Why Your RAG System Doesn't Need Embeddings

Research comparing seven retrieval strategies across 17,000+ chunks found that agentic approaches matter more than choosing between BM25 and vector search in RAG systems. Testing on D&D rulebooks and French RPG materials showed simple BM25 with an AI agent achieved near-perfect results (10/10 scores), while sophisticated vector embeddings offered minimal improvement since the LLM already handles semantic query reformulation. Agents using multiple queries consistently outperformed single-pass systems, and switching language models had more impact than changing search engines entirely.