Should you fine-tune an LLM on your documents to fix wr...

Fine-tuning is often a misguided response to RAG errors. It teaches the model style and format, not factual knowledge. Before scheduling training, perform an oracle-context ablation—providing the model with the correct document—to see if it can answer correctly. If the retriever’s recall@k is low, no amount of generator fine-tuning will improve the end-to-end accuracy beyond that ceiling.