RAG Performance Gains Start Before Ranking, Not During It

Most RAG optimization advice focuses on the ranking stage, but a new perspective argues that filtering irrelevant vectors before ranking begins can significantly reduce latency, memory use, and noisy context. Boundaries already present in a request — such as tenant ID, product, language, or document version — can serve as pre-ranking scopes to narrow the candidate set early. Applying these constraints upstream means fewer vectors to score, less memory to read, and fewer irrelevant inputs reaching the reranker or language model prompt. However, the approach requires that the boundary be stable, known at request time, and genuinely exclusive of unrelated data — otherwise a smaller scope may simply be fast because it is incorrect. KoutenDB, an open-source document and vector database written in Nim, implements this concept through application-defined "rings" that confine vector retrieval to a relevant, authorized subset of the corpus.