Is RAG actually less expensive than fine-tuning for small datasets?

I’m currently evaluating the best approach for a customer support bot that needs to answer questions based on around 500 technical manuals. The texts are dense and heavily jargon-filled, which makes me nervous about standard LLM hallucinations. My initial instinct was to fine-tune a 7B parameter model using Llama 3. However, after looking at the costs associated with GPU hours for training and the time required for data cleaning and formatting, I’m starting to question if that’s the right move. I’ve read mixed opinions on whether Retrieval-Augmented Generation (RAG) is superior for smaller, static datasets like this. Specifically, I’m worried about the latency and cost of running vector searches on every single query versus the upfront cost of fine-tuning. Has anyone actually benchmarked these two approaches for a dataset of this size? I want to know if the accuracy gain from fine-tuning justifies the higher operational complexity and potential for overfitting, or if a well-constructed RAG pipeline with a good embedding model like BGE M3 is sufficient and more cost-effective in the long run. I tried a quick proof of concept with LangChain and a local instance of Mistral 7B. The RAG setup was surprisingly easy to deploy, but the answers sometimes missed the nuance of the specific error codes mentioned in the manuals. Does anyone have advice on how to improve retrieval precision without resorting to full fine-tuning? I’d appreciate any insights from people who have gone through this exact bottleneck.