1.3.1. The Knowledge Cutoff Problem
💡 First Principle: Every base model has a training cutoff date and no access to private or post-cutoff information. This isn't a bug to be patched — it's an inherent property of how models are trained, which is precisely why retrieval-at-request-time exists as a standard pattern.
Ask a model about a contract signed last week or your company's internal refund policy, and it has three options: refuse, admit it doesn't know, or — worst case — fabricate a plausible-sounding answer. The fabrication risk is what makes the cutoff dangerous rather than merely limiting. Grounding converts an unanswerable or hallucination-prone question into a reading-comprehension task: retrieve the contract clause or the policy paragraph, hand it to the model in the prompt, and ask it to answer from that. The model is excellent at reading comprehension; it's unreliable at recalling specifics it was never given.
⚠️ Exam Trap: Scenarios describing "the model gives outdated answers" or "the model doesn't know our internal data" are grounding/RAG problems, not model-selection or fine-tuning problems. Swapping to a newer or larger model doesn't add your private data; only retrieval does.
Reflection Question: Why is a fabricated-but-plausible answer more dangerous in production than an honest "I don't know," and how does grounding address the root cause?
