Comment by DanielVZ

They went from indexing with embeddings + LLM to just using a biased embedding for their use case. This should save them most of their costs.