增强任务控制器智能体模式和模型配置
- 新增TaskController简化模式和分数驱动模式支持 - 添加phi4模型配置选项 - 优化主程序参数配置和默认设置 - 完善工作流和步骤执行器功能 - 更新.gitignore忽略规则 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
parent
a4eca4897d
commit
cc59034f1d
8
.gitignore
vendored
8
.gitignore
vendored
@ -53,4 +53,10 @@ results/
|
||||
|
||||
file_analyzer.py
|
||||
file_deleter_renumber.py
|
||||
files_to_delete.txt
|
||||
files_to_delete.txt
|
||||
|
||||
output0902-qwen3
|
||||
.claude
|
||||
|
||||
|
||||
.nfs*
|
||||
|
||||
@ -20,16 +20,25 @@ class TaskController(BaseAgent):
|
||||
Attributes:
|
||||
model_type (str): 使用的大语言模型类型，默认为 gpt-oss:latest
|
||||
llm_config (dict): LLM模型配置参数
|
||||
simple_mode (bool): 简化模式标志，True时自动选择第一个任务并返回固定指导
|
||||
"""
|
||||
|
||||
def __init__(self, model_type: str = "gpt-oss:latest", llm_config: dict = None):
|
||||
def __init__(self, model_type: str = "gpt-oss:latest", llm_config: dict = None,
|
||||
simple_mode: bool = False, score_driven_mode: bool = False):
|
||||
"""
|
||||
初始化任务控制器智能体
|
||||
|
||||
Args:
|
||||
model_type (str): 大语言模型类型，默认使用 gpt-oss:latest
|
||||
llm_config (dict): LLM模型的配置参数，如果为None则使用默认配置
|
||||
simple_mode (bool): 简化模式，如果为True则自动选择第一个任务并返回固定指导，默认为False
|
||||
score_driven_mode (bool): 分数驱动模式，如果为True则选择当前任务组中分数最低的任务，默认为False
|
||||
|
||||
Note:
|
||||
score_driven_mode和simple_mode不能同时为True，如果同时为True则优先使用score_driven_mode
|
||||
"""
|
||||
self.simple_mode = simple_mode
|
||||
self.score_driven_mode = score_driven_mode
|
||||
super().__init__(
|
||||
model_type=model_type,
|
||||
description="医疗任务控制器，负责任务选择和预问诊询问指导",
|
||||
@ -46,7 +55,8 @@ class TaskController(BaseAgent):
|
||||
chief_complaint: str,
|
||||
hpi_content: str = "",
|
||||
ph_content: str = "",
|
||||
additional_info: str = "") -> ControllerDecision:
|
||||
additional_info: str = "",
|
||||
task_manager = None) -> ControllerDecision:
|
||||
"""
|
||||
执行任务控制决策
|
||||
|
||||
@ -69,6 +79,14 @@ class TaskController(BaseAgent):
|
||||
Exception: 当LLM调用失败时，返回包含默认信息的ControllerDecision
|
||||
"""
|
||||
try:
|
||||
# 分数驱动模式：选择当前任务组中分数最低的任务
|
||||
if self.score_driven_mode and task_manager is not None:
|
||||
return self._get_score_driven_result(pending_tasks, task_manager)
|
||||
|
||||
# 简化模式：直接选择第一个任务并返回固定指导
|
||||
elif self.simple_mode:
|
||||
return self._get_simple_mode_result(pending_tasks)
|
||||
|
||||
# 构建决策提示词
|
||||
prompt = self._build_decision_prompt(
|
||||
pending_tasks, chief_complaint, hpi_content, ph_content, additional_info
|
||||
@ -103,6 +121,90 @@ class TaskController(BaseAgent):
|
||||
# 如果类型不匹配，返回默认结果
|
||||
return self._get_fallback_result([])
|
||||
|
||||
def _get_score_driven_result(self, pending_tasks: List[Dict[str, str]], task_manager) -> ControllerDecision:
|
||||
"""
|
||||
分数驱动模式下生成决策结果
|
||||
|
||||
在分数驱动模式下，从当前任务组的未完成任务中选择分数最低的任务，
|
||||
并返回相应的询问指导。这是基于数值比较的算法选择，无需LLM参与。
|
||||
|
||||
Args:
|
||||
pending_tasks (List[Dict[str, str]]): 待执行的任务列表，包含name和description
|
