Runs the same embedding and vector search the chat uses, without calling the chat model. Use it to tune chunk size, top-k and the score threshold.