text.anonymize.legal-de 0.1.0: WASM bridge to the judge-ner host service

NER anonymization of German legal texts (plain text + DOCX) via the JuraNER model by Harshil Darji (MIT). The transformer model runs in the judge-ner host service; this module forwards documents over loopback HTTP and maps responses onto the capability contract.
2026-07-10 12:44:32 +02:00 · 2026-07-10 12:44:32 +02:00 · 3e018f528a
commit 3e018f528a
15 changed files with 1781 additions and 0 deletions
--- a/module.yaml
+++ b/module.yaml
@ -0,0 +1,112 @@
+schema_version: 3
+provider: judge
+
+name: text-anonymize-legal-de
+version: 0.1.0
+
+# Bridge module: forwards the document to the judge-ner host service
+# (FastAPI, JuraNER transformers model) over loopback HTTP and maps
+# the response onto the frozen capability contract below.
+provides:
+  - capability: text.anonymize.legal-de
+    version: 0.1.0
+
+inputs:
+  request:
+    type: json
+    description:
+      en: |
+        Anonymization request:
+          {
+            "mode": "text" | "docx",
+            "options": {                              // all optional
+              "importance_levels": ["High"],          // High|Mid|Low
+              "confidence_threshold": 0.8,            // 0.0..1.0
+              "manual_phrases": [],                   // extra phrases to redact
+              "maintain_consistency": true,           // same surface -> same token
+              "remove_rubrum": false                  // drop the rubrum block
+            }
+          }
+      de: |
+        Anonymisierungs-Anfrage:
+          {
+            "mode": "text" | "docx",
+            "options": {                              // alle optional
+              "importance_levels": ["High"],          // High|Mid|Low
+              "confidence_threshold": 0.8,            // 0.0..1.0
+              "manual_phrases": [],                   // zusätzlich zu schwärzende Phrasen
+              "maintain_consistency": true,           // gleiche Oberfläche -> gleiches Token
+              "remove_rubrum": false                  // Rubrum-Block entfernen
+            }
+          }
+  document:
+    type: bytes
+    description:
+      en: |
+        The document to anonymize. UTF-8 plain text for mode=text,
+        a DOCX file for mode=docx.
+      de: |
+        Das zu anonymisierende Dokument. UTF-8-Klartext bei mode=text,
+        eine DOCX-Datei bei mode=docx.
+  endpoint:
+    type: text
+    description:
+      en: |
+        Optional base URL of the judge-ner host service.
+        Default: "http://127.0.0.1:8756".
+      de: |
+        Optionale Basis-URL des judge-ner-Host-Services.
+        Default: "http://127.0.0.1:8756".
+
+outputs:
+  result:
+    type: json
+    description:
+      en: |
+        {
+          "replacements": [ { "text", "label", "label_description",
+                              "start", "end", "score", "line_number",
+                              "anonymized", ...run-id fields for docx }, ... ],
+          "text": "..."          // anonymized text, only for mode=text
+        }
+        `replacements` lists all detected entities; entries with
+        "anonymized": null were detected but not replaced. A
+        "statistics" object (and "rubrum" when remove_rubrum matched)
+        is passed through from the service.
+      de: |
+        {
+          "replacements": [ { "text", "label", "label_description",
+                              "start", "end", "score", "line_number",
+                              "anonymized", ...Run-ID-Felder bei docx }, ... ],
+          "text": "..."          // anonymisierter Text, nur bei mode=text
+        }
+        `replacements` enthält alle erkannten Entitäten; Einträge mit
+        "anonymized": null wurden erkannt, aber nicht ersetzt. Ein
+        "statistics"-Objekt (und "rubrum" bei erkanntem Rubrum) wird
+        vom Service durchgereicht.
+  document:
+    type: bytes
+    description:
+      en: |
+        The anonymized document: the anonymized UTF-8 text for
+        mode=text, the rewritten DOCX for mode=docx.
+      de: |
+        Das anonymisierte Dokument: bei mode=text die anonymisierte
+        UTF-8-Textfassung, bei mode=docx das umgeschriebene DOCX.
+
+# MIME allow-list for the bytes input (Studio file-picker filter).
+accepts_mime:
+  - text/plain
+  - application/vnd.openxmlformats-officedocument.wordprocessingml.document
+
+# Outbound HTTP to the local judge-ner host service only.
+permissions:
+  - "net: localhost"
+  - "net: 127.0.0.1"
+
+# The hub refuses to dispatch a step to this module unless the
+# operator declared a `judge-ner` entry in the service registry
+# (~/.chain/config.yaml `services:`). See
+# fai_chain/docs/architecture/host-services.md.
+requires_services:
+  - judge-ner