||||
task_manager: 任务管理器实例，用于获取任务分数信息
|
||||
|
||||
Returns:
|
||||
ControllerDecision: 包含分数驱动模式任务选择和指导的结果
|
||||
"""
|
||||
if not pending_tasks:
|
||||
return ControllerDecision(
|
||||
selected_task="基本信息收集",
|
||||
specific_guidance="当前没有待执行任务，请按照标准医疗询问流程进行患者评估。"
|
||||
)
|
||||
|
||||
# 获取当前任务阶段
|
||||
current_phase = task_manager.get_current_phase()
|
||||
|
||||
# 获取当前阶段的任务分数
|
||||
phase_scores = task_manager.get_task_scores(current_phase)
|
||||
|
||||
# 在待执行任务中找到分数最低的任务
|
||||
lowest_score_task = None
|
||||
lowest_score = float('inf')
|
||||
|
||||
for task in pending_tasks:
|
||||
task_name = task.get("name", "")
|
||||
task_score = phase_scores.get(task_name, 0.0)
|
||||
|
||||
if task_score < lowest_score:
|
||||
lowest_score = task_score
|
||||
lowest_score_task = task
|
||||
|
||||
# 如果没有找到合适的任务，选择第一个并记录错误日志
|
||||
if lowest_score_task is None:
|
||||
# 使用logger记录错误，如果没有logger则使用print作为后备
|
||||
error_msg = f"Controller-ScoreDriven警告：在阶段{current_phase.value}中未找到合适任务，使用默认第一个任务"
|
||||
try:
|
||||
import logging
|
||||
logger = logging.getLogger(__name__)
|
||||
logger.error(error_msg)
|
||||
except:
|
||||
print(f"[ERROR] {error_msg}")
|
||||
|
||||
lowest_score_task = pending_tasks[0]
|
||||
lowest_score = phase_scores.get(lowest_score_task.get("name", ""), 0.0)
|
||||
|
||||
selected_task_name = lowest_score_task.get("name", "未知任务")
|
||||
|
||||
# 使用和simple模式相同的固定指导
|
||||
return ControllerDecision(
|
||||
selected_task=selected_task_name,
|
||||
specific_guidance="请按照标准医疗询问流程进行患者评估，基于患者临床信息选择最重要的询问任务，提供针对性的、具体的、可操作的询问指导建议，确保指导内容仅限于医生可以通过询问获取的信息。"
|
||||
)
|
||||
|
||||
def _get_simple_mode_result(self, pending_tasks: List[Dict[str, str]]) -> ControllerDecision:
|
||||
"""
|
||||
简化模式下生成决策结果
|
||||
|
||||
在简化模式下，直接选择第一个待执行任务，并返回固定的询问指导。
|
||||
|
||||
Args:
|
||||
pending_tasks (List[Dict[str, str]]): 待执行的任务列表
|
||||
|
||||
Returns:
|
||||
ControllerDecision: 包含简化模式任务选择和固定指导的结果
|
||||
"""
|
||||
# 如果有待执行任务，选择第一个作为默认任务
|
||||
if pending_tasks:
|
||||
selected_task = pending_tasks[0]
|
||||
selected_task_name = selected_task.get("name", "未知任务")
|
||||
else:
|
||||
selected_task_name = "基本信息收集"
|
||||
|
||||
return ControllerDecision(
|
||||
selected_task=selected_task_name,
|
||||
specific_guidance="请按照标准医疗询问流程进行患者评估，基于患者临床信息选择最重要的询问任务，提供针对性的、具体的、可操作的询问指导建议，确保指导内容仅限于医生可以通过询问获取的信息。"
|
||||
)
|
||||
|
||||
def _get_fallback_result(self, pending_tasks: List[Dict[str, str]]) -> ControllerDecision:
|
||||
"""
|
||||
生成决策失败时的默认结果
|
||||
|
||||
@ -24,6 +24,14 @@ LLM_CONFIG = {
|
||||
"api_key": "gpustack_d402860477878812_9ec494a501497d25b565987754f4db8c" # Ollama不需要真实API密钥，任意字符串即可
|
||||
}
|
||||
},
|
||||
"phi4": {
|
||||
"class": "OpenAILike",
|
||||
"params": {
|
||||
"id": "microsoft/phi-4",
|
||||
"base_url": "http://127.0.0.1:8000/v1", # Ollama OpenAI兼容端点
|
||||
"api_key": "gpustack_d402860477878812_9ec494a501497d25b565987754f4db8c" # Ollama不需要真实API密钥，任意字符串即可
