Instruction-like content
Supported patterns that appear to direct an AI, override a task or influence how retrieved content is handled.
Document security for AI knowledge bases
Check source documents before RAG ingestion.
Check incoming source documents before they are prepared, chunked or indexed. Follow each version’s decision and retrieve the current Approved copy for your knowledge base as sources arrive and change.
- Outside exact versionOutside source document
- Organise the recordChoose a Collection for the workflow
- Completed Standard ScanCheck and route
Three destinations
- ApprovedLibrary → current eligible version → RAG knowledge base
- Needs reviewInvestigate, replace or recheck
- QuarantineKeep this exact version out
The Collection keeps all three outcomes organised. Membership does not grant approval or retrieval.
No decision: The scan could not complete, so no security decision was made. Retry is available and no new scan credit is consumed.
The ingestion problem
Once a source is indexed, its influence can spread.
A knowledge base can retrieve content from the same source across many prompts, users and tasks. Hidden, active or instruction-like material in an outside document may survive preparation and later appear in retrieved context.
Make source admission a separate decision from parsing, chunking and indexing. Keep the exact version and its scan record identifiable while the wider RAG system continues to own permission, provenance and runtime controls.
See the wider RAG security modelWhat deserves attention
Check the source before its content becomes reusable context.
Focus on the document-level conditions that should influence whether this exact version proceeds.
Hidden and machine-readable surfaces
Supported content or features that may not be obvious during visual review but can still affect downstream processing.
Version drift
An updated, regenerated or replaced source is a new exact version and needs its own decision before ingestion.
Your document workflow
Separate document approval from RAG ingestion.
Use .mdSiren as the document checkpoint and record, then keep preparation and system controls in the wider pipeline.
- 1Record the source, owner, purpose and permission for its intended use.
- 2Choose the Collection for the knowledge base or project; assignment is explicit rather than inferred from the document.
- 3Submit the exact source file for a Standard Scan before parsing, chunking or indexing.
- 4Follow Approved, Needs review or Quarantine, together with the findings and coverage.
- 5Keep the source record organised in its Collection while pending, Approved, Needs review or in Quarantine; membership never makes a version eligible for retrieval.
- 6Retrieve only the current eligible Approved exact version from the Library into your authorised preparation and ingestion process.
When a replacement is accepted, it becomes current immediately and needs its own scan; the earlier version does not remain available as a fallback while the replacement is pending or held. Pro API access supports document submission, status checks and retrieval of eligible current Approved exact-version copies under the same rules.
Before ingestion
Know exactly what you are admitting.
A short admission record makes it easier to connect indexed content to the source and decision that allowed it in.
- Confirm ownership, permission, sensitivity and intended audience.
- Scan the source version that will actually be prepared for ingestion.
- Assign it to the right Collection for the knowledge base or project.
- Review findings, completed checks and coverage before acting on the status.
- Link the Library version and scan record used for ingestion to the downstream source record.
- Treat replacements, re-exports and policy changes as reasons to check again.
This document checkpoint does not replace access control, data classification, RAG evaluation or system-wide security.
Related resources
Before documents become AI context
Put a document checkpoint before your knowledge base.
Check each source version and retrieve the current Approved copy for your knowledge base. Pro API submission and eligible Approved exact-version retrieval are available with Pro.