TechPrufer
Get started
Dropbox
Dropbox logoDropbox · Feature trail

Universal Content Processing Platform

13 posts · Sep 2018 – Sep 2026

What this trail covers

Dropbox Dash has evolved from a traditional RAG pipeline to an agentic AI system. This evolution is driven by advancements in context engineering, focusing on improving the precision and relevance of information provided to AI models. Key strategies include consolidating retrieval tools into a single universal search index, filtering context through a knowledge graph to prioritize relevance, and introducing specialized agents for complex tasks like query construction. The integration of Mobius L Reclaim, a calendar assistant, has evolved to be AI-native by building an agent platform that connects LLMs to Reclaim's scheduling logic and calendar context. This platform allows users to describe scheduling goals in natural language, which an agent interprets and translates into Reclaim's internal 'Schedule Action' operations. A 'Preview Mode' was developed to allow users to review AI-proposed changes before applying them to their actual calendars, ensuring consistency and user control. The system connects directly to model providers and owns the agent loop, tools, and context gathering to maintain flexibility and control.

All Dropbox trails
  1. Evolving our calendar assistant Reclaim to be AI-native without starting overReclaim rebuilt its calendar assistant to be AI-native by developing an agent platform. This platform connects LLMs to Reclaim's existing scheduling logic and calendar context. Key technical contributions include: 1. An agent loop that manages the interaction between the model, tools (for information lookup and actions), and calendar context. 2. Direct integration with model providers, owning the agent loop, tools, and context gathering for flexibility. 3. A 'Schedule Action' abstraction that unifies how user requests, automated scheduling, and AI agents modify calendar events, ensuring consistent behavior. 4. A 'Preview Mode' that uses a pure-function automated scheduler and Redis for fast, reviewable simulation of proposed calendar changes before they are applied.
  2. How our universal content processing platform Riviera evolved for AI and beyondThis post details the evolution of Dropbox's Riviera platform from a file preview service to a universal content processing platform. It highlights the architectural shift towards reusable transformations instead of product-specific pipelines, enabling support for over 300 file formats and diverse workloads. The adoption by teams like Search, Replay, Sign, and Dash, and its crucial role in powering AI features like Dash by preparing content for models, are emphasized. The post also announces the external availability of these capabilities via APIs and Model Context Protocol tools, extending reuse beyond Dropbox.
  3. How we used DSPy to turn AI evaluations into better responses in Dash chatThis post details the application of DSPy, an open-source framework for optimizing AI systems, to improve the quality of AI-generated responses in Dropbox's Dash chat agent. It describes the process of using LLM-as-judge evaluations, calibrating these judges with human-labeled examples, and then employing DSPy's optimization algorithms (GEPA, MIPROv2) to refine the judges themselves. Subsequently, these improved judges were used to optimize the chat agent's system prompt through an automated feedback loop, replaying historical chats and scoring agent outputs to identify and correct failure modes such as wrong context selection, incomplete answers, and missed ambiguity. This resulted in a reduction of incomplete answers and token usage.
  4. How we optimized Dash's relevance judge with DSPyThis post details the adaptation of Dropbox Dash's relevance judge from a proprietary model (OpenAI's o3) to an open-weight model (gpt-oss-120b) using the DSPy framework. It outlines the process of defining a measurable objective (NMSE against human judgments), using DSPy's GEPA optimizer for prompt adaptation, and generating structured feedback for iterative prompt improvement. The contribution focuses on making relevance judging more cost-effective and reliable at scale by systematically optimizing prompts against a defined metric.
  5. Using LLMs to amplify human labeling and improve Dash search relevanceThis post details how Dropbox is using LLMs to amplify human labeling for training Dash's search relevance models. It describes a process where a small set of human-labeled data is used to tune LLM prompts and parameters. The LLM then generates a large volume of relevance labels, acting as a force multiplier for human effort. The post highlights the trade-offs between human and LLM evaluation, the importance of evaluation metrics like Mean Squared Error (MSE), and the use of LLMs offline to generate training data for smaller, more efficient relevance models.
  7. How low-bit inference enables efficient AIThis post details how Dropbox leverages low-bit inference techniques, specifically quantization, to improve the efficiency and reduce the cost of running large AI models for products like Dropbox Dash. It explains the trade-offs between different quantization formats (e.g., 16-bit to 4-bit), the role of hardware acceleration (Tensor Cores, Matrix Cores), and the challenges of bitpacking for sub-byte formats. The post highlights how these optimizations are crucial for delivering responsive user experiences and managing resource usage in production AI workloads.
  8. Engineering VP Josh Clemm on how we use knowledge graphs, MCP, and DSPy in DashThis post details Dropbox Dash's context engine, which ingests data from third-party apps, performs content understanding (including multimodal analysis for videos), models information as a knowledge graph, and stores it in secure data stores (lexical index with BM25 and vector store) for hybrid retrieval. It discusses the trade-offs between federated and indexed retrieval, highlighting the benefits of indexed retrieval for company-wide access and pre-processing. The post also addresses challenges with MCP (Model Context Protocol) tool definitions consuming context window space and proposes solutions like a 'super tool' for index-based retrieval, knowledge graph modeling to reduce token usage, storing tool results locally, and using sub-agents for complex queries. It elaborates on the value of knowledge graphs for modeling relationships across apps and providing canonical IDs for entities like people, which aids in relevance and retrieval.