|
||||
}
|
||||
},
|
||||
"Qwen3-7B": {
|
||||
"class": "OpenAILike",
|
||||
"params": {
|
||||
|
||||
16
main.py
16
main.py
@ -101,7 +101,7 @@ def parse_arguments() -> argparse.Namespace:
|
||||
parser.add_argument(
|
||||
'--log-dir',
|
||||
type=str,
|
||||
default='results/results0905-gemma3',
|
||||
default='results/results09010-score_driven',
|
||||
help='日志文件保存目录'
|
||||
)
|
||||
parser.add_argument(
|
||||
@ -115,7 +115,7 @@ def parse_arguments() -> argparse.Namespace:
|
||||
parser.add_argument(
|
||||
'--num-threads',
|
||||
type=int,
|
||||
default=40,
|
||||
default=45,
|
||||
help='并行处理线程数'
|
||||
)
|
||||
parser.add_argument(
|
||||
@ -149,7 +149,7 @@ def parse_arguments() -> argparse.Namespace:
|
||||
'--model-type',
|
||||
type=str,
|
||||
choices=available_models,
|
||||
default='Gemma3-4b',
|
||||
default='openai-mirror/gpt-oss-20b',
|
||||
help=f'使用的语言模型类型，可选: {", ".join(available_models)}'
|
||||
)
|
||||
parser.add_argument(
|
||||
@ -163,6 +163,13 @@ def parse_arguments() -> argparse.Namespace:
|
||||
default=None,
|
||||
help='模型配置JSON字符串（可选，覆盖默认配置）'
|
||||
)
|
||||
parser.add_argument(
|
||||
'--controller-mode',
|
||||
type=str,
|
||||
choices=['normal', 'sequence', 'score_driven'],
|
||||
default='score_driven',
|
||||
help='任务控制器模式：normal为智能模式（需要LLM推理），sequence为顺序模式（直接选择第一个任务），score_driven为分数驱动模式（选择当前任务组中分数最低的任务）'
|
||||
)
|
||||
|
||||
|
||||
# 调试和日志
|
||||
@ -347,7 +354,8 @@ def process_single_sample(sample_data: Dict[str, Any], sample_index: int,
|
||||
llm_config=llm_config,
|
||||
max_steps=args.max_steps,
|
||||
log_dir=args.log_dir,
|
||||
case_index=sample_index
|
||||
case_index=sample_index,
|
||||
controller_mode=args.controller_mode
|
||||
)
|
||||
|
||||
# 执行工作流
|
||||
|
||||
@ -12,7 +12,7 @@ class MedicalWorkflow:
|
||||
|
||||
def __init__(self, case_data: Dict[str, Any], model_type: str = "gpt-oss:latest",
|
||||
llm_config: Optional[Dict] = None, max_steps: int = 30, log_dir: str = "logs",
|
||||
case_index: Optional[int] = None):
|
||||
case_index: Optional[int] = None, controller_mode: str = "normal"):
|
||||
"""
|
||||
初始化医疗问诊工作流
|
||||
|
||||
@ -23,6 +23,7 @@ class MedicalWorkflow:
|
||||
max_steps: 最大执行步数，默认为30
|
||||
log_dir: 日志目录，默认为"logs"
|
||||
case_index: 病例序号，用于日志文件命名
|
||||
controller_mode: 任务控制器模式，'normal'为智能模式，'sequence'为顺序模式，'score_driven'为分数驱动模式
|
||||
"""
|
||||
self.case_data = case_data
|
||||
self.model_type = model_type
|
||||
@ -31,7 +32,7 @@ class MedicalWorkflow:
|
||||
|
||||
# 初始化核心组件
|
||||
self.task_manager = TaskManager()
|
||||
self.step_executor = StepExecutor(model_type=model_type, llm_config=self.llm_config)
|
||||
self.step_executor = StepExecutor(model_type=model_type, llm_config=self.llm_config, controller_mode=controller_mode)
|
||||
self.logger = WorkflowLogger(case_data=case_data, log_dir=log_dir, case_index=case_index)
|
||||
|
||||
# 重置历史评分，确保新的工作流从零开始
|
||||
|
||||
