fix: remove think=false (Ollama bug #14645), bump max_tokens to 32k

services/extractor/client.py

@@ -228,7 +228,12 @@ class OllamaClient:
             ],
             "format": json_schema,
             "stream": True,
-            "think": False,
+            # NOTE: Do NOT set "think": False here. Ollama has a known bug
+            # (issues #14645, #15260) where think=false silently disables
+            # the format constraint for qwen3.5 and gemma4 models, causing
+            # the model to output plain text instead of valid JSON.
+            # Omitting "think" lets the model use thinking mode (slightly
+            # slower but structured output actually works).
         }
 
         url = f"{self._config.base_url}/api/chat"