Software
RAG leaves the model unchanged and supplies passages from a named library at question time. Fine-tuning changes the model using examples, which is a different project with its own data, cost, and evaluation. Novatelia’s AI solution can scope retrieval. It does not publish a fine-tuned or proprietary model, and it does not state an accuracy figure.
Revised 2026-10-10. Published by Novatelia. No individual author is named, because no personal credential is claimed for this explanation.
Key points
- Changing documents is easier in a retrieval library than in model weights.
- Fine-tuning needs a dataset and an evaluation the project can defend.
- Neither method makes an answer automatically correct.
Definition
Use RAG when the facts live in documents that change. Consider fine-tuning only when the task is a stable behavior that examples can teach and the customer is allowed to train.
How it works
Retrieval indexes documents and fetches them per question. Fine-tuning runs a training job, produces a model version, and then still needs tests against cases the project cares about. The two can be combined, which is two scopes, not one slogan.
Architecture
- RAG: corpus, index, retrieval, model call.
- Fine-tuning: dataset, training job, held-out checks.
- A decision record of which method the requirement actually needs.
Use cases
- Policy text changes often and must be quoted.
- A stable classification task has labeled examples and a permitted training path.
- A buyer is being offered fine-tuning when a document index would answer the question.
Comparison
- RAG
- Documents supplied at question time. The model weights stay as issued.
- Fine-tuning
- A training job changes a model. Not a published Novatelia model.
Selection guidance
- If the facts change, start with retrieval.
- Do not commission training to avoid organizing the documents.
- Keep infrastructure for a training job on the AI infrastructure page, and only if that job is in scope.
Questions
Will Novatelia fine-tune a model as part of AI solutions?
Only if a written scope includes a dataset, a permitted base model, and an evaluation. The published page does not include a trained Novatelia model or a success rate.
Which is more accurate?
Neither has a published accuracy rate. Retrieval is limited by the documents. Training is limited by the examples. The scope should say how the result will be checked.
Related
More in Software
This page defines a term so a project can be scoped. It is not a certification, a test result, or a price. Partnerships are stated only on the partners page. The published address and telephone are the company contact, not a service-level commitment.
Describe the requirement
An RFQ saves a browser draft. It does not publish a price or issue a quotation reference.