tl;dr: Most RAG implementations over-engineer with embeddings and vector databases when BM25 full-text search plus LLM-based query rewriting (~$0.001/query) would solve 60% of use cases. The author recommends a tiered approach: start with full-text search, add query rewriting for vocabulary mismatches, move to hybrid search (BM25 + embedding rerank) only when data proves it's needed, and reserve full pre-embedding for massive-scale stable corpora. Key trade-offs involve data freshness, model deprecation risk, and latency—on-the-fly embedding avoids costly re-indexing when models change, while pre-embedding wins on latency but hurts on flexibility.