Mintlify ChromaFs: RAG Is Dead, Give Agents a Filesystem Instead
Mintlify just published a technical deep dive that should reframe how we think about agents accessing documentation. They replaced their entire RAG pipeline with a virtual filesystem called ChromaFs — and the results are dramatic.
The problem with RAG for documentation agents is fundamental: RAG retrieves chunks of text that match a query, but if the answer spans multiple pages or requires exact syntax that doesn't land in a top-K result, the agent is stuck. Mintlify's insight is that agents are converging on filesystems as their primary interface. grep, cat, ls, and find are all an agent needs. If each doc page is a file and each section is a directory, the agent can search for exact strings, read full pages, and traverse the structure autonomously.
ChromaFs maps standard UNIX commands to queries against Mintlify's Chroma vector database. The agent thinks it's navigating a filesystem, but every operation is actually a semantic query under the hood. Session creation dropped from 46 seconds p90 to about 100 milliseconds. Marginal per-conversation compute cost went to zero because the indexed DB gets reused. RBAC enforcement comes free — the filesystem layer enforces per-user access without sandboxes.
This isn't a toy experiment. ChromaFs powers 30,000+ daily conversations for Mintlify's documentation assistant, serving hundreds of thousands of users. The blog post hit 309 points on Hacker News today.
The broader implication is that the filesystem abstraction may be the right universal interface for agents — not REST APIs, not GraphQL, not RAG. Files and directories are the lowest common denominator that every agent framework already understands. Details at mintlify.com/blog/how-we-built-a-virtual-filesystem-for-our-assistant.
← Back to all articles
The problem with RAG for documentation agents is fundamental: RAG retrieves chunks of text that match a query, but if the answer spans multiple pages or requires exact syntax that doesn't land in a top-K result, the agent is stuck. Mintlify's insight is that agents are converging on filesystems as their primary interface. grep, cat, ls, and find are all an agent needs. If each doc page is a file and each section is a directory, the agent can search for exact strings, read full pages, and traverse the structure autonomously.
ChromaFs maps standard UNIX commands to queries against Mintlify's Chroma vector database. The agent thinks it's navigating a filesystem, but every operation is actually a semantic query under the hood. Session creation dropped from 46 seconds p90 to about 100 milliseconds. Marginal per-conversation compute cost went to zero because the indexed DB gets reused. RBAC enforcement comes free — the filesystem layer enforces per-user access without sandboxes.
This isn't a toy experiment. ChromaFs powers 30,000+ daily conversations for Mintlify's documentation assistant, serving hundreds of thousands of users. The blog post hit 309 points on Hacker News today.
The broader implication is that the filesystem abstraction may be the right universal interface for agents — not REST APIs, not GraphQL, not RAG. Files and directories are the lowest common denominator that every agent framework already understands. Details at mintlify.com/blog/how-we-built-a-virtual-filesystem-for-our-assistant.
Comments