CORE GUIDE
Freshness, authority and source priority
The most similar passage is not necessarily the source the system should trust.
Mental model
Evidence selection needs a trust policy: relevance answers 'is this about the task?', while authority and freshness answer 'should this source govern the decision?'
Why it matters
Knowledge systems often contain duplicates, old policies and summaries derived from stronger primary sources. If retrieval ranks only by semantic similarity, a stale FAQ can outrank the current operating procedure and produce a confidently wrong answer.
01
Add authority to relevance
Represent source type, owner, effective date, version and supersession when possible. Retrieval can use these signals as filters or ranking features, and the final answer can retain provenance so operators can see which source governed the claim.
02
Example: policy changed last month
An old FAQ says refunds are allowed for 30 days; the current policy says 14. The old page has more user-like language and wins vector similarity. An authority policy marks the current policy as controlling and either filters or heavily down-ranks the legacy FAQ.
Common failure modes
- Ranking by similarity while ignoring effective date.
- Letting derived summaries outrank primary policy documents.
- Failing to retire or mark superseded knowledge.
Engineering heuristics
- Encode authority and freshness as first-class metadata.
- Prefer primary sources when claims affect actions or compliance.
- Expose provenance in evaluation and debugging.
Takeaways
- 01Similarity answers a different question from trust.
- 02Knowledge quality depends on lifecycle and authority, not only embeddings.
- 03A retrieval pipeline should be able to explain why one source outranked another.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.