1.2. Grounding and Retrieval
💡 First Principle: A language model is a reasoning engine, not a database. It generates the most plausible continuation of text — and "plausible" is not "true." Grounding closes that gap by injecting verified, retrieved content into the model's context at answer time, so plausibility gets anchored to evidence.
The stakes are concrete: an ungrounded agent asked about your return policy will confidently describe a return policy — a fluent blend of every policy it saw in training. For an enterprise agent, that is not a quirk; it is a liability. Grounding is the single mechanism behind the entire knowledge half of this exam — every knowledge source in Phase 2 (Copilot connectors, Power Platform connectors, Azure AI Search) is just a different answer to the question "where should retrieval look?"
The mental model to carry: the model is a brilliant analyst with no filing cabinet. Retrieval is the assistant who fetches the right three documents and puts them on the analyst's desk before asking for an answer.
⚠️ Common Misconception: RAG fine-tunes or retrains the model on your data. It does not — retrieval happens at query time, the model's weights never change, and your data is not memorized into the model. That is precisely why RAG-based knowledge updates instantly when the source changes, while fine-tuning would require retraining.
