Mental model
Treat every retrievable unit as an evidence-bearing artifact, not an anonymous string. Its text is useful only when the system can also identify its source, version, owner, access policy, transformation history, and verification time.
Theory
Provenance answers four questions: what entity existed, which activity transformed it, which agent was responsible, and how the resulting artifact relates to earlier versions. In a retrieval system, the chain normally begins with a source record, continues through parsing and chunking, and ends with an indexed chunk carrying stable identifiers.
This metadata must survive reprocessing. A citation to a current web address is insufficient when the indexed content came from an older revision. Store a source ID, source version or digest, ingestion run, transformation version, chunk coordinates, access scope, and timestamps. At generation time, preserve the retrieved chunk IDs beside claims so the final answer can be audited.
Alternatives and trade-offs
Full event-level lineage is powerful but expensive. Snapshot-level lineage is simpler and may be enough for low-risk corpora. Content hashes detect changes but do not explain ownership or authorization. Human-maintained source records improve meaning but require governance. Choose the minimum chain that can support access review, freshness decisions, deletion, and claim verification.
Failure modes and misconceptions
URLs are not immutable evidence. Embedding metadata is not an access-control system. Reindexing without a recorded transformation version makes regressions difficult to reproduce. Copying a citation supplied by a model does not prove that the cited passage supported the claim.
Knowledge check
Which identifiers must travel with a chunk so a reviewer can reproduce why it appeared in an answer?
Decision scenario
A policy assistant stores the policy document ID, approved revision, section path, ingestion run, parser version, chunk span, tenant scope, and source digest. A replacement revision invalidates derived chunks and creates a new evidence chain instead of silently overwriting history.
Learning outcomes
- Explain Knowledge Provenance as a system mechanism rather than a slogan.
- Compare its alternatives, trade-offs, and production failure modes.
- Apply the concept to a decision and identify evidence that would validate it.
Trade-offs
Using Knowledge Provenance can improve capability or control, but it also introduces cost, latency, complexity, and failure modes that must be measured against an explicit objective.