diff --git a/packages/dbgpt-core/src/dbgpt/agent/expand/actions/deep_search_action.py b/packages/dbgpt-core/src/dbgpt/agent/expand/actions/deep_search_action.py
index 6bd9ed16d..14bdc160c 100644
--- a/packages/dbgpt-core/src/dbgpt/agent/expand/actions/deep_search_action.py
+++ b/packages/dbgpt-core/src/dbgpt/agent/expand/actions/deep_search_action.py
@@ -13,12 +13,12 @@ from dbgpt._private.pydantic import BaseModel, Field, model_to_dict
logger = logging.getLogger(__name__)
-class DeepSearchModel(BaseModel):
- """Chart item model."""
- status: str = Field(
- ...,
- description="The status of the current action, can be split_query, summary, or reflection.",
- )
+class SplitQueryModel(BaseModel):
+ """Model for splitting queries in deep search actions."""
+ # status: str = Field(
+ # ...,
+ # description="The status of the current action, can be split_query, summary, or reflection.",
+ # )
tools: List[dict] = Field(
default_factory=list,
description="List of tools to be used in the action.",
@@ -37,11 +37,40 @@ class DeepSearchModel(BaseModel):
return model_to_dict(self)
+class ReflectionModel(BaseModel):
+ """Model for Reflection."""
+ status: str = Field(
+ default_factory=list,
+ description="List of tools to be used in the action.",
+ )
+ knowledge_gap: str = Field(
+ ...,
+ description="The intention of the current action, describing what you want to achieve.",
+ )
+ sub_queries: List[str] = Field(
+ default_factory=list,
+ description="List of sub-queries generated from the current action.",
+ )
+ tools: List[dict] = Field(
+ default_factory=list,
+ description="List of tools to be used in the action.",
+ )
+ thought: str = Field(
+ ...,
+ description="The thought of the current action, describing what you want to achieve.",
+ )
+
+ def to_dict(self):
+ """Convert to dict."""
+ return model_to_dict(self)
+
+
class DeepSearchAction(ToolAction):
"""React action class."""
def __init__(self, **kwargs):
"""Tool action init."""
+ self.state = "split_query"
super().__init__(**kwargs)
@property
@@ -73,8 +102,9 @@ class DeepSearchAction(ToolAction):
) -> ActionOutput:
"""Perform the action."""
try:
- action_param: DeepSearchModel = self._input_convert(
- ai_message, DeepSearchModel
+ # state = "split_query"
+ action_param: ReflectionModel = self._input_convert(
+ ai_message, ReflectionModel
)
except Exception as e:
logger.exception(str(e))
@@ -83,29 +113,34 @@ class DeepSearchAction(ToolAction):
content="The requested correctly structured answer could not be found.",
)
- if action_param.status == "split_query":
- sub_queries = action_param.sub_queries
- # execute knowledge search
- if not action_param.tools:
- return ActionOutput(
- is_exe_success=False,
- content="No tools available for knowledge search.",
- )
- if action_param.tools:
- for tool in action_param.tools:
- if tool.get("tool_type") == "KnowledgeRetrieve":
- knowledge_args = action_param.get("args", {})
- if not knowledge_args:
- return ActionOutput(
- is_exe_success=False,
- content="No arguments provided for knowledge search.",
- )
- act_out = await self.knowledge_retrieve(
- sub_queries,
- knowledge_args,
- self.resource,
+ sub_queries = action_param.sub_queries
