Modern Hybrid Search

Modern search tools, should leverage the semantic capabilities of LLMs without sacrificing the control of traditional keyword based search. Hybrid systems enable this.

In a modern hybrid search system, a user's keywords and filters can be combined with an optimized query that is then processed in two parallel branches. One interfaces with a vector search system, retrieving a semantic candidate pool. The other, a traditional system, returns a respective candidate pool. These results can then be combined and passed to a re ranker module that orders the pool by relevance. These ordered results can then be returned to the user.

The traditional arm of this system will typically use BM25, which addresses weaknesses in the classic TFIDF formula. Recently BM25s has been released which can offer up to a 45x speed improvement on some benchmarks.

In the following slides and video walkthrough, we discuss the design for a modern hybrid search system and provide a deep dive on BM25s.