Start with retrieval
Chunk the corpus, embed it, cite the chunks, refuse when nothing matches. You can update a PDF without a training run. This is the default in our AI builds.
Fine-tune when evals demand it
Classification into your taxonomy, or a format the base model will not hold. We keep a frozen eval set so we can say whether the fine-tune earned its keep.
Questions
Can we do both?
Yes. Retrieve for facts, fine-tune for form. That costs more and needs a reason in the quote.