A knowledge base is a maintained collection of documents, records or notes that an AI system searches through before answering a question, so its response draws on that specific material rather than only on what the underlying model already learned during training.
The mechanism behind most knowledge bases is usually called semantic search, and it works on meaning rather than exact wording. OpenAI’s own developer documentation makes the point directly: because its retrieval tooling matches on meaning, “this flexibility makes semantic search a powerful technique for querying knowledge bases of any size” — a search for one question can surface an answer that shares almost none of its exact words. The same documentation notes that this kind of retrieval “is powered by vector stores, which serve as indices for your data” — the searchable form a knowledge base takes once it has been prepared for an AI system to use.
None of that changes what applies once personal information sits inside the knowledge base itself. If a mortgage brokerage indexes old client files so a tool can search them, the same obligations described under de-identification and training data still apply to that material — putting a file into a knowledge base does not exempt it from them.
Consider a mortgage brokerage’s own procedures — how to handle a switch mortgage, what documents a self-employed applicant needs, how a particular lender wants a file packaged. Turned into a knowledge base, a support tool can search those procedures directly and answer a question about a self-employed applicant using the brokerage’s own wording, instead of guessing from whatever the underlying model picked up about mortgages in general during training. The knowledge base only stays useful for as long as someone keeps it current — a procedure that changed last month but was never re-indexed will keep getting served up as if it were still correct.
See also: what is a dataset, what is training data, what is an API.
Deciding what belongs in a knowledge base, and keeping it current, is a build decision — custom-ai-solutions covers what a system needs from the material behind it and who owns keeping that material accurate.
