|
How to stop bots from abusing free trials (Sponsored)
 |
Free trials help AI apps grow, but bots and fake accounts exploit them. They steal tokens, burn compute, and disrupt real users.
Cursor, the fast-growing AI code assistant, uses WorkOS Radar to detect and stop abuse in real time. With device fingerprinting and behavioral signals, Radar blocks fraud before it reaches your app.
Disclaimer: The details in this post have been derived from the details shared online by the Dropbox Engineering Team. All credit for the technical details goes to the Dropbox Engineering Team. The links to the original articles and sources are present in the references section at the end of the post. We’ve attempted to analyze the details and provide our input about them. If you find any inaccuracies or omissions, please leave a comment, and we will do our best to fix them.
You’re racing against a deadline, and you desperately need that specific image from last month’s campaign or that video clip from a client presentation. You know it exists somewhere in your folders, but where? Was it in that project folder? A shared team drive? Or nested somewhere three levels deep in an old archive?
We’ve all been in this situation at some point, as this is the daily reality for knowledge workers who lose countless hours hunting for the right files within their cloud storage.
The problem becomes even more frustrating with multimedia content. While documents often have descriptive titles and searchable text inside them, images and videos typically come with cryptic default names like IMG_6798 or VID_20240315. Without meaningful labels, these files become nearly impossible to locate unless you manually browse through folders or remember exactly where you saved them.
Dropbox solved this problem by building multimedia search capabilities into Dropbox Dash, their universal search and knowledge management platform.
The challenge their engineering team faced wasn’t just about finding a file anymore. It’s about finding what’s inside that file. And when the folder structure inevitably breaks down, when files get moved or renamed by team members, or when you simply can’t recall the location of what you need, traditional filename-based search falls short.
In this article, we’ll explore how the Dropbox engineering team implemented multimedia search features and the technical challenges they faced along the way.
Challenges of Multimedia Search
Building a search feature for images, videos, and audio files presents a fundamentally different set of problems compared to searching through text documents.
Some of the key challenges are as follows:
Storage Costs: The sheer size difference is significant. Image files average about 3X larger than typical documents, while video files clock in at roughly 13X larger. These size differences directly translate to increased storage demands and costs.
Compute Intensity: Beyond storage, multimedia files require substantially more processing power to extract useful features. The complexity goes beyond just handling larger files. Unlike text documents, multimedia search needs visual previews at multiple resolutions to be useful, dramatically increasing computational requirements.
Ranking Relevance: Dropbox Dash already operated a sophisticated multi-phase retrieval and ranking system optimized for textual content. Extending this to multimedia meant indexing entirely new types of signals, creating query plans that leverage these signals effectively, and handling edge cases to avoid irrelevant results appearing at the top.
Preview Generation Dilemma: Users need visual previews to quickly identify the right file, and they need these previews in multiple resolutions for a smooth experience. However, only a small fraction of indexed files actually get viewed during searches. Pre-generating previews for everything would be extremely wasteful, but generating them on demand during searches introduces latency challenges that could frustrate users.
The Dropbox engineering team had to ensure their solution supported seamless browsing, filtering, and previewing of media content directly within Dash. This meant confronting higher infrastructure costs, stricter performance requirements, and adapting systems originally designed for text-based retrieval.
The Architecture
To deliver fast and accurate multimedia search while keeping costs manageable, the Dropbox engineering team designed a solution built on three core pillars:
A metadata-first indexing pipeline
Intelligent location-aware search
A preview generation system that creates visuals only when needed
Indexing Pipeline for Metadata
The foundation of multimedia search begins with indexing, the process of cataloging files so they can be quickly retrieved later. Dropbox made a critical early decision to index lightweight metadata rather than performing deep content analysis on every single file. This approach dramatically reduces computational costs while still enabling effective search.
Before building this multimedia search capability, Dropbox had intentionally avoided downloading or storing raw media blobs to keep storage and compute costs low. However, this meant their existing search index lacked the necessary features to support rich, media-specific search experiences. To bridge this gap, the team added support for ingesting multimedia blob content to extract the required features. Importantly, they retain the raw content not just for preview generation, but also to enable computing additional features in the future without needing to re-ingest files.
To power this indexing pipeline, Dropbox leveraged Riviera, its existing internal compute framework that already processes tens of petabytes of data daily for Dropbox Search. By reusing proven infrastructure, the team gained immediate benefits in scalability and reliability without building something entirely from scratch.
The indexing process extracts several key pieces of information from each multimedia file. This includes basic details like file path and title, EXIF data such as camera metadata, timestamps, and GPS coordinates, and even third-party preview URLs when available from applications like Canva.
See the diagram below:
 |
The data flows through the system in the following way:
Raw files are stored in a blob store
Riviera extracts features and metadata from these files
Information flows through third-party connectors
Kafka message brokers transport the data
Transformers process and structure the