+ if action_param.status == "summarize":
+ return ActionOutput(
+ is_exe_success=True,
+ content=action_param.thought,
+ terminate=True,
+ )
+ if not action_param.tools:
+ return ActionOutput(
+ is_exe_success=False,
+ content="No tools available for knowledge search.",
+ )
+ if action_param.tools:
+ for tool in action_param.tools:
+ # state = "knowledge_search"
+ if tool.get("tool_type") == "KnowledgeRetrieve":
+ knowledge_args = tool.get("args", {})
+ if not knowledge_args:
+ return ActionOutput(
+ is_exe_success=False,
+ content="No arguments provided for knowledge search.",
)
-
+ act_out = await self.knowledge_retrieve(
+ sub_queries,
+ knowledge_args,
+ self.resource,
+ )
+ act_out.terminate = False
# if "parser" in kwargs and isinstance(kwargs["parser"], ReActOutputParser):
# parser = kwargs["parser"]
@@ -138,21 +173,17 @@ class DeepSearchAction(ToolAction):
lang=self.language, question=query
)
query_context_map[query] = resource_prompt
+ content = "\n".join([
+ f"{query}:{context}" for query, context in query_context_map.items()]
+ )
action_output = ActionOutput(
is_exe_success=True,
- content="\n".join([
- f"{query}:{context}" for query, context in query_context_map.items()]
- ),
- view="\n".join([
- f"{query}:{context}" for query, context in query_context_map.items()]
- ),
- observations=query_context_map,
+ content=content,
+ view=content,
+ observations=content,
)
return action_output
-
-
-
async def _do_run(
self,
ai_message: str,
diff --git a/packages/dbgpt-core/src/dbgpt/agent/expand/deep_search.py b/packages/dbgpt-core/src/dbgpt/agent/expand/deep_search.py
index b324ee76b..0f557937b 100644
--- a/packages/dbgpt-core/src/dbgpt/agent/expand/deep_search.py
+++ b/packages/dbgpt-core/src/dbgpt/agent/expand/deep_search.py
@@ -74,29 +74,36 @@ selecting the right search tools.
# The current time is: {{ now_time }}.
# """
_DEEPSEARCH_SYSTEM_TEMPLATE = """
-你是一个深度搜索助手。你的任务是你将用户原始问题一个或者多个子问题，并且给出可用知识库工具和搜索工具来回答问题或解决问题。
+你是一个深度搜索助手。
+<目标>
+你的任务是根据用户的问题或任务，选择合适的知识检索工具和搜索工具来回答问题或解决问题。
+你需要根据已经搜到的知识和搜索到的信息:
+{{most_recent_memories}}判断是否需要更多的知识或信息来回答问题。
+如果需要更多的知识或信息，你需要提出后续的子问题来扩展你的理解。
+
<可用工具>
1. KnowledgeRetrieve: 查询内部知识库以获取信息。\n可用知识库: {{knowledge_tools}}
2. WebSearch: 进行互联网搜索以获取最新或额外信息。\n 可用搜索工具: {{search_tools}}
-3. 总结: 对多个来源的信息进行总结和综合。
<流程>
1. 分析任务并创建搜索计划。
2. 选择使用一个或多个工具收集信息。
+3. 对收集到的信息进行反思，判断是否足够回答问题。
<回复格式>
严格按以下JSON格式输出，确保可直接解析：
{
- "status": "split_query (拆解搜索计划) | summary (仅当可用知识可以回答用户问题) | reflection (反思) "
"tools": [{
"tool_type": "工具类型"
"args": "args1",
}],
"intention": "当前你的意图,
- "sub_queries": [],
+ "sub_queries": ["子问题1", "子问题2"],
+ "knowledge_gap": "总结缺乏关于性能指标和基准的信息",
+ "status": "reflection(反思) | summarize(最后总结)",
}
@@ -116,13 +123,14 @@ _DEEPSEARCH_SYSTEM_TEMPLATE = """
: 2022年诺贝尔文学奖得主
返回
{
- "status": "split_query"
+ "status": "reflection"
"tools"?: [{
"tool_type": "KnowledgeRetrieve"
"args": "knowledge_name",
}],
"intention": "你的拆解意图,
- "sub_queries": [],
+ "knowledge_gap": "总结缺乏关于2022年诺贝尔文学奖得主的信息",
+ "sub_queries": ["子问题1","子问题2"],
}
@@ -135,6 +143,36 @@ _DEEPSEARCH_SYSTEM_TEMPLATE = """
"""
_DEEPSEARCH_USER_TEMPLATE = """"""
+_DEEPSEARCH_FINIAL_SUMMARY_TEMPLATE = """
+
+Generate a high-quality summary of the provided context.
