Add embedding caching to reduce inference costs.#140