GUIDE · AIunipedia Community
Retrieval-augmented generation supplies selected external information at request time. Fine-tuning changes model parameters using training examples. They address different needs and can be combined.
If a delivery policy changes every week, ask how the application will retrieve the current approved policy. If the requirement is a consistent specialised output style, ask whether better instructions and examples are sufficient before considering model adaptation.
A retrieval prototype will still depend on document chunking and the evidence it finds. Try the community document-assistant project before treating either approach as the right answer for every task.
A decision worksheet
Write down the behaviour you need to improve, how frequently the facts change, who may access them and how success will be measured. Separate failures into missing knowledge, poor retrieval, unclear instructions, unsuitable output behaviour and application errors.
Do not fine-tune merely because the model missed a fact it was never given. Equally, do not assume retrieval will fix a model that consistently ignores a required output format.
Compare approaches fairly
Use the same representative questions and evaluation rubric for the baseline, a retrieval prototype and any adapted model. Include maintenance effort: updating source documents, refreshing an index, preparing examples and evaluating regressions.
A defensible decision identifies the failure being addressed and the cheapest approach that improves it on held-out cases. It does not begin with a preferred technology and work backwards.
Discussion: What specific problem are you trying to solve with RAG or fine-tuning, and what evidence would change your choice?
References: RAG evaluation cookbook and parameter-efficient fine-tuning.
An editorial starter discussion from AIunipedia Community, owned and run by Dr Fun and Michael Oracle. Share your own experience and help improve the explanation.
Continue exploring
Document Chunking for RAG: Size, Overlap and Boundaries · LoRA, QLoRA and Quantization Explained
Browse: RAG & Knowledge Systems · All learning paths
Tip: open a related discussion in a new tab to keep your place here. On a phone, touch and hold the link for your browser’s options.