Separate model knowledge from current external information

A model can explain stable concepts without proving what is true today. Prices, office holders, policies, product versions, schedules and incident ownership can change after training or between requests. Current claims need current evidence.

Classify the information need

The offline foundations lab labels three authored tasks.

from foundation_cases import freshness_case

result = freshness_case()
assert result["requires_retrieval"] == 2
assert result["model_memory_used_as_current_source"] is False
assert result["review_date"] == "2026-09-20"
print(result["tasks"])

A question about the definition inside a supplied 2024 contract can use that document. Today’s API price and the current incident owner require external current sources. The check verifies routing logic, not any retrieved fact.

Make freshness part of the contract

For each task, define acceptable source types, maximum age, effective date and fallback when evidence is unavailable. A current policy answer might require the approved policy repository and an effective version. A price answer should cite the vendor’s official page and observation time.

Retrieved content also needs validation. Check source authority, publication/effective date, version and relevance. Two current sources can conflict; surface the disagreement or route it to an owner instead of averaging text.

Keep claims traceable

Return citations or source IDs next to the fields they support. Store retrieval time and query with the run where policy permits. If no eligible evidence exists, abstain and explain what source is needed. Avoid phrases like “as of today” unless the workflow actually checked today’s source.

Knowledge cutoffs are not the only issue. A model may misremember older facts or blend sources. Use supplied evidence and independent checks for high-impact claims even when the subject is historically stable.

The Generative & Agentic AI course connects freshness routing to RAG, grounded answers and tool contracts.

Exercise

Classify 30 questions as stable concept, supplied-document fact, current external fact or private operational fact. Define source and freshness rules, then test missing, stale and conflicting evidence without allowing model memory to fill the gap.

Continue learning

This article is part of the LLM fundamentals and prompt design sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Compress a conversation without losing task constraints.
• Continue with Handle an LLM refusal in an application workflow.

Reference: NIST Generative AI Profile.