Skip to content

CHANGELOG.md

Latest commit

 

History

History
149 lines (127 loc) · 7.55 KB

CHANGELOG.md

File metadata and controls

149 lines (127 loc) · 7.55 KB
Raw

Changelog

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning. While the version is below 1.0.0, minor releases may include breaking changes; each such change is called out under a Changed or Removed heading.

Unreleased

Changed

  • An explicit score_threshold of 0 is now rejected. When a request sets score_threshold_set = true, score_threshold must be > 0; an explicit 0 (or negative) returns InvalidArgument instead of silently falling through to the service default. Previously the field was documented as "0 = include everything" but the service treated it as "use default", so an explicit 0 was silently ignored. Clients that want the server default should leave score_threshold_set unset (or false) rather than sending 0. Affects the REST, Connect, and gRPC transports for create/anonymize, synthesize, and PDF redaction.
  • The NER build tag hugot is renamed to ner. Self-hosters who build the NER variant from source now use go build -tags ner ./... (was -tags hugot); the published anonde-ner image is unaffected. The old name was misleading once the hugot transformer backend was removed (see below) — the tag only ever gated GLiNER. The anonde_build_info metric label for NER builds is now ner (was hugot).

Removed

  • anonde-ner-stack image and the gliner-stack backend. The NER lineup consolidates to a single NER image: anonde-ner (base GLiNER, ~770 MB) alongside the patterns-only anonde (~12 MB). The premium BASE+LARGE stack image (~2.1 GB) and its bundled ANALYZER_BACKEND=gliner-stack ensemble are no longer built, pushed, or documented. The LARGE GLiNER model itself stays reachable via ANALYZER_BACKEND=gliner-flat for self-hosters who want it, so no detection capability is lost — only the prebuilt two-model image and its ensemble wiring. Dockerfile.anonde-ner-stack, the make docker-build-ner-stack target, the ner-stack compose profile, and the stack jobs in the release / build CI are all removed.
  • hugot / XLM-R transformer NER backend (ANALYZER_BACKEND=hugot and the HUGOT_MODEL / HUGOT_MODELS_DIR env vars). GLiNER strictly outperformed it on every benched corpus and was the production default; the hugot recognizer carried a second ONNX code path for no recall benefit. The hugot library stays a dependency — GLiNER reuses its model downloader and on-disk cache layout. The public library entry points anonde.DefaultAnalyzerEngineWithHugot / ...WithHugotConfig and the recognizers.HugotNERConfig type are removed; use the GLiNER constructors instead.
  • Ollama NER backend (ANALYZER_BACKEND=ollama) and the OLLAMA_ENDPOINT / OLLAMA_MODEL env vars. The LLM-over-HTTP backend was never shipped in either image, carried no leak-rate bench numbers, and depended on an external daemon — off-brand for a deterministic local-redaction tool and pure installation-surface clutter. The public library entry point anonde.DefaultAnalyzerEngineWithOllama is removed. Using anonde as a privacy proxy in front of a local Ollama (ANONDE_OPENAI_BASE_URL) is unaffected — that is a separate feature.

0.1.1 - 2026-06-13

Maintenance release. Re-publishes the three image variants (anonde, anonde-ner, anonde-ner-stack) at 0.1.1.

Added

  • Selectable GLiNER label sets (GLINER_LABEL_SET) — the NER image ships four curated open-set label sets and selects one at inference time: chat (default), clinical, finance, and legal. All four map onto the same canonical entity types the pattern recognizers emit, so anonymizer operators and reveal/detokenize behave identically regardless of which set is active; an unrecognised value falls back to chat. Go-library callers set GLiNERConfig.Labels / LabelToEntity directly (e.g. recognizers.FinancePIILabels). See the README for the per-set coverage table.

Changed

  • Default NER label set is now chat (was the full clinical set in 0.1.0). chat drops age, profession, job title, date / date of birth, and the clinical / German-insurance labels because they over-redact ordinary conversational text ("18 years of experience" → AGE, "tech" → PROFESSION). Deployments that need the old behavior set GLINER_LABEL_SET=clinical.

0.1.0 - 2026-05-28

First tagged release. Three image variants published to ghcr.io: ghcr.io/anonde-io/anonde:0.1.0 (patterns-only, ~12 MB), ghcr.io/anonde-io/anonde-ner:0.1.0 (BASE GLiNER, ~770 MB), and ghcr.io/anonde-io/anonde-ner-stack:0.1.0 (BASE + LARGE GLiNER, ~2.1 GB). Multi-arch (linux/amd64 + linux/arm64).

Added

  • PII analyzer — 52 region-aware pattern recognizers covering 12+ jurisdictions (international IDs, US, UK, Germany, Italy, Spain, Australia, India, Poland, Singapore, Finland, Korea), with parallel dispatch and score-based conflict resolution.
  • Optional in-process NER behind the -tags hugot build — GLiNER and hugot ONNX recognizers. Requires CGO and a reachable libonnxruntime.so. NER beats patterns for PERSON/ORG/LOC/AGE/PROFESSION/NRP on conflict.
  • Anonymizer operators — Replace, Redact, Mask, Hash, Encrypt, and Synthesize, with adjacent-span merging.
  • Reversible token vault — in-memory and bbolt backends behind one interface. Tokens are stable per (tenant, doc) and reversible only through the reveal path.
  • HTTP server — three transports on one port: REST/JSON (grpc-gateway), Connect (Connect/JSON, Connect/Protobuf, gRPC-Web), and native gRPC. Wire JSON is snake_case; camelCase inputs are also accepted.
  • Reveal / detokenize gated by actor + purpose audit metadata — the only path that returns cleartext.
  • OpenAI-compatible proxy at POST /v1/chat/completions — anonymizes the prompt, forwards to the upstream provider, and de-anonymizes the response in OpenAI shape. v0.1 proxies OpenAI only, non-streaming only.
  • Content formats — text, JSON, NDJSON, logs, and PDF, with format negotiation. Scanned (image-only) PDFs are OCR'd via pdftoppm + tesseract when both are on PATH; bundled in the NER image variants, no-op fallback in the patterns-only image. Tunable via ANONDE_OCR_* env vars.
  • Docker images — three multi-arch (linux/amd64 + linux/arm64) variants: anonde (patterns-only, ~12 MB, pure Go), anonde-ner (BASE GLiNER, ~770 MB, CGO + bundled libonnxruntime), and anonde-ner-stack (BASE + LARGE GLiNER, ~2.1 GB).
  • Public benchmark matrix — leak-rate and F1 scoring across gold-annotated clinical, finance, legal, and general-PII corpora, re-run on every relevant PR with a guard rail against silent NER fallback.

Security

  • NER models are baked into the NER image; there is no outbound HuggingFace traffic at request time.
  • All anonymization and de-anonymization runs locally — no third-party calls except the upstream provider when the OpenAI-compatible proxy is explicitly configured.

Planned (post-v0.1.0)

  • Secret recognizers — API keys, tokens, credentials.
  • Streaming SSE support for the OpenAI-compatible proxy (stream: true).
  • Anthropic and Gemini upstreams for the proxy, selected by model prefix.