Mintlify replaced RAG chunking and sandbox VMs with ChromaFs — a virtual filesystem over Chroma — cutting session creation from 46s to milliseconds at 850K conversations/month.
Mintlify's engineering team replaced their RAG-based documentation assistant with a virtual filesystem called ChromaFs, built on top of their existing Chroma vector database. Instead of spinning up sandboxes to clone repos (which took ~46s p90 and would cost $70K+/year at scale), each doc page becomes a file and each section a directory — letting agents use grep, cat, ls, and find natively. The result: instant session creation, zero marginal compute cost, and built-in RBAC. ChromaFs now powers 30,000+ conversations per day across hundreds of thousands of users.
This is a concrete architectural pattern shift: instead of embedding chunks and hoping cosine similarity surfaces the right syntax, expose docs as a traversable filesystem so agents can cat, grep, and find their way to exact answers. The key insight is that agents already know how to use UNIX tools — stop fighting that and build the interface they expect. ChromaFs achieves this without spinning up real VMs by mapping Chroma records to filesystem primitives, killing the 46s sandbox cold-start problem entirely.
If you're building a docs or codebase agent this week, prototype the filesystem interface pattern: structure your vector DB records with path metadata (page → file, section → directory), then expose grep and ls as tool calls instead of a single semantic search endpoint. Benchmark whether your agent reaches correct answers in fewer tool calls than RAG.
Open a Python file and install: pip install chromadb
Tags
Signals by role
Also today
Tools mentioned