Best architecture for a chatbot over a frequently-updated company knowledge base?

This question targets one of the most common misconceptions in applied LLMs: that **fine-tuning** is how you teach a model new facts. It isn't. Fine-tuning runs gradient updates on examples, so it shapes HOW a model behaves. It teaches tone, output format, refusal rules, and domain reasoning. It is bad at storing specific facts, because a fact crammed into weights goes stale the moment your knowledge base changes. RAG keeps the model frozen and instead retrieves relevant text at query time, then puts that text in the prompt. The model reads it and answers. When a document changes, you re-index that chunk in seconds. Nothing is retrained. For a knowledge base that updates daily, option B (RAG) is correct. You can still add a light fine-tune on top, but only for company voice and format, never as the way the model knows facts. The wrong options all try to push facts into the weights. They are slow to update, expensive, and the model starts half-remembering details and confidently making things up.

The trap in this question is the belief that fine-tuning teaches facts. It teaches behavior. So the architecture for a frequently-updated knowledge base is **RAG**, option B. Fine-tuning runs gradient updates and changes the output distribution. It is sample-efficient for style, format, and reasoning patterns. It is sample-inefficient and unreliable for facts. A specific fact is one data point, not a pattern, so the model only partially memorises it and confabulates the rest. RAG puts facts where they belong: an external index. A daily update is a re-embed of the changed chunks, done in seconds. The model reads retrieved text and grounds its answer in it. Citations and live freshness fall out for free. The distractors all fight this. Nightly re fine-tuning a 70B model is costly and still stale between runs. Per-document tokens are not a real technique. Per-document LoRA adapters explode to fifty thousand and just reinvent retrieval, minus the citations. The production pattern is RAG for knowledge plus an optional light fine-tune for voice. Behavior in weights, knowledge in the index.