How to Build Reliable Code Review Retrieval Using Semantic and Keyword Search

A technical approach to code review retrieval combines lexical and embedding-based search to avoid missing exact identifiers or policy references. Rather than merging raw relevance scores, the method uses reciprocal rank fusion to combine ranked candidates before passing a bounded set to a reranker. Each retrieval result must carry document identity, revision, access scope, and chunk identity to remain operationally useful beyond a simple demo. Authorization checks are applied before any search, and findings are only generated when evidence points to an accessible, current policy passage. The core lesson is that a high similarity score does not constitute evidence, and retrieval success does not guarantee review success.