Data Governance & Privacy

Data governance and privacy are critical for shipping enterprise-grade LLM features.

1) Principles

• Least data: send the minimum needed; redact PII/secrets before model calls.
• Purpose limitation: only use data for the stated task; log consent where relevant.
• Separation: env-separated keys, storage, and logs; avoid prod data in dev.

2) Data Handling Pipeline

• Ingress filters: reject unsupported file types, excessive size/duration.
• Redaction: email/phone/account IDs; configurable patterns per tenant/region.
• Classification: tag sensitivity level (public/internal/secret/PII).
• Egress filters: scrub outputs that include sensitive source snippets unless allowed.

3) Storage & Retention

• TTL for temp artifacts (transcripts, intermediate parses).
• Encryption at rest & in transit; rotate keys.
• Avoid storing raw prompts/responses if they contain PII; hash user IDs.

4) Regionality & Residency

• Route by region; keep data in-region where required.
• Per-tenant policies: some tenants opt out of training or logging.
• Document data flows for compliance reviews.

5) Access Control & Auditing

• RBAC/ABAC on datasets and tools; tenant_id filters on retrieval.
• Audit logs: who accessed what, when; config changes to prompts/models.
• Break-glass procedures for emergency access with approvals.

6) Third-Party Models/Tools

• Provider DPA and data retention settings; disable training on your data if possible.
• Sanitize tool inputs/outputs; allowlist domains/APIs.
• For self-hosted models: patch cadence, network egress controls, isolated VPC.

7) Safety Filters

• Prompt-injection detection for user-supplied docs/inputs.
• Content moderation (toxicity/abuse); refuse unsafe requests.
• Output filters for secrets/credentials patterns.

8) Minimal Checklist

• Data classification + redaction before LLM.
• Region-aware routing; tenant filters on retrieval.
• Encrypted storage with TTL; audited access.
• Provider settings reviewed (no training, retention limits).