It would be helpful to document your architecture. Eg what kind of text search you use, eg trigrams, some sort of benchmark for searching with/without expensive embeddings.