Update README.md

This commit is contained in:
PSBigBig 2025-09-01 17:57:07 +08:00 committed by GitHub
parent f16b114376
commit 4843318dcc
No known key found for this signature in database
GPG key ID: B5690EEEBB952194

View file

@ -1,62 +1,70 @@
# Enterprise Knowledge Governance — Global Fix Map
A compact hub to make enterprise data safe and auditable across RAG, agents, and long-running workflows. Use this folder to define the policy layer and route symptoms to the exact repair page. No infra change required.
A compact hub to keep enterprise knowledge **safe, auditable, and policy-true** across RAG, agents, and long-running workflows.
Use this folder to define the policy layer and route symptoms to the exact repair page. No infra change required.
---
## What this folder is
- A minimal but complete **policy layer** for RAG and agent stacks.
- Guardrails that prevent leakage and region violations.
- Copy-paste contracts for tagging, retrieval fences, and audits.
- Acceptance targets you can actually measure on live traffic.
---
## When to use this folder
- You have mixed-sensitivity corpora and must stop accidental leakage.
- Regional data residency is a contract requirement.
- Stale SOPs or outdated policies keep showing up in answers.
- Legal retention vs. developer convenience creates drift.
- You need verifiable access trails and regulator-ready exports.
- You have mixed-sensitivity corpora and must stop accidental leakage.
- Regional data residency is a contract requirement.
- Stale SOPs or outdated policies keep showing up in answers.
- Legal retention vs developer convenience keeps drifting.
- You need verifiable access trails and regulator-ready exports.
---
## Orientation: pages and what they solve
| Page | What it solves | Typical symptom |
|------|----------------|-----------------|
| [access_control.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/access_control.md) | Role, tenant, region, sensitivity intersections at retrieval time | Answer cites a snippet the user cannot see |
| [audit_and_traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/audit_and_traceability.md) | Immutable trails for who asked, what was cited, why it was allowed | Regulator asks for proof and you cannot produce it |
| [compliance.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/compliance.md) | End-to-end policy surface and controls | Policy defined in docs but not enforced in code |
| [compliance_audit.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/compliance_audit.md) | Audit checklists and export packs | You cannot prove who accessed PII last quarter |
| [data_residency.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/data_residency.md) | Region pinning for shards, embeddings, logs | Cross-region egress appears in billing or logs |
| [data_sensitivity.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/data_sensitivity.md) | Sensitivity model and redaction gates | PII or secrets slip through after parsing |
| [knowledge_expiry.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/knowledge_expiry.md) | Freshness and deprecation workflow | Bot answers with outdated SOP or revoked policy |
| [retention_policy.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/retention_policy.md) | TTL queues, deletion attestations, proof of purge | Items that should be gone still show up in runs |
---
## Acceptance targets
- Zero unauthorized citation of PII or restricted snippets in production evals.
- Policy tags present on ≥ 0.95 of onboarded documents.
- Residency violations equal 0 across seven days of traffic.
- Retention SLA respected for 100 percent of expired items within 24 hours.
- All retrievals carry a trace record with `citations`, `ΔS`, `λ_state`, `policy_eval`.
## Quick routes to pages
- Access control and role fences
→ [access_control.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/access_control.md)
- Audit trail and trace schema
→ [audit_and_traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/audit_and_traceability.md)
- Compliance overview and controls
→ [compliance.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/compliance.md)
- Compliance audit checklist and exports
→ [compliance_audit.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/compliance_audit.md)
- Data residency and region pinning
→ [data_residency.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/data_residency.md)
- Data sensitivity model and redaction gates
→ [data_sensitivity.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/data_sensitivity.md)
- Knowledge expiry and policy freshness
→ [knowledge_expiry.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/knowledge_expiry.md)
- Retention policy and deletion jobs
→ [retention_policy.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/retention_policy.md)
- Zero unauthorized citation of PII or restricted snippets in production evals
- Policy tags present on ≥ 0.95 of onboarded documents
- Residency violations equal 0 across seven days of traffic
- Retention SLA respected for 100 percent of expired items inside 24 hours
- Every answer carries a trace with `citations`, `ΔS`, `λ_state`, `policy_eval`
---
## Map symptoms → structural fixes
- Wrong snippet shows up from a restricted area
→ [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md)
→ Contract the payload with [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md)
- Prompt or tool step bypasses the policy and leaks PII
→ [prompt-injection.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/prompt-injection.md)