+
+
+
+When creating a NEW summary:
+1. Highlight the most relevant information related to the user topic from the search results
+2. Ensure a coherent flow of information
+
+When EXTENDING an existing summary:
+{{most_recent_memories}}
+1. Read the existing summary and new search results carefully.
+2. Compare the new information with the existing summary.
+3. For each piece of new information:
+ a. If it's related to existing points, integrate it into the relevant paragraph.
+ b. If it's entirely new but relevant, add a new paragraph with a smooth transition.
+ c. If it's not relevant to the user topic, skip it.
+4. Ensure all additions are relevant to the user's topic.
+5. Verify that your final output differs from the input summary.
+< /REQUIREMENTS >
+
+< FORMATTING >
+- Start directly with the updated summary, without preamble or titles. Do not use XML tags in the output.
+< /FORMATTING >
+
+
+Think carefully about the provided Context first. Then generate a summary of the context to address the User Input.
+
+"""
_REACT_WRITE_MEMORY_TEMPLATE = """\
@@ -176,42 +214,7 @@ class DeepSearchAgent(ConversableAgent):
"""Init indicator AssistantAgent."""
super().__init__(**kwargs)
- self._init_actions([DeepSearchAction, Terminate])
-
- # async def _a_init_reply_message(
- # self,
- # received_message: AgentMessage,
- # rely_messages: Optional[List[AgentMessage]] = None,
- # ) -> AgentMessage:
- # reply_message = super()._init_reply_message(received_message, rely_messages)
- #
- # tool_packs = ToolPack.from_resource(self.resource)
- # action_space = []
- # action_space_names = []
- # action_space_simple_desc = []
- # if tool_packs:
- # tool_pack = tool_packs[0]
- # for tool in tool_pack.sub_resources:
- # tool_desc, _ = await tool.get_prompt(lang=self.language)
- # action_space_names.append(tool.name)
- # action_space.append(tool_desc)
- # if isinstance(tool, BaseTool):
- # tool_simple_desc = tool.description
- # else:
- # tool_simple_desc = tool.get_prompt()
- # action_space_simple_desc.append(f"{tool.name}: {tool_simple_desc}")
- # else:
- # for action in self.actions:
- # action_space_names.append(action.name)
- # action_space.append(action.get_action_description())
- # # self.actions
- # reply_message.context = {
- # "max_steps": self.max_retry_count,
- # "action_space": "\n".join(action_space),
- # "action_space_names": ", ".join(action_space_names),
- # "action_space_simple_desc": "\n".join(action_space_simple_desc),
- # }
- # return reply_message
+ self._init_actions([DeepSearchAction])
async def preload_resource(self) -> None:
await super().preload_resource()
@@ -289,14 +292,44 @@ class DeepSearchAgent(ConversableAgent):
"knowledge_desc": self.resource.retriever_desc,
})
- # new_resource = self.resource.apply(apply_func=_remove_tool)
- # if new_resource:
- # resource_prompt, resource_reference = await new_resource.get_prompt(
- # lang=self.language, question=question
- # )
- # return resource_prompt, resource_reference
return json.dumps(abilities, ensure_ascii=False), []
+ async def build_system_prompt(
+ self,
+ question: Optional[str] = None,
+ most_recent_memories: Optional[str] = None,
+ resource_vars: Optional[Dict] = None,
+ context: Optional[Dict[str, Any]] = None,
+ is_retry_chat: bool = False,
+ ):
+ """Build system prompt."""
+ system_prompt = None
+ if self.bind_prompt:
+ prompt_param = {}
+ if resource_vars:
+ prompt_param.update(resource_vars)
+ if context:
+ prompt_param.update(context)
+ if self.bind_prompt.template_format == "f-string":
+ system_prompt = self.bind_prompt.template.format(
+ **prompt_param,
+ )
+ elif self.bind_prompt.template_format == "jinja2":
+ system_prompt = Template(self.bind_prompt.template).render(prompt_param)
+ else:
+ logger.warning("Bind prompt template not exsit or format not support!")
