Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions openwiki/.last-update.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"updatedAt": "2026-08-03T01:06:15.779Z",
"updatedAt": "2026-08-11T03:50:11.351Z",
"command": "update",
"gitHead": "22f4decb23425c5c085e4723e23231116468b2da",
"gitHead": "faa2a4cab16ffbc588f0229d02f89e8a1c9f5594",
"model": "zai.glm-5"
}
29 changes: 21 additions & 8 deletions openwiki/architecture/overview.md
Original file line number Diff line number Diff line change
Expand Up @@ -147,17 +147,30 @@ See [api-integration/overview.md](../api-integration/overview.md) for authentica

Amazon Bedrock Guardrails wraps all AI interactions:
- Blocks violence, hate speech, sexual content
- Input and output filtering
- PII detection and tokenization before provider requests
- Input and output filtering with configurable thresholds
- Topic-based detection (drugs, weapons, self-harm) with telemetry-only mode

### PII Protection
### Zero-Data-Retention Inference

Automatically detects and tokenizes:
- Student names, emails, phone numbers
- Location data
- Identifiable information
AI providers do not retain inference data. Student names, emails, and other context reach the selected model byte-identical and are not stored by the provider. This is critical for tool calls such as district-data queries where the model must use real identifiers supplied by authorized users.

See `/docs/features/k12-content-safety.md` for implementation details.
### PII Detection Gates

Amazon Comprehend detection runs at two durable-content boundaries:

1. **Nexus memory writes** — Detected entity types logged as telemetry (never values). Writes proceed; detector errors are non-fatal.
2. **Published agent content** — Detected entity types logged as telemetry. Content remains unmodified.

Automatic memory extraction excludes third-party identifiers at the prompt level, not via refusal, because it runs unattended after each persisted Nexus turn.

| PII Type | Source |
|----------|--------|
| Names, emails, phone numbers, addresses | Amazon Comprehend |
| SSN, dates, ages | Amazon Comprehend |
| Student IDs (7-digit patterns) | Custom patterns |
| District-specific identifiers | Custom patterns |

See `/docs/features/k12-content-safety.md` for implementation details and filtering configuration.

## Request Flow Example

Expand Down
1 change: 1 addition & 0 deletions openwiki/quickstart.md
Original file line number Diff line number Diff line change
Expand Up @@ -98,6 +98,7 @@ The following areas have substantial existing documentation and are deferred fro
- **API Reference**: See `/docs/API_REFERENCE.md`
- **Deployment Guide**: See `/docs/DEPLOYMENT.md`
- **Testing Guide**: See `/docs/guides/TESTING.md`
- **Open Adaptive District Skill**: Skill category description in agent-platform/overview.md uses outdated "adaptive learning platform" label. Should align with simplified OAD model (six-week build cycle) when agent-platform docs next refreshed. Source: `/infra/agent-image/skills/psd-open-adaptive-district/SKILL.md`.
- **K-12 Content Safety**: See `/docs/features/k12-content-safety.md`
- **Individual Skill Docs**: See `/infra/agent-image/skills/*/SKILL.md`
- **Database Migrations**: See `/docs/database/drizzle-migration-guide.md`