→ Lock tool schemas in [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md)
| Symptom | Open this |
|--------|-----------|
| Wrong snippet shows up from a restricted area | [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md) · [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md) |
| Prompt or tool bypasses policy and leaks PII | [prompt-injection.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/prompt-injection.md) · lock tool args in [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md) |
| Sensitive text survives parsing and chunking | [chunking-checklist.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/chunking-checklist.md) · [ocr-parsing-checklist.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/ocr-parsing-checklist.md) |
| Live runs drift from policy or regions | [ops/live_monitoring_rag.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/ops/live_monitoring_rag.md) · [ops/debug_playbook.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/ops/debug_playbook.md) |
- Sensitive text survives parsing and chunking
→ [chunking-checklist.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/chunking-checklist.md)
→ OCR and export checks in [ocr-parsing-checklist.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/ocr-parsing-checklist.md)
- Live runs drift from policy or regions
→ [ops/live_monitoring_rag.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/ops/live_monitoring_rag.md)
→ [ops/debug_playbook.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/ops/debug_playbook.md)
---
## 60-second setup checklist
1) **Tag the corpus**
Attach `sensitivity`, `region`, `owner`, `retention_tier` to every doc. Enforce schema with [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md).
2) **Fence retrieval**
At retrieve time require the intersection of `{tenant_id, role, region, sensitivity}` and drop non-matching snippets. Verify with [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md).
Require the intersection of `{tenant_id, role, region, sensitivity}` at retrieve time and drop non-matching snippets. Verify with [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md).
3) **Pin residency**
Keep embeddings and shards in the source region. Block cross-region egress unless policy allows. See [data_residency.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/data_residency.md).
@ -67,9 +75,12 @@ A compact hub to make enterprise data safe and auditable across RAG, agents, and
5) **Audit everything**
Emit `actor`, `question`, `citations`, `ΔS`, `λ_state`, `policy_eval`, `region` for each answer. Route to an immutable sink. See [audit_and_traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/audit_and_traceability.md).
---
## Copy-paste policy probe for your LLM step
```txt
You have TXTOS and the WFGY Problem Map loaded.
You have TXT OS and WFGY pages loaded.
Question: "{user_question}"
Context carries fields {sensitivity, region, retention_tier, owner} for each snippet.
@ -77,11 +88,39 @@ Context carries fields {sensitivity, region, retention_tier, owner} for each sni
Do:
1) Enforce cite-then-explain. Refuse if a cited snippet breaks role or region.
2) Return {"citations":[...], "policy_eval":{"allow":true|false,"reason":"..."}, "ΔS":0.xx, "λ_state":"→|←|<>|×"}.
3) If blocked, give the minimal fix and the exact WFGY page to open.
3) If blocked, output the smallest change to comply and the exact WFGY page to open.
````
---
## FAQ
**Q: We tag sensitivity at ingest but leaks still happen. What now?**
A: Enforce the tag at **retrieve time** too. Use access intersections `{tenant_id, role, region, sensitivity}` and verify with [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md).
**Q: Our vendor requires EU residency. How do we prove compliance?**
A: Pin vectors, shards, and logs to EU regions and export an egress report weekly. Steps in [data\_residency.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/data_residency.md).
**Q: Outdated SOPs keep showing up. Where is the crack?**
A: You need a freshness control. Mark deprecated docs and enforce a deny-list at retrieval. See [knowledge\_expiry.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/knowledge_expiry.md).
**Q: Legal wants proof of deletion after TTL.**
A: Implement TTL queues and write deletion attestations to an immutable sink. Details in [retention\_policy.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/retention_policy.md).
**Q: We cannot reconstruct who saw which PII last quarter.**
A: Add the audit contract from [audit\_and\_traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/audit_and_traceability.md). Every answer should log `actor`, `citations`, `ΔS`, `λ_state`, `policy_eval`, `region`.
**Q: Prompt injection bypassed our fences.**
A: Tighten tool schemas and add role-ordered templates. See [prompt\_injection.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/prompt-injection.md) and the contracts in [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/data-contracts.md).
**Q: Coverage is high but restricted snippets sometimes get ranked.**
A: Coverage is not policy. Add **policy\_eval** to the trace and drop candidates before rerank. Use [access\_control.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/GlobalFixMap/Enterprise_Knowledge_Gov/access_control.md).
**Q: How do we set pass or fail before release?**
A: Use a governance gate: zero restricted citations, zero residency violations, coverage ≥ 0.70, ΔS ≤ 0.45, and audit completeness 100 percent for a 7-day canary.
---
### 🧭 Explore More
| Module | Description | Link |