How an AI Code Review Bot Learned to Shut Up

Greptile's AI code review bot initially faced criticism for generating excessive comments. To address this, they experimented with prompt engineering and having the LLM evaluate its own comments, but these methods proved ineffective. Their breakthrough came from vectorizing past comments, clustering them in a vector database, and filtering out new comments similar to those previously downvoted. This approach boosted the developer address rate from 19% to over 55%, significantly reducing LLM noise.