feat: add MiniMax model compatibility support

.env.example

@@ -5,6 +5,14 @@ LLM_API_KEY=your_api_key_here
 LLM_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
 LLM_MODEL_NAME=qwen-plus
 
+# ===== 使用 MiniMax 模型（可选）=====
+# MiniMax M2.5: 高性能、高性价比，支持 204,800 tokens 上下文
+# 获取 API Key: https://platform.minimax.io/
+# LLM_API_KEY=your_minimax_api_key_here
+# LLM_BASE_URL=https://api.minimax.io/v1
+# LLM_MODEL_NAME=MiniMax-M2.5
+# 国内用户可使用: LLM_BASE_URL=https://api.minimaxi.com/v1
+
 # ===== ZEP记忆图谱配置 =====
 # 每月免费额度即可支撑简单使用：https://app.getzep.com/
 ZEP_API_KEY=your_zep_api_key_here

README-EN.md

@@ -127,6 +127,28 @@ LLM_MODEL_NAME=qwen-plus
 ZEP_API_KEY=your_zep_api_key
 ```
 
+<details>
+<summary><b>Using MiniMax Models</b></summary>
+
+[MiniMax](https://platform.minimax.io/) provides high-performance, cost-effective LLM models with OpenAI-compatible API:
+
+```env
+LLM_API_KEY=your_minimax_api_key
+LLM_BASE_URL=https://api.minimax.io/v1
+LLM_MODEL_NAME=MiniMax-M2.5
+```
+
+| Model | Description |
+|-------|-------------|
+| `MiniMax-M2.5` | Flagship model, 204K context window |
+| `MiniMax-M2.5-highspeed` | Same performance, faster and more agile |
+
+For users in China: `LLM_BASE_URL=https://api.minimaxi.com/v1`
+
+API Documentation: [OpenAI Compatible API](https://platform.minimax.io/docs/api-reference/text-openai-api)
+
+</details>
+
 #### 2. Install Dependencies
 
 ```bash

README.md

@@ -127,6 +127,28 @@ LLM_MODEL_NAME=qwen-plus
 ZEP_API_KEY=your_zep_api_key
 ```
 
+<details>
+<summary><b>使用 MiniMax 模型</b></summary>
+
+[MiniMax](https://platform.minimax.io/) 提供高性能、高性价比的 LLM 模型，支持 OpenAI 兼容 API：
+
+```env
+LLM_API_KEY=your_minimax_api_key
+LLM_BASE_URL=https://api.minimax.io/v1
+LLM_MODEL_NAME=MiniMax-M2.5
+```
+
+| 模型 | 说明 |
+|------|------|
+| `MiniMax-M2.5` | 旗舰模型，204K 上下文窗口 |
+| `MiniMax-M2.5-highspeed` | 同等性能，更快更敏捷 |
+
+国内用户可使用：`LLM_BASE_URL=https://api.minimaxi.com/v1`
+
+API 文档：[OpenAI 兼容接口](https://platform.minimax.io/docs/api-reference/text-openai-api)
+
+</details>
+
 #### 2. 安装依赖
 
 ```bash

backend/app/services/oasis_profile_generator.py

@@ -20,6 +20,7 @@
 
 from ..config import Config
 from ..utils.logger import get_logger
+from ..utils.llm_client import _is_minimax, _clamp_temperature, _inject_json_instruction, parse_json_from_response
 from .zep_entity_reader import EntityNode, ZepEntityReader
 
