RAG vs fine-tuning
Separate the need for current company knowledge from changes to response behaviour before selecting an approach.
Author：WeTech technical team · Updated： (Hong Kong time (GMT+8))
Evaluate RAG when answers need current documents and supporting sources. For consistent format, style or task behaviour, establish a baseline before considering fine-tuning. The approaches can be combined, but fine-tuning does not automatically provide current information, document permissions or verifiable citations.
When to use this guide
For buyers offered a company-trained model who need to distinguish knowledge, retrieval and behaviour problems.
Compare the approaches
| Need | RAG consideration | Fine-tuning consideration |
|---|---|---|
| Changing knowledge | Source connections and refresh | Still needs a freshness mechanism |
| Evidence | Retrieved passages and accessible sources | Does not replace citation checks |
| Behaviour consistency | Test prompts and output rules first | Representative training and held-out evaluation |
| Ongoing cost | Indexing, retrieval, answers, upkeep | Training, hosting, evaluation, retraining |
Locate the failure first
Classify failures as missing retrieval, obsolete sources, unsupported claims, format errors or unauthorised access. A retrieval problem may not require a model change, and format errors may respond to simpler controls. Compare an explainable change against the same evaluation set.
Use a common baseline
Hold back questions covering absent answers, conflicting sources and revoked permissions. Compare completion, source support, latency, review effort and cost per task. Require data handling, model version and failure handoff details rather than only selected successful demonstrations.
Separate responsibilities in a combined design
In a combined design, manage knowledge sources, permissions, training data and output evaluation separately. Do not put all company documents into training simply for convenience. Confirm permitted use, deletion arrangements and responsibility when a model is retired.
Illustrative scenario
Hypothetical design: a support team retrieves a daily updated manual and applies a fixed response structure to drafts. Evaluate fine-tuning separately only if behavioural errors persist and suitable training examples exist. No delivered product capability is implied.
Frequently asked questions
Does RAG eliminate wrong answers?
No. Test retrieval, source support, permissions and unanswerable cases.
Will fine-tuning know our latest policy?
Not automatically. Freshness and verification need a separate design.
Must we choose only one?
No. Use baseline tests to show what each additional component solves.
Brand and public case scope
WeTech Consulting, also called WeTech or WeTech AI HK, provides AI and IT consulting in Hong Kong. WeProperty HK and WeCredit HK are self-developed products. Their cases support the observable features listed: property, estate and transaction-reference entry points for WeProperty; submission, review and public-data masking for WeCredit. WeCredit is not a TU credit-rating service or a final determination of facts.
WeProperty HK · WeCredit HK (case studies in Traditional Chinese)
Reference sources
Official documents support the technical, risk and measurement concepts. Checklists, estimation methods and illustrative scenarios are WeTech planning recommendations, not official prices or client results. Check the supplier’s current plan and terms. Sources reviewed: (GMT+8)
Related guides
Discuss your project
Share your workflow, systems and goals. We will review scope before discussing deliverables, acceptance and a quotation.
Request an initial consultationExplore this service