Brilliant approach! Using LLMs offline to train embeddings is such a clever way to get the intelligence without the latency hit. Curious about the training data pipeline - did you find semantic search worked better than keyword matching for domain suggestions?