 logger = get_logger('mirofish.oasis_profile')
@@ -523,43 +524,53 @@ def _generate_profile_with_llm(
         # 尝试多次生成，直到成功或达到最大重试次数
         max_attempts = 3
         last_error = None
-
+        use_minimax = _is_minimax(self.model_name, self.base_url)
+
         for attempt in range(max_attempts):
             try:
-                response = self.client.chat.completions.create(
-                    model=self.model_name,
-                    messages=[
-                        {"role": "system", "content": self._get_system_prompt(is_individual)},
-                        {"role": "user", "content": prompt}
-                    ],
-                    response_format={"type": "json_object"},
-                    temperature=0.7 - (attempt * 0.1)  # 每次重试降低温度
+                import re as _re
+                temperature = _clamp_temperature(0.7 - (attempt * 0.1), self.model_name, self.base_url)
+                messages = [
+                    {"role": "system", "content": self._get_system_prompt(is_individual)},
+                    {"role": "user", "content": prompt}
+                ]
+
+                kwargs = {
+                    "model": self.model_name,
+                    "messages": _inject_json_instruction(messages) if use_minimax else messages,
+                    "temperature": temperature,
                     # 不设置max_tokens，让LLM自由发挥
-                )
-
+                }
+                if not use_minimax:
+                    kwargs["response_format"] = {"type": "json_object"}
+
+                response = self.client.chat.completions.create(**kwargs)
+
                 content = response.choices[0].message.content
-
+                # 移除 <think> 标签
+                content = _re.sub(r'<think>[\s\S]*?</think>', '', content).strip()
+
                 # 检查是否被截断（finish_reason不是'stop'）
                 finish_reason = response.choices[0].finish_reason
                 if finish_reason == 'length':
                     logger.warning(f"LLM输出被截断 (attempt {attempt+1}), 尝试修复...")
                     content = self._fix_truncated_json(content)
-                
+
                 # 尝试解析JSON
                 try:
-                    result = json.loads(content)
-                    
+                    result = parse_json_from_response(content)
+
                     # 验证必需字段
                     if "bio" not in result or not result["bio"]:
                         result["bio"] = entity_summary[:200] if entity_summary else f"{entity_type}: {entity_name}"
                     if "persona" not in result or not result["persona"]:
                         result["persona"] = entity_summary or f"{entity_name}是一个{entity_type}。"
-                    
+
                     return result
-                    
-                except json.JSONDecodeError as je:
+
+                except (json.JSONDecodeError, ValueError) as je:
                     logger.warning(f"JSON解析失败 (attempt {attempt+1}): {str(je)[:80]}")
-                    
+
                     # 尝试修复JSON
                     result = self._try_fix_json(content, entity_name, entity_type, entity_summary)
                     if result.get("_fixed"):

backend/app/services/simulation_config_generator.py

@@ -20,6 +20,7 @@
 
 from ..config import Config
 from ..utils.logger import get_logger
+from ..utils.llm_client import _is_minimax, _clamp_temperature, _inject_json_instruction, parse_json_from_response
 from .zep_entity_reader import EntityNode, ZepEntityReader
 
 logger = get_logger('mirofish.simulation_config')
@@ -433,42 +434,51 @@ def _summarize_entities(self, entities: List[EntityNode]) -> str:
     def _call_llm_with_retry(self, prompt: str, system_prompt: str) -> Dict[str, Any]:
         """带重试的LLM调用，包含JSON修复逻辑"""
         import re
-        
+
         max_attempts = 3
         last_error = None
-
+        use_minimax = _is_minimax(self.model_name, self.base_url)
+
         for attempt in range(max_attempts):
             try:
-                response = self.client.chat.completions.create(
-                    model=self.model_name,
-                    messages=[
-                        {"role": "system", "content": system_prompt},
-                        {"role": "user", "content": prompt}
-                    ],
-                    response_format={"type": "json_object"},
-                    temperature=0.7 - (attempt * 0.1)  # 每次重试降低温度
+                temperature = _clamp_temperature(0.7 - (attempt * 0.1), self.model_name, self.base_url)
+                messages = [
+                    {"role": "system", "content": system_prompt},
+                    {"role": "user", "content": prompt}
+                ]
+
+                kwargs = {
+                    "model": self.model_name,
+                    "messages": _inject_json_instruction(messages) if use_minimax else messages,
+                    "temperature": temperature,
                     # 不设置max_tokens，让LLM自由发挥
-                )
-
+                }
+                if not use_minimax:
+                    kwargs["response_format"] = {"type": "json_object"}
+
+                response = self.client.chat.completions.create(**kwargs)
+
                 content = response.choices[0].message.content
+                # 移除 <think> 标签
+                content = re.sub(r'<think>[\s\S]*?</think>', '', content).strip()
                 finish_reason = response.choices[0].finish_reason
-                
+
                 # 检查是否被截断
                 if finish_reason == 'length':
                     logger.warning(f"LLM输出被截断 (attempt {attempt+1})")
                     content = self._fix_truncated_json(content)
-                
+
                 # 尝试解析JSON
                 try:
-                    return json.loads(content)
-                except json.JSONDecodeError as e:
+                    return parse_json_from_response(content)
+                except (json.JSONDecodeError, ValueError) as e:
                     logger.warning(f"JSON解析失败 (attempt {attempt+1}): {str(e)[:80]}")
-                    
+
                     # 尝试修复JSON
                     fixed = self._try_fix_config_json(content)
                     if fixed:
                         return fixed
-                    
+
                     last_error = e
 
             except Exception as e: