Every AI chat product says you can train it on your own content. The word is wrong, and it is worth being pedantic about, because what people expect from "training" is not what they get.
Nothing about the model changes when you add a document. Your text is indexed, and the model reads a few passages of it at the moment somebody asks a question.
What the word makes people expect
The fourth line is the one that matters commercially. Teams budget for retraining that never happens, or worse, avoid updating a document because they think it is expensive.
The three ways your content gets in
There are only three, and picking the wrong one for a given source is the most common setup mistake.
Whichever you pick, the same thing happens next: the text is cut into passages and each one becomes a list of numbers.1 That list is what gets searched when a question arrives.
What training would actually mean
There is a real technique the word describes. Fine-tuning adjusts the model's weights on examples you supply, and it changes how the model writes: its format, its register, the shape of its output.
It is the wrong tool for facts. A fine-tuned model cannot tell you which page an answer came from, and a price that changes means paying to train again. Retrieval keeps the facts outside the model, where they can be edited, deleted and cited.
welcomeai.devWhat is RAG? Retrieval-augmented generation explainedWhat changes once you stop using the word
Four practical consequences, and they are the reason the distinction is worth arguing about.
- An edit is live on the next question. No retraining, no waiting, no version of the assistant that still believes last quarter's price.
- Deleting a source removes it. The passage stops being retrievable, which is not something you can say about anything baked into a model.
- A wrong answer is traceable. You read the passage behind it and see whether the document was wrong or the retrieval missed.
- Adding content is cheap and answering is not. Indexing runs at a small fraction of the per-token price of answering, so the instinct to ration uploads is backwards.
The passages were copied at index time. Removing the PDF from your drive changes nothing until you delete the source in the knowledge base and let the index rebuild.
Where to start
Connect what you already publish, upload only what lives nowhere else, and write Q&A entries for the handful of questions where the wording is not yours to improvise. Then read the answers before anyone else does.
welcomeai.devKnowledge bases: sources, indexing and what each type is forQuestions people ask
If you take one thing from the vocabulary: you are not teaching it, you are giving it something to read. Everything about how you run it follows from that.
