tuning LLMs, supporting multiple languages, and fostering collaboration across diverse teams.
Integrating RAG and AI agents has enhanced Dash's AI capabilities, aiming to meet businesses' needs and advance the future of knowledge work.