From 49c39a182b200f00460c6fcce55c680af4718d57 Mon Sep 17 00:00:00 2001
From: Hyungi Ahn <hyungiahn@Hyungiui-MacBookPro.local>
Date: Thu, 26 Mar 2026 13:34:20 +0900
Subject: [PATCH] =?UTF-8?q?fix:=20LLM=20thinking=20=EC=B6=9C=EB=A0=A5=20?=
 =?UTF-8?q?=EB=8C=80=EC=9D=91=20=E2=80=94=20max=5Ftokens=20=EC=A6=9D?=
 =?UTF-8?q?=EA=B0=80=20+=20JSON=20=EC=B6=94=EC=B6=9C=20=EA=B0=95=ED=99=94?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 scripts/pkm_utils.py | 21 ++++++++++++++-------
 1 file changed, 14 insertions(+), 7 deletions(-)

diff --git a/scripts/pkm_utils.py b/scripts/pkm_utils.py
index f901701..eb66b45 100644
--- a/scripts/pkm_utils.py
+++ b/scripts/pkm_utils.py
@@ -111,22 +111,29 @@ def llm_generate(prompt: str, model: str = "mlx-community/Qwen3.5-35B-A3B-4bit",
     import requests
     messages = []
     if json_mode:
-        messages.append({"role": "system", "content": "You must respond ONLY with valid JSON. No thinking, no explanation, no markdown."})
-    # Qwen3.5: /nothink 접미사로 thinking 출력 억제
-    messages.append({"role": "user", "content": prompt + " /nothink"})
+        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": 0.3,
-        "max_tokens": 1024,
-    }, timeout=120)
+        "temperature": 0.1 if json_mode else 0.3,
+        "max_tokens": 2048,
+    }, timeout=180)
     resp.raise_for_status()
     content = resp.json()["choices"][0]["message"]["content"]
-    # JSON 블록 추출 (```json ... ``` 감싸기 대응)
+    # thinking 블록 제거 (Qwen3.5 thinking 모델 대응)
+    if "<think>" in content and "</think>" in content:
+        content = content.split("</think>")[-1].strip()
+    # 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
+    json_match = re.search(r'\{[\s\S]*\}', content)
+    if json_match:
+        content = json_match.group(0)
     return content