  9. How Dash uses context engineering for smarter AIThis post details the evolution of Dropbox Dash from a RAG pipeline to an agentic AI system by introducing context engineering strategies. It describes the consolidation of multiple retrieval tools into a single universal search index, the use of a knowledge graph to filter and rank context for relevance, and the creation of specialized agents (e.g., for search query construction) to handle complex tasks. The goal is to improve model reasoning, planning, and action by providing focused and relevant context, addressing issues like analysis paralysis and context rot.
  10. With Mobius Labs' Aana models, we're bringing deeper multimodal understanding to Dropbox DashThis post details the integration of Mobius Labs' Aana multimodal models into Dropbox Dash. Aana provides an ultra-efficient architecture for understanding rich media (audio, video, images) at scale by analyzing inter-modal relationships. It combines open-source, fine-tuned foundation models for speech (Whisper-like), vision (transformer-based, MoE), and language. Key technical components include HQQ for low-bit inference (8-bit, 4-bit) to reduce compute and memory costs, and Gemlite for accelerating core AI operations with custom GPU kernels. The Aana SDK orchestrates these optimizations, handling batching, model coordination, and efficient GPU utilization, enabling flexible configuration and deployment of multimodal processing pipelines.
  11. A practical blueprint for evaluating conversational AI at scaleThis post details a practical blueprint for evaluating conversational AI at scale, emphasizing an "evaluation-first" approach. It covers curating datasets (public and internal), defining actionable metrics using LLMs as judges, and setting up an evaluation platform (Braintrust). The process treats every experiment like production code, with hard gates for boolean metrics and scalar budgets, and continuous monitoring in production. This ensures consistency, traceability, and reliable quality control for AI systems.
  12. How we brought multimedia search to Dropbox DashThis post details the engineering behind adding multimedia search (images, video, audio) to Dropbox Dash. It covers the challenges of storage and compute costs, relevance, and responsiveness for media files. The solution involves indexing lightweight metadata (file path, title, EXIF) using the Riviera framework, backfilling the index with raw content for preview generation, and optimizing storage. At query time, it matches input against metadata, applies custom logic for geographic references, and tokenizes camel case/hyphenated strings. Previews are generated just-in-time on the read path to balance latency and resource efficiency.
  13. Building Dash: How RAG and AI agents help us meet the needs of businessesThis post details the evolution of Dropbox Dash, moving from a traditional RAG pipeline to an agentic AI system. It highlights the challenges of data diversity, fragmentation, and modalities in business data environments. The core technical contributions involve the application of Retrieval-Augmented Generation (RAG) and AI agents. Specific technical trade-offs discussed include latency vs. quality in retrieval systems, data freshness vs. scalability, and budget vs. user experience. The post also mentions the engineering of a minimal Python interpreter for AI agents and the use of semantic search with vector indexes versus lexical search. It emphasizes the importance of the retrieval system in setting the bounds of an LLM's knowledge and impacting latency and perceived quality.
  14. Selecting a model for semantic search at Dropbox scaleThis post details the integration of semantic search into Dropbox's existing Nautilus search engine. It describes the shift from keyword-based search to understanding meaning and intent through embeddings. The authors evaluated text embedding models using a customized version of the Massive Text Embedding Benchmark (MTEB), adapting it for Dropbox's infrastructure, handling multiple embeddings per document, optimizing precision and dimensionality, and incorporating filename embeddings. The post highlights the observed metrics of a nearly 17% reduction in empty search sessions and a 2% lift in search session success rate.
  15. Bringing AI-powered answers and summaries to file previews on the webThis post details the integration of AI-powered summarization and Q&A features into Dropbox's web file preview experience. It describes how the existing Riviera file conversion framework was extended to handle AI embeddings. The summarization plugin uses k-means clustering on text embeddings to identify key concepts and generate summaries, improving topic diversity and reducing LLM hallucinations compared to map-reduce approaches. The Q&A plugin uses embedding similarity to find relevant text chunks for answering user questions.
  16. Putting everything in its right place with ML-powered file organizationThis post details the development of the 'Smart Move' feature, an AI-assisted file organization capability. It describes the ML-driven approach to analyzing user's existing subfolder structures and suggesting optimal file placements. Key technical challenges addressed include product design and understanding diverse user organization patterns, handling sensitive data with new workflows and storage solutions, identifying appropriate datasets for training, and optimizing for low latency in synchronous UX flows. The post also discusses the development of a new ML model by treating existing organized folders as labeled end states, and the importance of low-overhead prototypes for gathering user feedback.
  17. Cannes: How ML saves us $1.7M a year on document previewsThis post details the application of a gradient-boosted classifier (Cannes v1) to predict the likelihood of document previews being viewed. The system integrates with the Riviera platform to optimize pre-warming of preview assets, saving $1.7M annually. Key technical contributions include the development of a prediction pipeline that fetches live signals (from Edgestore and User Profile Service) and feeds them into a model, the use of A/B testing via Stormcrow to validate performance, and the careful consideration of tradeoffs between cost-benefit, complexity, interpretability, and deployment cost. The model uses file extension, account type, and recent activity as features, achieving >70% accuracy in predicting preview usage within 60 days.
  18. Machine intelligence at Dropbox: An update from our DBXi teamThis post introduces the Dropbox Intelligence Initiative (DBXi) and its application to improve search quality and speed. It details the use of machine intelligence for organization, contextualization, and prioritization of work content. A prototype is presented that uses a user-specific graph connecting people, content, and activity signals to intelligently highlight important work and provide personalized activity feeds. The post also mentions foundational work in image recognition, OCR, and the rebuilding of the search infrastructure.