Guide
Adapters versus retrieval: two ways to give a model your material
Retrieval looks up documents when a question arrives and hands them to the model. An adapter changes the model itself through training. Retrieval suits facts that change often and need a visible source. An adapter suits terminology, formats, and ways of working that should show up in every answer. They are not exclusive.
Ask Gee about this guide
Gee is an AI demo and can be wrong. Don’t share confidential information. Conversations aren’t stored.
How retrieval works
Retrieval searches a set of documents when a question comes in. It places the passages it finds into the prompt, and the model answers from them.
Its strength is freshness and sources. Replace a document and the next answer reflects it, and the passage used can be shown to the reader. Its limit is that an answer is only as good as the search. The model itself still does not know your terms or your formats.
How an adapter works
An adapter is a smaller set of weights trained on your approved material. The base model is not retrained. The adapter carries your terminology, formats, and patterns into every answer without a passage being supplied each time.
Its limit is that the model learns at training time. When the material changes, you train a new version. A trained model also does not, by itself, point to the document an answer came from.
When each one fits
- Retrieval fits documents that change often, answers that must show a source, and large collections that no one wants to retrain on.
- An adapter fits stable terminology, house formats, and judgment patterns that should be present everywhere.
- Both fit when a team needs consistent behavior and current, citable documents.
A quick test for choosing
Ask which failure you can least afford. If it is a stale or unsourced answer, favor retrieval. If it is inconsistent terms or the wrong format, favor an adapter. If it is both, plan for both.
How Gpodz handles the adapter side
Gpodz Compute is designed around three layers: a shared base model, a company adapter, and a role adapter. Each adapter is versioned separately from the base. Your reviewers test a new version on examples you choose, and it is activated only when you approve it. A previous version is designed to be restored without retraining the base.
Adapter changes are designed to be recorded in an append-only log, with each entry chained to the one before it.
This site does not describe a retrieval component. If you need answers tied to current documents, say so during scoping.
Questions on this topic
Can I use both?
In general, yes. The two techniques solve different problems and are often combined. For Gpodz Compute, ask us during scoping what is available for your case.
What happens when a policy changes?
With retrieval, you replace the document. With an adapter, the new material goes into a new adapter version, which your reviewers test before it is activated. The previous version stays available to restore.
Which is easier to review?
Retrieval lets a reviewer see the passage behind an answer. An adapter is reviewed by testing a version on examples you choose, with versions you can compare and roll back.
Tell us which fits your material.
Bring an example of the documents and the questions. We will say what is available today.