fix: LLM thinking 허용 + 마지막 유효 JSON 추출 방식으로 변경

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

scripts/pkm_utils.py

@@ -109,32 +109,32 @@ def llm_generate(prompt: str, model: str = "mlx-community/Qwen3.5-35B-A3B-4bit",
                  host: str = "http://localhost:8800", json_mode: bool = False) -> str:
     """MLX 서버 API 호출 (OpenAI 호환)"""
     import requests
-    messages = []
+    messages = [{"role": "user", "content": prompt}]
-    if json_mode:
-        messages.append({"role": "system", "content": "IMPORTANT: Output ONLY valid JSON. No thinking process, no explanation, no markdown fences. Start your response with { and end with }."})
-    messages.append({"role": "user", "content": prompt})
     resp = requests.post(f"{host}/v1/chat/completions", json={
         "model": model,
         "messages": messages,
-        "temperature": 0.1 if json_mode else 0.3,
+        "temperature": 0.3,
-        "max_tokens": 2048,
+        "max_tokens": 4096,
-    }, timeout=180)
+    }, timeout=300)
     resp.raise_for_status()
     content = resp.json()["choices"][0]["message"]["content"]
-    # thinking 블록 제거 (Qwen3.5 thinking 모델 대응)
+    if not json_mode:
-    if "<think>" in content and "</think>" in content:
+        return content
-        content = content.split("</think>")[-1].strip()
+    # JSON 모드: thinking 허용 → 마지막 유효 JSON 객체 추출
-    # JSON 블록 추출
-    if "```json" in content:
-        content = content.split("```json")[1].split("```")[0].strip()
-    elif "```" in content:
-        content = content.split("```")[1].split("```")[0].strip()
-    # { 로 시작하는 JSON 추출
     import re
+    import json as _json
+    # 배열이 포함된 JSON 객체 매칭
+    all_jsons = re.findall(r'\{[^{}]*(?:\[[^\]]*\])?[^{}]*\}', content)
+    for j in reversed(all_jsons):
+        try:
+            parsed = _json.loads(j)
+            if any(k in parsed for k in ("domain_db", "tags", "domain", "classification")):
+                return j
+        except _json.JSONDecodeError:
+            continue
+    # 폴백: 전체에서 가장 큰 JSON 추출
     json_match = re.search(r'\{[\s\S]*\}', content)
-    if json_match:
+    return json_match.group(0) if json_match else content
-        content = json_match.group(0)
-    return content
 
 
 # 하위호환 별칭