feat(cra): Datenblatt-Extraktion auf lokales 35B + llm_status-Fix

llm_cascade additiv modell-faehig (optionaler model-Param, Cache-Key kennt
model_hint → keine Kollision; Default unveraendert für alle anderen Nutzer).
Datenblatt-Extraktor nutzt jetzt qwen3.5:35b-a3b (CRA_DATASHEET_MODEL, gleiches
Modell wie der Compliance Advisor) für bessere semantische Zuordnung. Plus
llm_status (ok|empty|unavailable) + Logging statt stillem except; Frontend zeigt
bei 'unavailable' einen Hinweis statt leerer Felder (wichtig auf prod ohne
lokales Ollama → Cascade-Fallback bzw. Hinweis).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

admin-compliance/app/sdk/cra/_components/DatasheetExtract.tsx

@@ -8,6 +8,7 @@ interface ExtractResult {
   limits: Record<string, string>
   provenance: Record<string, string>
   detected: { interfaces: string[]; units: string[] }
+  llm_status?: string
   filled: string[]
   missing: string[]
   followup: Followup[]
@@ -73,6 +74,12 @@ export function DatasheetExtract() {
 
       {res && (
         <div className="mt-5 space-y-4">
+          {res.llm_status === 'unavailable' && (
+            <div className="rounded-lg border border-amber-300 bg-amber-50 dark:bg-amber-900/20 text-amber-900 dark:text-amber-200 p-3 text-xs">
+              KI-Extraktion gerade nicht verfügbar (lokales Modell lädt oder offline). Unten stehen nur
+              deterministisch erkannte Werte — bitte „Grenzen extrahieren" erneut klicken oder Felder manuell ergänzen.
+            </div>
+          )}
           {(res.detected.interfaces.length > 0 || res.detected.units.length > 0) && (
             <div className="text-xs text-gray-600 dark:text-gray-300">
               <span className="font-medium">Deterministisch erkannt:</span>{' '}

backend-compliance/compliance/services/cra_datasheet_extractor.py

@@ -10,8 +10,15 @@ Pure + testable: detect_signals / parse_grenzen_json / compute_followups. The
 async extract_grenzen() wraps the LLM call (llm_cascade, same as vendor extractor).
 """
 import json
+import logging
+import os
 import re
-from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+# Datasheet extraction uses the local 35B (same model as the Compliance Advisor) —
+# higher-quality semantic mapping than the default cascade model. Env-overridable.
+_DATASHEET_MODEL = os.getenv("CRA_DATASHEET_MODEL", "qwen3.5:35b-a3b")
 
 # IACE Grenzen field keys (must match admin LimitsFormData). label + whether it
 # is essential for a usable risk assessment (=> asked as follow-up if empty).
@@ -150,6 +157,7 @@ async def extract_grenzen(text: str, max_chars: int = 20000) -> dict:
     signals = detect_signals(text or "")
     limits: dict = {}
     provenance: dict = {}
+    llm_status = "skipped"  # skipped | ok | empty | unavailable
     excerpt = (text or "")[:max_chars]
     if len(excerpt) >= 200:
         try:
@@ -157,20 +165,25 @@ async def extract_grenzen(text: str, max_chars: int = 20000) -> dict:
             res = await call_with_cascade(
                 system=_system_prompt(),
                 user=f"Datenblatt-Text:\n\n{excerpt}",
-                min_confidence=0.5, max_tokens=4000,
+                min_confidence=0.5, max_tokens=4000, model=_DATASHEET_MODEL,
             )
             parsed = parse_grenzen_json(res.get("text", "") if isinstance(res, dict) else "")
             for key, entry in parsed.items():
                 limits[key] = entry["value"]
                 provenance[key] = entry.get("source", "")
-        except Exception:
+            llm_status = "ok" if parsed else "empty"
-            pass  # extraction is best-effort; fall back to detector + follow-ups
+        except Exception as e:
+            # best-effort: keep the deterministic facts, but surface the failure so
+            # a cold-start/timeout doesn't masquerade as "nothing on the datasheet".
+            logger.warning("datasheet LLM extraction failed: %s (%s)", e, type(e).__name__)
+            llm_status = "unavailable"
 
     _merge_detected(limits, provenance, signals)
     return {
         "limits": limits,
         "provenance": provenance,
         "detected": signals,
+        "llm_status": llm_status,
         "filled": sorted(limits.keys()),
         "missing": [k for k in _FIELD_KEYS if not (limits.get(k) or "").strip()],
         "followup": compute_followups(limits),

backend-compliance/compliance/services/llm_cascade.py

@@ -104,9 +104,10 @@ def _heuristic_confidence(response_text: str, input_len: int) -> float:
 
 async def _call_ollama(system: str, user: str,
                         max_tokens: int = 6000,
-                        timeout: float = 90.0) -> str:
+                        timeout: float = 90.0,
+                        model: str = "") -> str:
     base = os.getenv("OLLAMA_URL", "http://host.docker.internal:11434")
-    model = os.getenv("CMP_LLM_MODEL", "qwen3:30b-a3b")
+    model = model or os.getenv("CMP_LLM_MODEL", "qwen3:30b-a3b")
     payload = {
         "model": model, "stream": False, "format": "json",
         "messages": [{"role": "system", "content": system},
@@ -188,10 +189,11 @@ async def call_with_cascade(
     user: str,
     min_confidence: float = 0.6,
     max_tokens: int = 6000,
+    model: str = "",
 ) -> dict:
     """Returns {'text': str, 'confidence': float, 'source': str,
-    'cached': bool}."""
+    'cached': bool}. `model` overrides the local Tier-1 (Ollama) model only."""
-    key = _cache_key(system, user)
+    key = _cache_key(system, user, model)
     cached = _cache_get(key)
     if cached:
         cached["cached"] = True
@@ -211,7 +213,7 @@ async def call_with_cascade(
                 "or ANTHROPIC_API_KEY to enable fallbacks."
             )
     # Tier 1: Qwen lokal
-    text = await _call_ollama(system, user, max_tokens=max_tokens)
+    text = await _call_ollama(system, user, max_tokens=max_tokens, model=model)
     conf = _heuristic_confidence(text, input_len)
     if text and conf >= min_confidence:
         out = {"text": text, "confidence": conf,