Permission parity is an index property
If a retrieval index does not enforce the same entitlements as the systems it was built from, it is a data incident waiting for the right question. Filtering results after generation is not a control, because the model has already read the content it should never have seen.
The architecture therefore carries the entitlement decision into ingestion and query time: every chunk stores the access predicate of its source, and every query is evaluated as the requesting user.
Evidence before fluency
An enterprise answer without a source is unusable in a dispute. Synthesis is constrained to cite retrieved passages, and the interface shows the passage rather than only a reference, so a reviewer can judge the answer without re-running the query.
- Store stable identifiers and versions so a citation still resolves after the source changes
- Surface retrieval confidence and coverage, not only the generated text
- Log the retrieved set with the answer for audit and later evaluation
Refusal as a first-class outcome
Systems that cannot refuse will invent. The architecture defines refusal thresholds per use case, routes refused cases to a named queue, and treats refusal rate as an operating metric rather than a defect — a falling refusal rate with unchanged content is a signal to re-check evaluation, not a success.
Operating the platform
- Re-index on source change, with lineage from chunk back to document version
- Re-run the evaluation set on every model, prompt, chunking or ranking change
- Track cost per answered question alongside quality, so scale decisions stay informed
Licence and citation
Free to read, quote and cite with attribution to SJAIN Tech. Redistribution of the PDF in modified form is not permitted.
Cite as: SJAIN Tech (Sjain Ventures Ltd), “Retrieval architecture for regulated enterprises: entitlements, evidence and refusal”, 30 June 2026. https://tech.sjain.io/research/retrieval-architecture-for-regulated-enterprises