April 3, 2026
Grep beats vibes, change my mind
We replaced RAG with a virtual filesystem for our AI documentation assistant
Devs cheer old-school tools, skeptics side-eye real-world use, bean‑counters roast the $70k math
TLDR: A startup faked a file browser over its database so its AI can “search like a dev,” cutting wait times from 46 seconds to 100 ms with almost no extra cost. Devs loved the back‑to‑basics move, skeptics questioned if it works in messy companies, and everyone dunked on the $70k estimate.
The internet’s dev crowd just went full nostalgia after a startup ditched “RAG” (a fancy AI look-up trick) for a pretend file system called ChromaFs. Translation: instead of hoping the AI “vibes” its way to an answer, it now browses docs like a human—searching for exact text (aka grep), listing folders, and peeking at files—while secretly querying a database. The result? From a painful 46‑second wait to about 100 milliseconds, and near‑zero added costs. Cue the meme: grep > vibes.
Fans say this is a back‑to‑basics win. One commenter compared it to librarians shelving books—simple, structured, interpretable. Another hailed it as common sense: spinning up full virtual machines just to read docs is like renting a bulldozer to plant a daisy. But not everyone’s clapping. Skeptics warn it may crumble in messy, real‑world companies where information isn’t tidy; if your docs are chaos, a fake file tree might just be a prettier map of chaos. Meanwhile, the money subplot stole the show: the article’s $70k sandbox estimate got absolutely roasted—“just drop a zero?”—as the thread turned into Finance vs. Latency smack talk. And somewhere in the crossfire, the DIY crowd asked the wholesome question: can I run a tiny version on my laptop to search my emails? Internet, never change.
Key Points
- •The team replaced semantic-search-only retrieval and slow sandboxes with ChromaFs, a virtual filesystem over a Chroma vector database.
- •ChromaFs intercepts UNIX commands (e.g., grep, ls, cat, find) and translates them into vector DB queries, providing a filesystem illusion.
- •Session creation latency dropped from ~46 seconds to ~100 milliseconds, improving frontend user experience.
- •Estimated marginal compute cost per conversation fell to near zero by reusing the existing database instead of micro-VMs.
- •Technical implementation uses Vercel Labs’ just-bash (TypeScript), yargs-parser for grep flag parsing, Chroma for coarse filtering, and Redis for prefetch and in-memory filtering.