From 0ccc6c4047015543f2c9f4a3b2e581b7002d1d8c Mon Sep 17 00:00:00 2001
From: Benjamin Admin <benjaminadmin@MacBook-Pro.local>
Date: Tue, 28 Apr 2026 00:51:06 +0200
Subject: [PATCH] fix: handle Qwen think mode in classification, add German
 term matching

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 .../compliance/api/agent_analyze_routes.py    | 21 +++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/backend-compliance/compliance/api/agent_analyze_routes.py b/backend-compliance/compliance/api/agent_analyze_routes.py
index 246d27d..b4cb5da 100644
--- a/backend-compliance/compliance/api/agent_analyze_routes.py
+++ b/backend-compliance/compliance/api/agent_analyze_routes.py
@@ -121,18 +121,35 @@ async def _classify(client: httpx.AsyncClient, text: str) -> str:
         resp = await client.post(f"{SDK_URL}/sdk/v1/llm/chat", headers=SDK_HEADERS, json={
             "messages": [
                 {"role": "system", "content": (
+                    "/no_think\n"
                     "Klassifiziere das Dokument in GENAU EINE Kategorie: "
                     "privacy_policy, cookie_banner, terms_of_service, imprint, dpa, other. "
-                    "Antworte NUR mit dem Kategorienamen, nichts anderes."
+                    "Antworte NUR mit dem Kategorienamen, nichts anderes. Kein Denken, keine Erklaerung."
                 )},
                 {"role": "user", "content": text[:2000]},
             ],
         })
         data = resp.json()
-        raw = data.get("response", data.get("content", "other")).strip().lower()
+        # Qwen 3.5 may use think mode — content can be in message.content or response
+        raw = (
+            data.get("response", "")
+            or data.get("content", "")
+            or (data.get("message", {}) or {}).get("content", "")
+            or ""
+        ).strip().lower()
+        # Strip Qwen think tags if present
+        raw = re.sub(r"<think>.*?</think>", "", raw, flags=re.DOTALL).strip()
+        logger.info("Classification raw response: %s", raw[:200])
         for cat in ["privacy_policy", "cookie_banner", "terms_of_service", "imprint", "dpa"]:
             if cat in raw:
                 return cat
+        # Also check German terms
+        if "datenschutz" in raw:
+            return "privacy_policy"
+        if "cookie" in raw:
+            return "cookie_banner"
+        if "impressum" in raw:
+            return "imprint"
         return "other"
     except Exception as e:
         logger.warning("Classification failed: %s", e)