more commonly you use approximate KNN vector search with LLM based embeddings, which can find many fitting documents bm25 and similar would never manage to the

more commonly you use approximate KNN vector search with LLM based embeddings, which can find many fitting documents bm25 and similar would never manage to the tricky part if to properly combine the results