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