Freshness

Freshness is the dimension that gets no attention until it causes an incident. A document was correct when it was ingested. It was superseded three weeks ago. The retriever still finds it, the reranker still scores it well, the grounding check passes because the answer is faithfully drawn from the passage. Every metric is green and the answer is wrong.

Stale content is worse than missing content, because missing content produces "I don't know" and stale content produces a confident answer with a citation.

Measuring freshness needs the pipeline to have captured the right metadata: effective dates, expiry dates, version identifiers, supersession links, and a source-of-truth rank when multiple sources cover the same subject. With those, freshness compliance is measurable directly: the share of retrieved results that are current as of the query time, and the share of the corpus that is within its freshness window by content class.

Mechanisms:

Effective-dated retrieval. Every retrieval carries an as-of time, defaulting to now. Superseded and expired content is filtered out by default, available on request for historical questions. This is a metadata filter like any other, and it only works if supersession was recorded at ingestion.

Freshness windows by content class. Not all content ages at the same rate. Regulatory notices might need same-day ingestion; product documentation weekly; internal procedures quarterly. Define the window per class, monitor the lag between source change and index update against it, and alert on breach. This is a service-level objective on the pipeline, surfaced as a retrieval metric.

Source-of-truth hierarchy. When two sources conflict, the system has to know which wins. That ordering — authoritative system of record above derived content, official publication above internal copy, newer version above older — should be explicit, recorded as metadata, and applied at ranking time. Without it, the retriever's tie-break is whatever scored higher, which is arbitrary.

Contradiction detection. Near-duplicate detection in the pipeline finds documents that say the same thing. A related pass can find documents that should say the same thing and do not: two currently-effective sources with overlapping scope and conflicting statements. These are flagged for a human, because the resolution is a governance decision. The retriever's job is not to hide the conflict; it is to surface it.

Propagation on change. When a source document changes, every derived representation — chunks, summaries, entities, embeddings — must be regenerated or marked stale, and the old ones must stop being retrievable. This is a pipeline property, but it is a retrieval failure when it does not happen, and the retrieval metrics are where it shows up.