We built a state-of-the-art RAG system for code review. In Qodo 2.4, we took most of it out.

Over the past year, the approach to code review using retrieval-augmented generation (RAG) has evolved significantly, shifting from an extensive indexing strategy to a more selective memory-based retrieval system. Initially, the focus was on creating searchable chunks of code to provide models with the necessary context for large-scale codebase Q&A and reviews, which worked well when models had limited context windows. However, improvements in agent capabilities and expanded context windows have reduced the necessity for comprehensive indexes, as models can now effectively fetch and interpret the relevant code themselves. This shift prompted a reevaluation of the retrieval system's cost-effectiveness, revealing that the maintenance of heavy retrieval infrastructure was no longer justified by its benefits. The transition to a memory-focused system, which leverages pull request histories to maintain team-specific knowledge, has proven to be more efficient and effective, reducing infrastructure demands while enhancing code review quality. This change underscores the importance of continually assessing the relevance and cost of existing systems in light of technological advancements.

No tracked trend matches for this post yet.