No Canadian regulator has published a definition of retrieval-augmented generation, or RAG. The clearest working description available comes from a vendor whose own product implements it. Microsoft’s documentation puts it plainly: “Retrieval-augmented generation (RAG) is a pattern that extends LLM capabilities by grounding responses in your proprietary content.” In practice: instead of relying only on what a large language model picked up during training, a RAG system searches a separate, current set of documents at the moment of the question and hands the model the relevant passages to work from before it writes an answer.
“Your proprietary content”, in the vendor’s own phrase above, is frequently a business’s own client records. The moment an organization connects a generative AI tool to those records instead of relying only on the model’s original training, it has usually created exactly the situation PIPEDA’s accountability principle already covers. Schedule 1 of the statute states it directly: “An organization is responsible for personal information in its possession or custody, including information that has been transferred to a third party for processing. The organization shall use contractual or other means to provide a comparable level of protection while the information is being processed by a third party.” Building a retrieval layer over a customer database does not take that data outside PIPEDA’s reach — it changes who is touching it and how, and the organization that built the retrieval system remains accountable for it regardless.
A mortgage brokerage builds a retrieval layer over its own file notes so a staff chat tool can answer “what’s the status of this file” using the brokerage’s real, current records instead of whatever a general-purpose model happened to learn in training. Because the passages the system retrieves for any given question usually name a specific client and their file details, the accountability obligation quoted above attaches to that retrieval step itself, not only to the original file it was pulled from. Grounding an answer in real records also narrows — without eliminating — the risk of a hallucinated answer, since the model has something real to draw from rather than only a pattern it learned months earlier.
See also large language model and AI hallucination.
This is one term in a plain-English glossary on how AI actually works and where it fits in a Canadian business.
