Why Your RAG System Doesn't Need Embeddings

After benchmarking BM25, vector, and hybrid search across 2 corpora and 7 agents: the LLM does the semantic work that embeddings are supposed to do. A good agent with BM25 scores 10/10 where a single-pass vector query scores 8. Ingestion quality and model choice matter more than your search engine. Start with BM25.