Your content is split into short passages and turned into vectors. A question is turned into a vector the same way, the closest passages are found, and only those passages are given to the language model with the question. The model never searches the web.
From a document to passages
A fifty-page manual is too long to hand to a model with every question, and most of it is irrelevant to any one question. So the manual is split into short passages that overlap slightly, so a sentence cut at a boundary still appears whole in one of them. Each passage remembers which document and section it came from, which keeps a passage about fees from the price list distinguishable from one about fees in the terms.
From passages to vectors
Each passage is converted into a list of numbers called an embedding. Passages about similar things end up with similar numbers, even when they use different words or different languages. This step runs on our own servers. Nothing is sent to a third party for it.
From a question to the right passages
When a visitor asks something, the question gets an embedding the same way, and the closest passages are found by comparing numbers. The search works on meaning, so wording and language can differ between question and passage, and content in the visitor's own language is preferred when it exists. Exact terms such as a product code or a name are matched as well, so they are found even when the wording around them is unusual.
What the model actually sees
Only the most relevant passages, the visitor's question and the recent turns of the conversation are given to the language model. It may answer from those passages only, in the visitor's language, and it says so when they do not cover the question. The model has no browser and no knowledge of other clients.
Why this is checkable
Because every answer starts from a retrieval, every answer can be graded. If the best passage was only loosely related, the outcome is recorded as a weak match. If the reply admits it has nothing, it is recorded as unanswered. Those records are what the dashboard shows, and they tell you which content to add next.