Enterprise knowledge: validate and carry forward
- Source
- From the Staff AI System Design course
- By
- Sourav Banerjee
- Last updated
- 8 Nov 2026
Self-Test and Memory Card
- What is immutable? Each tenant/document/version object identity and its content, not the current publication pointer.
- What does a tombstone prevent? Older observed events from reviving visible deleted content.
- When is a relevant hit unusable? Wrong tenant, unreadable policy, deleted or mismatched current version.
- Why re-fetch trusted text? A caller can forge a search-hit body even when it names an existing allowed document.
- What does the local fixture not prove? Real connector freshness, distributed publication, model faithfulness, or physical erasure.
The 60-second card: durable source event; immutable version; coherent publication; rank within trusted scope; recheck before context; cite exact passage; refuse without evidence. Review after about one, seven, and 21 days.
- Explain ingestion, publication, and answer grounding in two minutes without notes.
- Draw the durable Queue/Worker path and the trusted context gate.
- Defend lexical versus semantic/hybrid retrieval under this corpus.
- Diagnose stale, deleted, revoked, and partially published evidence with a concrete signal.
- Run/modify the version, relevance, and negative-access tests; learner completion remains unassessed.
Record the Change and Its Limits
ADR-003 records the versioned retrieval choice. The stage snapshot preserves actual evidence.
A later T18.5 may add stateless app replicas and caching when this path faces broader enterprise traffic. It must preserve authoritative version/permission checks and citeable source identities; a cache hit cannot grant access. No advanced governance, embedding deployment, or evaluation platform is added here.
Primary References
- SQLite FTS5: full-text matching and ranking used by the runnable index.
- OWASP authorisation guidance: trusted per-resource checks, not prompt-based access.
- Elastic semantic text: an implementation example of embedding/chunking concerns for the alternative, not an installed dependency.
Sources checked 2 November 2026. Numbers and fixture text are authored examples. Execution evidence and untested integrations are in validation.
Key Takeaways
- Each tenant, document and version object and its content is immutable. The current publication pointer is not.
- A relevant hit is unusable when it belongs to the wrong tenant, the reader is not allowed, or the document is deleted or no longer the current version. Re-fetch stored text, because a caller can forge a hit body.
- The local fixture does not prove real connector freshness, distributed publication, model faithfulness or physical erasure.
- A cache hit can never grant access. Any caching added for scale must keep the authoritative version and permission checks and citeable source identities.