Skip to content

security_governance.md

Latest commit

 

History

History
37 lines (29 loc) · 2.55 KB

security_governance.md

File metadata and controls

37 lines (29 loc) · 2.55 KB
Raw

Security And Governance

RAGScaleGuard handles the kind of data that enterprises normally classify as confidential: emails, support tickets, customer context, internal decisions, source-code discussions, meeting transcripts, and policy documents.

Data Handling Requirements

  • Keep raw enterprise corpora in data/, which is ignored by git.
  • Keep generated reports in reports/, which is ignored by git.
  • Do not log raw document text in CI.
  • Redact reports before sharing outside the authorised team.
  • Treat document IDs as sensitive when they encode customer names, ticket IDs, employee names, or repository paths.
  • Treat diagnostic artefacts as sensitive because they can expose ranking decisions, source relationships, document freshness, and authority signals.

Governance Controls The Toolkit Should Support

  • Source authority scoring: final specs and resolved tickets should outrank stale chat.
  • Freshness scoring: current artefacts should be preferred where facts change over time.
  • Metadata-aware constraints: project, customer, source type, date, owner, and department should be usable as retrieval signals.
  • Conflict detection: contradictory evidence should be surfaced rather than hidden behind a single generated answer.
  • Citation support checks: generated claims should not be returned when citations do not support them.
  • Diagnostic artefacts: high-risk findings should explain why the guard flagged the case and which evidence was involved.
  • Auditability: reports should show which retrieval strategy produced which evidence set.
  • Evaluation integrity: held-out evaluation data must not be used for model training.

Enterprise Deployment Checklist

  • Run evaluations inside a controlled working directory or container.
  • Pin dependencies and run pytest, mypy, and ruff in CI.
  • Decide whether hosted rerankers or answer evaluators are allowed to receive document text.
  • Keep HTTP retrieval adapters behind internal networking and server-side authorisation.
  • Pass adapter secrets through environment variables rather than committed config files.
  • Record corpus version, corpus size, retriever settings, embedding model, reranker model, and top-k.
  • Review report artefacts before publishing external summaries or case studies.
  • Decide whether diagnostic artefacts may be exported to third-party observability, ticketing, or adviser tools.
  • Redact or hash sensitive document IDs before sharing reports outside the authorised team.

See reporting_schema.md for the report and artefact fields that should be reviewed during export approval.