feat(pii): regex-only block-output redaction + drop GLiNER/GPU image

[TheodoreSpeaks]

Summary

• Restrict the block-output PII stage to regex/checksum entities only — drop the 4 spaCy-NER types (PERSON/LOCATION/NRP/DATE_TIME). Enforced at the UI grid, the API contract (migration-safe strip, never reject → existing rules stay saveable), and the resolver (so un-migrated rules can't feed NER into execution).
• Add a Presidio skip-spaCy fast path: regex-only requests reuse a blank NlpArtifacts to bypass the spaCy NER pass entirely — the per-leaf cost that made bulk block-output redaction slow. redact_batch logs nlp=skip|full.
• Remove the GLiNER/GPU image: strip torch/CUDA/GLiNER from docker/pii.Dockerfile, delete engines.py/requirements-gliner/bench+tests, restore server.py to a self-contained CPU spaCy service (keeps the NER-only pipe-disable ~2x win + the logging-emit fix), and drop the Helm pii.engine/pii.device wiring.

Type of Change

• New feature + cleanup

Testing

• 43 unit tests pass (regex-only grid + strip, contract transform, resolver strip); type-check, lint, check:api-validation:strict clean; py_compile server.py; helm lint + helm unittest (10/10) pass.
• ⚠️ The server skip-spaCy path is logic/syntax-verified but not runtime-tested locally (Presidio isn't installable here). Validate on staging: a regex-only block-output redaction should log nlp=skip and run dramatically faster than a NER (nlp=full) request.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
docs	Skipped		Jul 15, 2026 11:49pm

[cursor]

PR Summary

Medium Risk
Block-output masking no longer covers NER entity types (behavior change for in-flight redaction), and the skip-spaCy fast path is not covered by removed integration tests—staging validation of nlp=skip is advised. GLiNER deployments must move to spaCy-only images.

Overview
Block-output PII redaction is now regex/checksum only: spaCy-NER types (PERSON, LOCATION, NRP, DATE_TIME, plus ORGANIZATION) are stripped from that stage in the UI picker, API schema transform, retention resolver, and save/hydration paths. Existing rules that only selected NER for block outputs are migration-safe (entities removed, stage disabled rather than rejected).

The Presidio service adds a regex-only analyze fast path that skips the spaCy NLP pass when the request lists only non-NER entities and has no positive score_threshold, using prebuilt blank NlpArtifacts; /redact_batch logs nlp=skip|full. Analyzer setup is inlined in server.py (large spaCy models, VIN + locale recognizers, disabled non-NER spaCy pipes, fixed request timing logs).

GLiNER/GPU support is removed: engines.py, GLiNER deps/bench/tests, torch/CUDA/small spaCy models from the PII image, and Helm pii.engine / pii.device wiring.

^{Reviewed by Cursor Bugbot for commit 47ca743. Bugbot is set up for automated code reviews on this repo. Configure here.}

[TheodoreSpeaks]

@greptile review

[greptile-apps]

Greptile Summary

This PR makes block-output PII redaction regex-only and simplifies the PII service. The main changes are:

• Filters spaCy NER entities from block-output policies in the UI, API contract, and runtime resolver.
• Adds a spaCy-skipping path for eligible regex-only requests.
• Removes GLiNER, Torch, CUDA, and related image configuration.
• Updates the Presidio service and Helm deployment for CPU-only spaCy.

Confidence Score: 5/5

This looks safe to merge.

• The latest changes cover spaCy-backed entities such as ORGANIZATION through the live recognizer registry and an explicit fallback set.
• Requests with a positive score threshold use the full NLP path, preserving context-aware analysis.
• No blocking issue was found in the updated code.

Important Files Changed

Filename	Overview
apps/pii/server.py	Adds the regex-only fast path, derives spaCy-backed entities from the recognizer registry, and preserves full analysis for positive score thresholds.
apps/sim/lib/guardrails/pii-entities.ts	Defines and filters the spaCy NER entity set used by block-output policies.
apps/sim/lib/api/contracts/primitives.ts	Normalizes block-output policies by stripping NER entities and disabling empty stages.
apps/sim/lib/billing/retention.ts	Applies the regex-only restriction while resolving stored PII policies.
docker/pii.Dockerfile	Removes the GLiNER and GPU layers and packages the standalone spaCy service.

_{Reviews (4): Last reviewed commit: "fix(pii): include ORGANIZATION in app-si..." | Re-trigger Greptile}

[greptile-apps]

Greptile Summary

This PR makes block-output PII redaction regex-only and removes the GLiNER deployment path. The main changes are:

• Filters spaCy NER entities from block-output policies in the UI, API contract, and runtime resolver.
• Adds a blank-artifacts fast path for regex-only Presidio requests.
• Consolidates the PII service into a CPU spaCy implementation.
• Removes GLiNER dependencies, image layers, tests, and Helm settings.

Confidence Score: 4/5

Thresholded Presidio requests can miss context-sensitive PII and need fixes before merging.

• Empty artifacts bypass the tokens and lemmas used for context scoring.
• Disabled spaCy pipes weaken context scoring on every analysis path.
• The UI, contract, resolver, image, and Helm changes otherwise remain consistent.

apps/pii/server.py

Security Review

The changed Presidio paths can expose PII in thresholded requests. Empty artifacts and globally disabled language pipes prevent context enhancement, allowing context-sensitive pattern matches to fall below the requested threshold.

Important Files Changed

Filename	Overview
apps/pii/server.py	Adds regex-only analysis helpers and globally disables spaCy pipes, creating false-negative paths for thresholded requests.
apps/sim/lib/guardrails/pii-entities.ts	Defines the shared NER entity set and filters block-output choices and hydrated policies.
apps/sim/lib/api/contracts/primitives.ts	Strips NER entities from block-output policies and disables stages left empty.
apps/sim/lib/billing/retention.ts	Enforces regex-only block-output policies during runtime resolution.
apps/sim/ee/data-retention/components/data-retention-settings.tsx	Restricts block-output selections and sanitizes policies before saving.
docker/pii.Dockerfile	Removes GLiNER, Torch, CUDA, development-test dependencies, and deleted source files.
helm/sim/templates/deployment-pii.yaml	Removes obsolete engine and device wiring while retaining generic environment variables.

_{Reviews (2): Last reviewed commit: "feat(pii): restrict block-output redacti..." | Re-trigger Greptile}

[TheodoreSpeaks]

@greptile

[TheodoreSpeaks]

@cursor review

[TheodoreSpeaks]

@greptile

[TheodoreSpeaks]

@cursor review

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 47ca743. Configure here.}