If I used Gemini 2.0 for extraction and chunking to feed into a RAG that I maintain on my local network, then what sort of locally-hosted LLM would I need to gain meaningful insights from my knowledge base? Would a 13B parameter model be sufficient?