Key takeaways
- Inventory source authority, owner, jurisdiction, effective dates, review cadence, and removal rules before ingestion.
- Carry source, version, chunk, and transformation provenance into retrieval and the final answer.
- Test retrieval separately from answer quality using dated, ambiguous, conflicting, and withdrawn material.
- Monitor freshness lag and deletion propagation; a successful index job is not proof of current answers.
01
Retrieval makes knowledge replaceable; it does not make it true
A sales assistant quotes last year’s pricing after the website changed. A support agent finds the regional policy but misses the global exception. The index completed without errors, yet the answer is stale. RAG joins a model with retrieved material; the original research highlights provenance and updating knowledge as important motivations, not guarantees that every retrieved passage is authoritative or current.
Begin with a source register. For each repository, record the accountable owner, intended audience, territory, language, authority level, creation and effective dates, review cadence, superseding relationship, retention rule, and whether the material may be shown to the user.
Swipe to compare every column
| Layer | Version to record | Freshness failure to test |
|---|---|---|
| Source | Document ID and effective date | Superseded policy remains active |
| Transform | Parser, OCR, and cleanup version | Table or qualifier disappears |
| Chunk | Boundaries and parent source | Exception separated from rule |
| Index | Embedding model and build ID | Deletion or update fails to propagate |
02
Make provenance travel with the passage
W3C PROV describes entities, activities, and agents involved in producing an object so people can assess quality, reliability, and trustworthiness. In practical terms, a retrieved chunk should still know its source URL or record, version, section, effective period, transformation, access policy, and responsible owner.
Do not let a copied paragraph become an orphan. Preserve headings, table relationships, footnotes, and nearby exceptions where they change meaning. Give the application enough metadata to prefer current authoritative material and to explain which source supported a consequential answer.
03
Evaluate the retriever before grading the prose
Build queries that exercise exact facts, paraphrases, acronyms, multiple markets, conflicting versions, withdrawn material, and questions with no supported answer. Measure whether the right evidence appears in the candidate set, at what rank, with the required context. Then evaluate whether the answer uses it faithfully and declines when support is absent.
Keep a frozen regression set and a rolling set drawn from real, redacted questions. A prompt change cannot repair evidence the retriever never supplied. An embedding change may improve average relevance while breaking the few policy questions that matter most.
04
Operate freshness as a service level
Define maximum lag by source class: an emergency statement may need minutes; a brand guide may tolerate days. Trace publication, ingestion, transformation, indexing, cache expiry, retrieval, and answer generation. Alert on age and missing deletions, not only job failure.
Provide a visible “last reviewed” signal, correction route, and safe fallback. When two current sources conflict, route the conflict to an owner instead of asking the model to invent hierarchy. The honest answer may be that the knowledge base cannot yet support the decision.
Primary sources and further reading
Use the source material to validate details against your own context and current platform configuration.
- Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
- W3C: PROV overview
- NIST: Generative AI Profile
This field note follows the XenGrowth editorial policy: primary sources where available, visible limitations, material review dates, and no invented first-hand experience.
Stay with the problem



