That helps a bit but doesn't counteract training. The LLM still reasons through the web query/db response with it's trained parameters