·3 min read
Fine-tuning is not grounding
Weights are a poor filing cabinet. If the answer has to match this week's notes, retrieve them. Do not hope the model memorised the PDF.
Photo: Susan Q Yin on Unsplash
Fine-tuning is for behaviour. Grounding is for facts that change. Mixing them up is how a product confidently recites last semester’s reading list.
The RAG paper is from 2020 and still describes the split: generate from retrieved evidence instead of hoping parameters stored the document.
Vendors have since productised the same idea. OpenAI’s file search docs and Anthropic’s contextual retrieval are implementation notes, not a new religion. The product decision is which corpus is allowed.
A course, a policy pack, a contract set. Those files move. Fine-tuning on Monday’s PDF means you are wrong on Thursday unless you retrain. Retrieval can be updated when the file is.
Fine-tunes still earn their keep: house style, tool use, refusal patterns. They are a poor database. Do not store the student handbook in weights.
Citations fall out of grounding if you keep the passages. They do not fall out of a fine-tune. “The model has seen this” is not a footnote.
If you need both, retrieve first. Tune for how the answer is written, not for what is allowed to be true.
Sources
The TruFyre Way
Fine-tuning is not grounding. TruFyre retrieves the live pack for facts. Fine-tunes, when we use them, are for tone.