The Relationship Blindspot in Vector Retrieval Every optimization in this series so far — better chunking, reranking, query rewriting, CRAG — works within a fundamental assumption: retrieval is about finding similar text. But a whole category of questions doesn't fit that assumption: Questions that require reasoning across multiple entities. Consider: "bge-large-zh-v1.5 and bge-reranker-v2-m3 both come from the same organization. What is that organization, and what role does each model play in a RAG pipeline?" Vector search will find documents mentioning BAAI or BGE — that part works. But retrieval only returns text chunks. The LLM then has to figure out "these two models are both from BAAI" by reading disconnected paragraphs. The relationship isn't in the retrieval result; it's implicit in the text. Or consider: "Trace the evolution of retrieval quality evaluation from basic RAG to CRAG." This requires placing RAG → Rerank → Self-RAG → CRAG in order along a progression chain.…