Security
Data Flow Diagrams
How data flows through OGuardAI, where PII exists, what crosses the trust boundary, and what each component sees
Data Flow Diagrams
For the complete data flow documentation covering all 8 operations (Transform, Rehydrate, RAG, Proxy, Streaming, File Processing, Image/OCR, Batch), see the canonical reference:
Architecture: Data Flow Diagrams
Security-Relevant Summary
All data flows follow the same trust boundary model:
- Trusted Zone (OGuardAI runtime, plus the Python NER sidecar when deployed): Raw PII exists here in memory during request processing, and never crosses to the untrusted zone
- Untrusted Zone (LLMs, tools, logs, vector stores): Only semantic tokens
{{type:id:cap}}and safe metadata cross the boundary, plus any value an operator explicitly lists in a policywhitelist(the shipped strict policies have an empty whitelist) - Session state: Encrypted with AES-256-GCM, client-held (sealed mode), in-process (memory mode), or encrypted at rest in a shared Redis (redis mode).
- Revocation table: Stores HMAC-SHA-256 hashes only, no raw PII
- Audit events: Entity types and counts only, no raw values logged
For detailed trust boundary analysis, see the Security Whitepaper.