introducing knowledge bases. greptile maintains detailed docs about how your codebase works, and uses them during review. for each part of the codebase, it maintains a running list of past bugs. this way, the next time it reviews that part of the code, it knows what to look for.... greptile catches 12.2% more bugs with knowledge bases enabled! now available to all users :)show more

reinventing RAG from first principles

Giving this out for free is insane, could’ve been a standalone product

Gamechanger, exctied to test

@ycombinator I haven’t finished to understand the Graph yet. But congrats for the launch!

the interesting bit is that review history becomes a living prior. the hard part is keeping that prior from going stale: when a bug is fixed or an invariant changes, old “what to look for” guidance needs provenance, otherwise context becomes another source of false confidence.

Where tokens get used the most is the real deal. Cool stuff mate!

I’m curious how you choose what from the knowledge base actually makes it into a review. If you have docs, past bugs, etc for the whole repo, how do you figure out which pieces are worth showing the agent for a change?

@ycombinator So yc are very interested in agents nowadays seems logical :))

nice release!

bug history as context is such a good primitive. every production mistake basically makes the next review slightly less stupid...

So interesting, I need to try it now

This is the real unlock, an agent that remembers what already went wrong beats a smarter model that reviews every file from zero.

the 12.2% improvement is measured across real codebases. This suggests that the knowledge base approach has practical, measurable value—not just theoretical appeal. The ability to track past bugs per codebase segment is a form of "attention" that mimics human expertise.

huge release, this is going to be very valuable to so many teams!

it did a great job of documenting our 1M+ loc codebase with 6 frontends and a bunch of backend services, really cool

lovely