Specialist-model research desk

Direct questions

RAG versus fine-tuning: which is right for proprietary data?

Frontier Development research deskSection 01 · Page 3 of 14Last updated 26 October 2026

Use retrieval when the immediate need is to supply relevant information at inference time; consider adaptation when repeatable behaviour remains deficient despite adequate context. A combined system may be appropriate, but adds components that need their own evaluation.

A support answer can fail because no document was retrieved, the wrong tenant’s document was retrieved, the right document was ignored, or the requested action was unauthorised. Do not reduce those outcomes to one “hallucination” score. Track retrieval eligibility, source relevance, answer support and action correctness separately.

For extraction, compare a specialist document service as well as a general model. Extend’s documentation describes schema-defined JSON extraction with confidence and source citations. Those features are useful comparison dimensions, not proof of accuracy; the buyer must test whether confidence predicts mistakes on its own documents.

TRACE THE EVIDENCE

Sources cited on this page

  1. Extend’s documentation →docs.extend.ai