+ if not system_prompt:
+ param: Dict = context if context else {}
+ system_prompt = await self.build_prompt(
+ question=question,
+ is_system=True,
+ most_recent_memories=most_recent_memories,
+ resource_vars=resource_vars,
+ is_retry_chat=is_retry_chat,
+ **param,
+ )
+ return system_prompt
+
def prepare_act_param(
self,
received_message: Optional[AgentMessage],
@@ -350,6 +383,8 @@ class DeepSearchAgent(ConversableAgent):
rely_action_out=last_out,
**kwargs,
)
+ if not last_out.terminate:
+ self.profile.system_prompt_template = _DEEPSEARCH_FINIAL_SUMMARY_TEMPLATE
span.metadata["action_out"] = last_out.to_dict() if last_out else None
if not last_out:
raise ValueError("Action should return value！")
@@ -366,7 +401,7 @@ class DeepSearchAgent(ConversableAgent):
) -> Union[str, List["AgentMessage"]]:
memories = await self.memory.read(observation)
not_json_memories = []
- messages = []
+ # messages = []
structured_memories = []
for m in memories:
if m.raw_observation:
@@ -381,46 +416,48 @@ class DeepSearchAgent(ConversableAgent):
except Exception:
not_json_memories.append(m.raw_observation)
- for mem_dict in structured_memories:
- question = mem_dict.get("question")
- thought = mem_dict.get("thought")
- action = mem_dict.get("action")
- action_input = mem_dict.get("action_input")
- observation = mem_dict.get("observation")
- if question:
- messages.append(
- AgentMessage(
- content=f"Question: {question}",
- role=ModelMessageRoleType.HUMAN,
- )
- )
- ai_content = []
- if thought:
- ai_content.append(f"Thought: {thought}")
- if action:
- ai_content.append(f"Action: {action}")
- if action_input:
- ai_content.append(f"Action Input: {action_input}")
- messages.append(
- AgentMessage(
- content="\n".join(ai_content),
- role=ModelMessageRoleType.AI,
- )
- )
-
- if observation:
- messages.append(
- AgentMessage(
- content=f"Observation: {observation}",
- role=ModelMessageRoleType.HUMAN,
- )
- )
-
- if not messages and not_json_memories:
- messages.append(
- AgentMessage(
- content="\n".join(not_json_memories),
- role=ModelMessageRoleType.HUMAN,
- )
- )
- return messages
+ # for mem_dict in structured_memories:
+ # question = mem_dict.get("question")
+ # thought = mem_dict.get("thought")
+ # action = mem_dict.get("action")
+ # action_input = mem_dict.get("action_input")
+ # observation = mem_dict.get("observation")
+ # if question:
+ # messages.append(
+ # AgentMessage(
+ # content=f"Question: {question}",
+ # role=ModelMessageRoleType.HUMAN,
+ # )
+ # )
+ # ai_content = []
+ # if thought:
+ # ai_content.append(f"Thought: {thought}")
+ # if action:
+ # ai_content.append(f"Action: {action}")
+ # if action_input:
+ # ai_content.append(f"Action Input: {action_input}")
+ # messages.append(
+ # AgentMessage(
+ # content="\n".join(ai_content),
+ # role=ModelMessageRoleType.AI,
+ # )
+ # )
+ #
+ # if observation:
+ # messages.append(
+ # AgentMessage(
+ # content=f"Observation: {observation}",
+ # role=ModelMessageRoleType.HUMAN,
+ # )
+ # )
+ #
+ # if not messages and not_json_memories:
+ # messages.append(
+ # AgentMessage(
+ # content="\n".join(not_json_memories),
+ # role=ModelMessageRoleType.HUMAN,
+ # )
+ # )
+ return "\n".join([
+ mem_dict.get("observation") for mem_dict in structured_memories
+ ])