The problem both approaches solve
A general-purpose model knows a great deal about language and very little about any one business. It does not know this season's range, the returns policy, the claims the legal team has approved or the way the brand writes, and when asked it fills those gaps with what is typical, as how language models generate text explains. The answers look right and are wrong in the details that matter most to customers.
Both approaches aim to close that gap using the brand's own First-Party DataData a business collects directly from its own customers, visitors and users.: product data, help content, policies, approved copy. They differ in where the brand's knowledge lives. With retrieval, it stays in the documents and is consulted fresh for each answer. With fine-tuning, it is absorbed into the model during an extra round of training. That difference drives almost every row of the comparison.
How grounding and retrieval work explains the retrieval mechanism in detail. Fine-tuning, in general terms, means taking an existing model and training it further on a set of examples, usually pairs of inputs and ideal outputs, so that it becomes more likely to produce outputs like those examples. It changes the model's tendencies; it does not give it a reliable store of facts.
Side by side
Retrieval vs Fine-tuning
| Row label | Retrieval | Fine-tuning |
|---|---|---|
| Where brand knowledge lives | In documents and data, consulted at question time | Inside the model, as patterns learned from examples |
| Freshness | Updating a document changes the next answer | Changes need another round of tuning |
| Accuracy on specific facts | Good when the right passage is found; answers can cite it | Unreliable for specifics; the model can produce plausible wrong details |
| Traceability | Answers can point to their source passages | No source to point to; output reflects training patterns |
| Voice and style | Depends on prompts, examples and instructions | Can build a consistent style or format into the model |
| Setup effort | Organising, cleaning and structuring the document library | Assembling a large set of high-quality examples and running training |
| Ongoing maintenance | Keeping documents current, deduplicated and owned | Re-tuning as content or the base model changes, and re-testing |
| Removing information | Remove or archive the document | Hard to remove what a model has learned without re-tuning |
| Data exposure | Documents are stored and searched by the tool | Examples are used in training; a tuned model may reproduce them |
| Exit | Documents remain the brand's own and can move to another tool | A tuned model is often tied to one vendor's platform |
The pattern is consistent. Retrieval keeps the brand's knowledge in a form the brand controls, can update, can cite and can take elsewhere. Fine-tuning changes how a model behaves in ways that are harder to inspect, update and reverse, in return for style and format that hold without long prompts.
When retrieval is the answer
Retrieval is the default for anything that is a fact and anything that changes.
- Products and catalogues. Names, specifications, compatibility, availability. A Product FeedA structured file of a store's products, prices and availability used to build shopping and catalogue ads. or product information system that the tool can draw on keeps answers in step with what is on sale.
- Prices, offers and terms. These change too often to train into a model, and errors are costly.
- Policies and help content. Returns, delivery, warranties, eligibility. Customers act on these answers.
- Approved claims. The substantiated list that legal or compliance has signed off; retrieval lets an answer stick to it and point to it.
- Anything that must be traceable. Where a reviewer, a customer or a regulator might ask where an answer came from, a cited passage is the answer.
Retrieval's weaknesses are in the documents, not the model: stale versions, duplicates and passages that make no sense on their own. Fixing those is ordinary content work, and it also improves the help centre and site for human readers. The academy exercise on grounding a help assistant for a bicycle retailer practises sorting what an assistant must know into facts and form, choosing the mechanism for each and turning a vendor's promise of a model trained on the brand into questions.
When fine-tuning is worth considering
Fine-tuning earns its cost in narrower situations, mostly about form rather than facts.
- Stable, distinctive style at volume. A team producing a large amount of output in a voice that prompts and examples do not capture well enough, measured rather than assumed.
- Consistent format. Outputs that must follow a precise structure every time, such as a fixed set of fields or a house pattern for titles, where prompt instructions keep slipping.
- Shorter prompts at scale. Where long instructions and examples in every prompt add cost or slow responses at high volume, building them into the model can help.
- Specialised classification. Sorting text into the brand's own categories, with many labelled examples available.
Even then, three conditions should hold. Retrieval and well-built prompts, of the kind in prompt patterns for marketing, have been tried and tested on real tasks. The gap that remains has been measured with the method in testing prompts and models. And the team has enough good, approved examples to train on and the capacity to re-tune and re-test when the base model changes. Most fine-tuned setups still use retrieval for facts; tuning shapes how the model writes, retrieval supplies what it writes about.
An outdoor furniture brand compares two vendor proposals
Ellerbrook Outdoor Furniture, an invented maker of garden tables and benches, gets two proposals for a product question assistant. Vendor A offers retrieval over its 140 product pages, care guides and delivery policy. Vendor B offers a model "trained on your brand" using 3 years of product copy. The team builds a test set of 50 real customer questions and runs both. Vendor A answers 41 correctly with a citation and says it does not know for 6. Vendor B answers in a closer match to the brand voice but gives wrong dimensions or finishes in 14 answers, 9 of them for products discontinued in the past year. Vendor B also confirms that removing a discontinued product would mean re-tuning. The team chooses retrieval, adds 3 approved voice examples to the instructions, and plans to re-test in 6 months.
What 'trained on your brand' should mean in a contract
Vendors use phrases like "trained on your brand", "learns your voice" and "your own AI" for very different things: retrieval over uploaded documents, a fine-tuned model, saved instructions and examples, or a mix. The mechanism determines the risks, so the first job is to find out which it is, and the second is to make sure the agreement reflects it. The questions below are practical due diligence, not legal advice; the disclosure and data privacy guide covers what vendor terms on inputs and training should say, and the tool evaluation guide covers how to run the wider procurement.
- Which mechanism? Retrieval, fine-tuning, stored instructions or a combination, and for which features.
- Where does the material go? Where documents and examples are stored, for how long and who at the vendor can access them.
- Is it used for anyone else? Whether the brand's material is used to train or improve models offered to other customers, and whether that can be switched off.
- How is it updated? How a changed price, policy or product reaches answers, and how quickly.
- How is it removed? How a document or example is deleted, and for a tuned model, what removal actually involves.
- What happens when the base model changes? Whether the vendor re-tunes and re-tests, and whether the brand is told.
- How is quality shown? Whether the vendor will run the brand's own test set, and report results.
- What can the brand take away? Whether documents, examples, instructions and any tuned model can be exported if the contract ends.
How to decide
- List what the tool must know. Separate facts (products, prices, policies, claims) from form (voice, format, structure).
- Put facts in retrieval. Organise the documents and data that hold them, give each an owner and a review date, and choose a tool that retrieves and cites.
- Handle form with prompts and examples first. A voice guide, approved examples and format instructions, stored in a shared prompt library.
- Test on real tasks. Use a test set and a rubric, and measure the share of usable answers as a KPIA metric tied directly to an objective. for the tool.
- Consider fine-tuning only for a measured, stable gap in form. And only if the team can supply the examples, accept the maintenance and agree acceptable terms on data and exit.
Common mistakes
- Fine-tuning to teach facts. A tuned model holds an impression of its examples; prices and specifications drift into plausible errors.
- Accepting "trained on your brand" without asking how. The phrase covers several mechanisms with very different risks.
- Blaming the model for retrieval failures. Wrong answers from a retrieval tool usually trace to stale, duplicated or badly structured documents.
- Tuning before trying prompts. Voice and format problems are often solved by a clear guide and good examples.
- Ignoring exit. A tuned model tied to one platform is hard to move; documents are not.
- Forgetting removal. Information built into a model is hard to take out; decide what goes into training with that in mind.
- Skipping the test. Without a test set from real questions, the choice rests on a demo.
