Developing a RAG based llama-index with a local model (llama2) and default embedding. The answers vary fairly much asking the same question repeatedly. How would you make the answer(s) a bit more deterministic / consistent for the same query