What happened
Mintlify, a documentation platform serving hundreds of thousands of users, has publicly shared how they replaced traditional RAG (Retrieval Augmented Generation) with what they call ChromaFs: a virtual filesystem that lets AI agents navigate documentation the same way developers navigate code.
The problem with their previous approach was straightforward: RAG could only retrieve chunks of text matching a query. When answers lived across multiple pages, or users needed exact syntax that didn't land in a top-K result, the system was stuck. The AI assistant couldn't explore or traverse documentation structure.
Their solution treats each documentation page as a file and each section as a directory. The AI agent can now use familiar commands: grep to search for exact strings, cat to read full pages, ls to list contents, and find to traverse structure. But instead of spinning up expensive sandbox environments with real filesystems, ChromaFs intercepts these commands and translates them into queries against their existing Chroma vector database.
Why this matters
The performance gains are striking. Session creation dropped from approximately 46 seconds to around 100 milliseconds. The marginal compute cost per conversation fell from roughly $0.014 to effectively zero, since ChromaFs reuses infrastructure they already pay for.
At scale, the savings compound. Mintlify estimates that running dedicated micro-VMs for 850,000 monthly conversations would cost over $70,000 per year. Their virtual filesystem approach eliminates this cost entirely while delivering faster response times.
But the architectural insight matters more than the numbers. AI agents are converging on filesystems as their primary interface because the primitives are universal. grep, cat, ls, and find are all an agent needs to explore structured information. The key insight: the agent doesn't need a real filesystem, just the illusion of one.
The Laava perspective
This approach validates what we've been building at Laava. RAG is a powerful tool, but it's not the only tool. Production AI systems often need hybrid approaches that combine semantic search with deterministic traversal.
Our 3 Layer Architecture separates Context, Reasoning, and Action precisely because different problems need different retrieval strategies. Sometimes you need semantic similarity search. Sometimes you need exact string matching. Sometimes you need to traverse a document hierarchy. A well-designed system supports all three.
Mintlify's approach also demonstrates a principle we apply with enterprise clients: reuse existing infrastructure. They already had their documentation indexed in Chroma. Rather than adding new systems, they built a thin translation layer on top. This is the boring, reliable engineering that production AI requires.
The access control integration is particularly elegant. Because file visibility is determined before the filesystem tree is built, the agent can't even reference paths a user isn't authorized to see. Security is architectural, not bolted on.
What you can do
If you're building AI assistants that need to navigate structured information, consider whether pure RAG is the right approach. Ask yourself: does the agent need semantic similarity, or does it need to traverse and explore? Often the answer is both.
Laava helps enterprises design AI architectures that match retrieval strategy to use case. If you're facing limitations with your current RAG implementation, or if your AI assistant struggles with multi-page answers and exact syntax lookups, we can help you evaluate hybrid approaches like the one Mintlify has deployed.