@ -41,25 +41,35 @@ class StepExecutor:
|
||||
"chief_complaint_similarity": 0.0
|
||||
}
|
||||
|
||||
def __init__(self, model_type: str = "gpt-oss:latest", llm_config: dict = None):
|
||||
def __init__(self, model_type: str = "gpt-oss:latest", llm_config: dict = None, controller_mode: str = "normal"):
|
||||
"""
|
||||
初始化step执行器
|
||||
|
||||
Args:
|
||||
model_type: 使用的语言模型类型（除Evaluator外的所有agent使用）
|
||||
llm_config: 语言模型配置
|
||||
controller_mode: 任务控制器模式，'normal'为智能模式，'sequence'为顺序模式，'score_driven'为分数驱动模式
|
||||
|
||||
Note:
|
||||
Evaluator agent 固定使用 gpt-oss:latest 模型，不受 model_type 参数影响
|
||||
"""
|
||||
self.model_type = model_type
|
||||
self.llm_config = llm_config or {}
|
||||
self.controller_mode = controller_mode
|
||||
|
||||
# 初始化所有agent
|
||||
self.recipient = RecipientAgent(model_type=model_type, llm_config=self.llm_config)
|
||||
self.triager = TriageAgent(model_type=model_type, llm_config=self.llm_config)
|
||||
self.monitor = Monitor(model_type=model_type, llm_config=self.llm_config)
|
||||
self.controller = TaskController(model_type=model_type, llm_config=self.llm_config)
|
||||
# 根据模式初始化TaskController
|
||||
simple_mode = (controller_mode == "sequence")
|
||||
score_driven_mode = (controller_mode == "score_driven")
|
||||
self.controller = TaskController(
|
||||
model_type=model_type,
|
||||
llm_config=self.llm_config,
|
||||
simple_mode=simple_mode,
|
||||
score_driven_mode=score_driven_mode
|
||||
)
|
||||
self.prompter = Prompter(model_type=model_type, llm_config=self.llm_config)
|
||||
self.virtual_patient = VirtualPatientAgent(model_type=model_type, llm_config=self.llm_config)
|
||||
# Evaluator 固定使用 gpt-oss:latest 模型
|
||||
@ -402,18 +412,28 @@ class StepExecutor:
|
||||
"pending_tasks": pending_tasks,
|
||||
"chief_complaint": recipient_result.chief_complaint,
|
||||
"hpi_content": recipient_result.updated_HPI,
|
||||
"ph_content": recipient_result.updated_PH
|
||||
"ph_content": recipient_result.updated_PH,
|
||||
"task_manager": task_manager # 传递task_manager用于score_driven模式
|
||||
}
|
||||
|
||||
result = self.controller.run(**input_data)
|
||||
execution_time = time.time() - start_time
|
||||
|
||||
# 为日志记录创建可序列化的input_data副本（移除TaskManager对象）
|
||||
log_input_data = {
|
||||
"pending_tasks": input_data["pending_tasks"],
|
||||
"chief_complaint": input_data["chief_complaint"],
|
||||
"hpi_content": input_data["hpi_content"],
|
||||
"ph_content": input_data["ph_content"]
|
||||
# 不包含task_manager，因为它不能JSON序列化
|
||||
}
|
||||
|
||||
output_data = {
|
||||
"selected_task": result.selected_task,
|
||||
"specific_guidance": result.specific_guidance
|
||||
}
|
||||
|
||||
logger.log_agent_execution(step_num, "controller", input_data, output_data, execution_time)
|
||||
logger.log_agent_execution(step_num, "controller", log_input_data, output_data, execution_time)
|
||||
|
||||
return result
|
||||
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user