Add embedding caching to reduce inference costs.

Add embedding caching to reduce inference costs. #904