release: v2.8.2 fix Ollama amount normalization

MDfile/RELEASE_NOTES_v2.8.2.md

@@ -0,0 +1,13 @@
+# BMC Hub v2.8.2
+
+## AI Template Builder
+
+- Rettet fejl når Ollama returnerer fakturabeløb som tekst i stedet for numerisk JSON.
+- Understøtter nu både danske beløbsformater som `1.471,20 DKK` og engelske formater som `1471.20`.
+- Ugyldige eller tomme beløb håndteres uden at afbryde AI-analysen.
+- Tilføjet regressionstest for beløbsnormalisering.
+
+## Verifikation
+
+- Regressionstest: 2 bestået.
+- Python-syntaks og diff-kontrol bestået.

VERSION

@@ -1 +1 @@
-2.8.1
+2.8.2

app/services/ollama_service.py

@@ -460,6 +460,18 @@ REGLER FOR title:
                 # Parse JSON from response
                 extraction = self._parse_json_response(raw_response)
 
+                # Models do not always respect the numeric JSON schema and may
+                # return Danish/English amount strings. Normalize them before
+                # applying sign rules so comparisons with zero are type-safe.
+                for amount_key in ('total_amount', 'vat_amount'):
+                    extraction[amount_key] = self._normalize_amount_value(extraction.get(amount_key))
+                if isinstance(extraction.get('lines'), list):
+                    for line in extraction['lines']:
+                        if not isinstance(line, dict):
+                            continue
+                        for amount_key in ('unit_price', 'line_total', 'vat_amount'):
+                            line[amount_key] = self._normalize_amount_value(line.get(amount_key))
+                
                 # CRITICAL: Fix amount signs based on document_type
                 # LLM sometimes returns negative amounts for invoices - fix this!
                 document_type = extraction.get('document_type', 'invoice')
@@ -528,6 +540,39 @@ REGLER FOR title:
                     "confidence": 0.0
                 }
 
+    @staticmethod
+    def _normalize_amount_value(value):
+        """Convert common Danish/English amount representations to numbers."""
+        if value is None or value == '':
+            return None
+        if isinstance(value, bool):
+            return None
+        if isinstance(value, (int, float)):
+            return value
+        if not isinstance(value, str):
+            return None
+
+        cleaned = re.sub(r'[^0-9,.-]', '', value.strip()).strip('.,')
+        if not cleaned or cleaned in {'-', '.', ',', '-.', '-,'}:
+            return None
+
+        if ',' in cleaned and '.' in cleaned:
+            if cleaned.rfind(',') > cleaned.rfind('.'):
+                cleaned = cleaned.replace('.', '').replace(',', '.')
+            else:
+                cleaned = cleaned.replace(',', '')
+        elif ',' in cleaned:
+            cleaned = cleaned.replace('.', '').replace(',', '.')
+        elif cleaned.count('.') > 1:
+            cleaned = cleaned.replace('.', '')
+        elif '.' in cleaned and len(cleaned.rsplit('.', 1)[1]) == 3:
+            cleaned = cleaned.replace('.', '')
+
+        try:
+            return float(cleaned)
+        except ValueError:
+            return None
+    
     def _parse_json_response(self, response: str) -> Dict:
         """Parse JSON from LLM response with aggressive fallback strategies"""
         logger.info(f"🔍 Response length: {len(response)}, preview: {response[:200]}")

tests/test_ollama_amount_normalization.py

@@ -0,0 +1,19 @@
+from app.services.ollama_service import OllamaService
+
+
+def test_normalizes_numeric_and_danish_amount_values():
+    normalize = OllamaService._normalize_amount_value
+
+    assert normalize(1471.2) == 1471.2
+    assert normalize("1471.20") == 1471.2
+    assert normalize("1.471,20 DKK") == 1471.2
+    assert normalize("-1 471,20 kr.") == -1471.2
+
+
+def test_rejects_non_amount_values_without_type_errors():
+    normalize = OllamaService._normalize_amount_value
+
+    assert normalize(None) is None
+    assert normalize("") is None
+    assert normalize("ukendt") is None
+    assert normalize({"value": "10,00"}) is None