Why would you need a vector database when your system response time is dominated by calls to off-prem LLMs? Linear search through flat-file of embeddings. Done.