feat(middleware): Use NER models instead of Regex for PII filtering by richiejp · Pull Request #10160 · mudler/LocalAI

fix(router): score classifier production-readiness

richiejp

committed

feat(backend): auto-size batch to context for embedding and rerank models

richiejp

committed

fix(gallery): raise arch-router scoring output cap via parallel:64

richiejp

committed

feat(pii): inbound encoder/NER detection tier

richiejp

committed

build(nix): add C++ gRPC to the dev shell for the llama-cpp backend

richiejp

committed

feat(llama-cpp): TokenClassify RPC for openai-privacy-filter NER

richiejp

committed

feat(config): add token_classify known_usecase for the PII NER tier

richiejp

committed

feat(gallery): add privacy-filter-multilingual token-classify model

richiejp

committed

refactor(pii): NER-centric PII filter; remove the regex tier

richiejp

committed

docs(pii): gallery pii_detection policy + NER-centric docs

richiejp

committed

feat(ui): NER-centric PII editor; drop the regex pattern UI

richiejp

committed

refactor(pii): drop deprecated pii.patterns / pii.ner config keys

richiejp

committed

feat(gallery): NER (token_classify) install filter + OpenMed icon

richiejp

committed

fix(gallery): correct privacy-filter HF repo to LocalAI-io; plain-text description

richiejp

committed

fix(backend): size batch to context for token_classify (NER) models

richiejp

committed

fix(config): don't paint FLAG_CHAT onto score/token_classify GGUF models

richiejp

committed

fix(middleware): list only PII-coverable models on the Filtering page

richiejp

committed

feat(pii): instance-wide default detector + per-usecase default-on

richiejp

committed

feat(pii): Middleware UI for the instance-wide default PII policy

richiejp

committed

feat(pii): extend request filtering to completions, embeddings, edits & Ollama

richiejp

committed

docs(pii): document instance-wide default policy & widened endpoint coverage

richiejp

committed

feat(pii): flag enabled-but-no-detector models as a no-op on the Filtering page

richiejp

committed

fix(ui): ModelMultiSelect added a model per keystroke; compact chip layout

richiejp

committed

fix(ui): close the detector dropdown gap; use the standard Toggle for PII defaults

richiejp

committed

feat(pii): drop the per-usecase default-on selector; cloud models stay default-on

richiejp

committed

feat(pii): observability for NER detections — backend trace, confidence, debug logs

richiejp

committed

feat(pii): pattern detector tier — bounded restricted-regex secret matching

richiejp

committed

feat(pii): model-editor pattern/builtins editor + Traces row

richiejp

committed

feat(pii): detector-models table with inline toggles in middleware UI

richiejp

committed

fix(modeladmin): show and persist map fields like entity_actions in the editor

richiejp

committed

fix(pii): label detections by their source (ner vs pattern)

richiejp

committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(middleware): Use NER models instead of Regex for PII filtering#10160

feat(middleware): Use NER models instead of Regex for PII filtering#10160
richiejp wants to merge 31 commits into
mudler:masterfrom
richiejp:feat/pii-ner-tier

Commits on Jun 7, 2026

fix(router): score classifier production-readiness

feat(backend): auto-size batch to context for embedding and rerank models

fix(gallery): raise arch-router scoring output cap via parallel:64

feat(pii): inbound encoder/NER detection tier

build(nix): add C++ gRPC to the dev shell for the llama-cpp backend

feat(llama-cpp): TokenClassify RPC for openai-privacy-filter NER

feat(config): add token_classify known_usecase for the PII NER tier

feat(gallery): add privacy-filter-multilingual token-classify model

refactor(pii): NER-centric PII filter; remove the regex tier

docs(pii): gallery pii_detection policy + NER-centric docs

feat(ui): NER-centric PII editor; drop the regex pattern UI

refactor(pii): drop deprecated pii.patterns / pii.ner config keys

feat(gallery): NER (token_classify) install filter + OpenMed icon

fix(gallery): correct privacy-filter HF repo to LocalAI-io; plain-text description

fix(backend): size batch to context for token_classify (NER) models

fix(config): don't paint FLAG_CHAT onto score/token_classify GGUF models

fix(middleware): list only PII-coverable models on the Filtering page

feat(pii): instance-wide default detector + per-usecase default-on

feat(pii): Middleware UI for the instance-wide default PII policy

feat(pii): extend request filtering to completions, embeddings, edits & Ollama

docs(pii): document instance-wide default policy & widened endpoint coverage

feat(pii): flag enabled-but-no-detector models as a no-op on the Filtering page

fix(ui): ModelMultiSelect added a model per keystroke; compact chip layout

fix(ui): close the detector dropdown gap; use the standard Toggle for PII defaults

feat(pii): drop the per-usecase default-on selector; cloud models stay default-on

feat(pii): observability for NER detections — backend trace, confidence, debug logs

feat(pii): pattern detector tier — bounded restricted-regex secret matching

feat(pii): model-editor pattern/builtins editor + Traces row

feat(pii): detector-models table with inline toggles in middleware UI

fix(modeladmin): show and persist map fields like entity_actions in the editor

fix(pii): label detections by their source (ner vs pattern)