feat: add dedicated tts_instruct task for auto-instruct

pyproject.toml

@@ -25,7 +25,7 @@ dependencies = [
     "pypinyin>=0.54.0",
     "python-dotenv>=1.1.1",
     "python-multipart>=0.0.20",
-    "quick-algo>=0.1.3",
+    "quick-algo @ file:///Users/xenon/MaiM-with-u/MaiMBot-LPMM/lib/quick_algo",
     "rich>=14.0.0",
     "ruff>=0.12.2",
     "setuptools>=80.9.0",

requirements.txt

@@ -18,7 +18,7 @@ pydantic>=2.11.7
 pypinyin>=0.54.0
 python-dotenv>=1.1.1
 python-multipart>=0.0.20
-quick-algo>=0.1.3
+quick-algo @ file:///Users/xenon/MaiM-with-u/MaiMBot-LPMM/lib/quick_algo
 rich>=14.0.0
 ruff>=0.12.2
 setuptools>=80.9.0
@@ -28,4 +28,4 @@ tomlkit>=0.13.3
 urllib3>=2.5.0
 uvicorn>=0.35.0
 msgpack
-zstandard
+zstandard

src/config/api_ada_configs.py

@@ -132,6 +132,9 @@ class ModelTaskConfig(ConfigBase):
     lpmm_rdf_build: TaskConfig
     """LPMM RDF构建模型配置"""
 
+    tts_instruct: TaskConfig = field(default_factory=TaskConfig)
+    """TTS instruct 生成模型配置（用于生成情绪/语速/停顿指令）"""
+
     def get_task(self, task_name: str) -> TaskConfig:
         """获取指定任务的配置"""
         if hasattr(self, task_name):

src/plugin_system/apis/generator_api.py

@@ -21,6 +21,8 @@ from src.chat.utils.utils import process_llm_response
 from src.chat.replyer.replyer_manager import replyer_manager
 from src.plugin_system.base.component_types import ActionInfo
 from src.chat.logger.plan_reply_logger import PlanReplyLogger
+from src.config.config import model_config
+from src.llm_models.utils_model import LLMRequest
 
 if TYPE_CHECKING:
     from src.common.data_models.info_data_model import ActionPlannerInfo
@@ -324,3 +326,34 @@ async def generate_response_custom(
     except Exception as e:
         logger.error(f"[GeneratorAPI] 生成自定义回复时出错: {e}")
         return None
+
+
+async def generate_tts_instruct(
+    prompt: str,
+    request_type: str = "tts_instruct",
+    temperature: Optional[float] = None,
+    max_tokens: Optional[int] = None,
+) -> Optional[str]:
+    """
+    生成 TTS 表演指令（instruct），使用独立的 model_task_config.tts_instruct 模型组。
+
+    设计目标：
+    - 不走 replyer（避免上下文/人设污染）
+    - 仅基于 prompt 生成短文本，便于在 WebUI 的 model_config.toml 单独配置模型
+    """
+    try:
+        task_cfg = model_config.model_task_config.get_task("tts_instruct")
+        if not task_cfg.model_list:
+            logger.error("[GeneratorAPI] tts_instruct 的 model_list 为空，请在 model_config.toml 配置 [model_task_config.tts_instruct]")
+            return None
+
+        llm = LLMRequest(model_set=task_cfg, request_type=request_type)
+        content, _meta = await llm.generate_response_async(
+            prompt=prompt,
+            temperature=temperature if temperature is not None else task_cfg.temperature,
+            max_tokens=max_tokens if max_tokens is not None else task_cfg.max_tokens,
+        )
+        return content.strip() if content else None
+    except Exception as e:
+        logger.error(f"[GeneratorAPI] 生成 tts_instruct 失败: {e}", exc_info=True)
+        return None

template/model_config_template.toml

@@ -156,6 +156,13 @@ max_tokens = 2048
 slow_threshold = 25.0
 selection_strategy = "random"           # 模型选择策略：balance（负载均衡）或 random（随机选择）
 
+[model_task_config.tts_instruct] # 生成 TTS 表演指令（情绪/语速/停顿）用的模型（独立于 replyer）
+model_list = ["qwen3-30b","qwen3-next-80b"]
+temperature = 0.2
+max_tokens = 128
+slow_threshold = 10.0
+selection_strategy = "random"
+
 [model_task_config.planner] #决策：负责决定麦麦该什么时候回复的模型
 model_list = ["siliconflow-deepseek-v3.2"]
 temperature = 0.3
@@ -194,4 +201,4 @@ model_list = ["siliconflow-deepseek-v3.2"]
 temperature = 0.2
 max_tokens = 800
 slow_threshold = 20.0
-selection_strategy = "random"           # 模型选择策略：balance（负载均衡）或 random（随机选择）
+selection_strategy = "random"           # 模型选择策略：balance（负载均衡）或 random（随机选择）