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"

services/shared/config.py

@@ -47,7 +47,7 @@ class OllamaConfig:
     retry_base_delay: float = 1.0
     retry_max_delay: float = 10.0
     retry_backoff_multiplier: float = 2.0
-    max_tokens: int = 4096
+    max_tokens: int = 32768
     stall_timeout: float = 30.0
     loop_window: int = 64
     loop_threshold: float = 0.5