“Do we need RAG or fine-tuning?” may come too early in an AI assistant procurement meeting. Without agreement on what the assistant gets wrong, the discussion compares technology names rather than solutions to a business problem. Start with samples of required outputs, then investigate the gap.
You may need an answer grounded in a document changed this morning, a summary in a fixed format, or both. These are different requirements. The word “training” in a proposal does not explain where knowledge resides, who updates it or how outputs are checked.
Test the business question first
Imagine a fictional supplier wanting an internal assistant to answer questions about a frequently changing catalog and turn staff requests into structured summaries. If a specification is wrong, did the assistant fail to find the right document or misunderstand it? If a summary omits a field, were instructions and examples clear initially?
The example alone does not determine the solution. The implementer should distinguish information-access problems from output-format problems and propose a limited comparison. No actual experiment or model comparison is reported here; this is a way to improve a request for proposals.
Understand the options enough to decide
With RAG, an application retrieves relevant content and supplies it alongside the question to ground generation. Fine-tuning uses training data to modify model behavior or task performance. AWS and Microsoft guidance identify changing documents and specific behavioral needs as comparison factors, with combinations possible.
Retrieving a passage does not guarantee a correct answer or citation, and index updates do not promise instant, cost-free freshness. Fine-tuning is not available for every model or an automatic substitute for current sources. A method's name cannot certify system accuracy or access permissions.
A decision sheet rather than terminology
| Illustrated need | Evidence to provide | Requested comparison |
|---|---|---|
| Changing specifications | Current and superseded documents with a distinguishing question | Source selection and effective date in the answer |
| Fixed-format summary | Varied requests and accepted output examples | Instructions and examples first, then justification for additional training |
| Role-restricted document | Business roles and permitted content boundaries | What each role can see and who reviews exceptions |
| Combined requirement | A question needing fresh information and a structured summary | Each component's contribution and maintenance cost |
This proposed sheet does not conclusively classify every project. Prioritize the simplest comparison addressing the gap: improved instructions may suffice, retrieval may be needed, or repeated behavioral needs may warrant studying fine-tuning. Ask what a more expensive option adds instead of accepting complexity as a predetermined outcome.
Separate demonstration success from operational success
Choose questions excluded from the preparation examples, including questions the documents cannot answer. In the catalog illustration, what happens when a specification disappears or two versions conflict? Request visible uncertainty or review rather than a complete-looking answer unsupported by approved information.
Also assign approval of new documents, removal of obsolete ones and review of changes against earlier cases. Purchasing a model or index does not automatically assign these jobs. Request separate estimates for data preparation, evaluation, updates and ongoing oversight, without assuming one approach is always cheaper.
Ensure samples represent real staff questions, including abbreviations and similar names, rather than polished wording alone. Request a recorded reason for rejecting each output so source errors remain distinct from wording errors. Avoid paying for model changes to compensate for an incorrect document.
What should the technical proposal justify?
Look for a described cause of the gap, a baseline option, data usable with permission, an explicit acceptance method and limitations of the selected model or service. If retrieval and fine-tuning are combined, ask which outcome would deteriorate without each component and how an agreed trial would establish that.
Review knowledge-base readiness for source preparation and AI assistant demonstration acceptance for reviewing a proposal. Discuss AI assistant development with Nextwave using work samples, without purchasing a promise that technology alone prevents errors.
References and comparison limits
The AWS comparison of RAG and fine-tuning and Microsoft's explanation of retrieval and indexes were opened on 4 November 2026. The original sheet and example involve no training, model calls or performance testing. AI assisted preparation, with information and translation review.
