Build retrieval as many small, independent wikis instead of one monolithic index, then compose a RAG on demand from any subset of them. Covers per-wiki BM25 indexing, Reciprocal Rank Fusion across separate indexes, markdown-aware chunking, optional dense/graph layers, and sensitivity enforcement via the composition manifest. Use when designing a RAG over heterogeneous docs, when a single big index gives mushy results, or when different corpora need different trust levels. Triggers include "design a RAG", "retrieval is returning junk", "combine multiple knowledge bases", "hybrid search over docs", "compose a knowledge base".