4
Fine-tuning beat RAG for our internal docs, and I'm tired of pretending otherwise
We spent 3 months building a RAG pipeline with vector search over our repair manuals, and it still gave techs mismatched answers maybe 30% of the time. Last month I finally tried fine-tuning a small 7B model on just 400 cleaned-up examples from those same manuals. The fine-tuned model answered correctly on 9 out of 10 test questions, while RAG only hit 6. Sure, fine-tuning took more setup and we have to retrain every time a manual changes, but for our narrow use case it's just plain better. I know everyone's pushing retrieval these days because it's cheaper to update, but has anyone else found fine-tuning more reliable for a specific, stable document set?
0 comments
Log in to join the discussion
Log In0 Comments
No comments yet
Be the first to share your thoughts on this discussion.