BM25 vs. Vector Search: Choosing the Right Retrieval Strategy for Production Systems

Search is deceptively complex. You can stand up Elasticsearch in an afternoon and have something that works. Whether it surfaces the right document when a user asks "how do I reset my subscription?" instead of typing "subscription reset steps" is an entirely different problem. The two dominant retrieval paradigms — BM25 and Vector Search — are both mature and production-proven. The real question is why one fails where the other succeeds, and how to combine them. BM25: The Probabilistic Workhorse BM25 scores documents using term frequency, inverse document frequency, and document length normalization. It is still the default ranking algorithm in Elasticsearch and OpenSearch. Excels at: exact keyword matching (SKUs, error codes, CLI flags), transparent debuggable ranking, sub-millisecond latency at scale, zero GPU cost. Breaks down at: vocabulary mismatch ("cancel membership" vs "terminate subscription"), semantic intent, cross-language queries, conceptual similarity. BM25 is fundamental