Embedding generation is computationally intensive. So when ML models are run on CPU, it takes longer to generate embeddings, which then holds up the firebase function