A knowledge base is not training: what changes and what doesn't

Giving an assistant your documents doesn't teach it anything. It looks them up when a question comes in and reads the matching passages. The documents sit in a library beside the model, not inside it.

Two different things

Training is how a language model gets its general skills. It's done by the company that makes the model, on a huge pile of text, before you ever use it. AWS points out that this data¹ is static and introduces a cut-off date on what the model knows.

A knowledge base is a separate library of your own material that the assistant can search. The model doesn't absorb it. In Anthropic's words, the approach pulls relevant information out of the library and adds it to what you typed.²

So the model gets a few relevant pages slipped in front of it each time. Wikipedia says this encourages the model to prioritize the supplied data over its pre-existing training knowledge.³ "Encourages" is the right word. It's a nudge, not a lock.

What this means in practice

Changing a file changes future answers. You don't wait for the model to be retrained. Wikipedia quotes a tech publication saying that, rather than retraining, all that's needed is to augment the external knowledge base with the updated information.³ Keeping a knowledge base current covers what you still need to do by hand.

Deleting a file is a library job, not a retraining job. Whether the tool really removes the old copy right away is something to check in its documentation.

It's cheaper than retraining. AWS says the computational and financial costs of retraining models for organization-specific information are high, and presents retrieval as a way to add your information without the need to retrain the model.¹

It doesn't make the model flawless. The model still writes the final words, and it can still get things wrong. Wikipedia notes that RAG reduces the need for frequent model retraining but does not remove it entirely.³ Reasons answers still go wrong are in Why an assistant still gets answers wrong.

A quick way to tell which one you're dealing with

Ask yourself where the knowledge lives.

• If it's in a library you can open, edit and delete from, it's a knowledge base.
• If it's baked into the model, you can't edit it. Only the maker can change it.

If a tool says it can "train on your data," read its help pages closely to see whether it means uploading files to be searched or something else. Makers document these separately.

Why the difference matters to a small business

Staff sometimes worry that an uploaded price list has been "taught" to the assistant forever, or that a mistake in a file is now permanent. With a knowledge base, the file is a reference in a library, and references can be replaced.

What does carry over is anything you've put in the library. If a document is wrong, the assistant may repeat it. If a document holds private details, those details can be fetched and shown to whoever can ask. What to put in a knowledge base goes through what belongs and what doesn't.

Where your files are stored, and whether the maker uses them for its own purposes, are separate questions. Those depend on the maker's terms, so read them for your own plan.