init wip prompt tuning

.pre-commit-config.yaml

@@ -56,6 +56,7 @@ repos:
     rev: 6.1.0
     hooks:
       - id: flake8
+        args: ["--extend-ignore=E203"]
   - repo: https://github.com/pylint-dev/pylint
     rev: v3.0.2
     hooks:

src/agentscope/agents/direct_prompt_opt_agent.py

@@ -0,0 +1,122 @@
+# -*- coding: utf-8 -*-
+# pylint: disable=C0301
+"""A agent that performs direct prompt optimization."""
+from typing import Union
+
+from agentscope.message import Msg
+from agentscope.agents.agent import AgentBase
+
+
+OPT_QUERY_PROMPT = """
+你是一个专业的prompt工程师，你擅长优化prompt优化。你的任务是优化用户提供的prompt, 使得优化后的prompt指示更清晰，结构更明确。
+
+请注意：
+1. 优化后的prompt必须与用户提供的prompt意图一致，可适当加入上下文或任何可以缩小范围并指导大模型能够更好地理解完成任务的附加信息，对用户的prompt进行重构。请注意不要作过多的拓展。
+2. 优化后的prompt要保留用户提供的prompt里的关键信息, 例如原prompt中的与任务相关的背景知识，文本分析任务中的原文本，关于输出格式的要求等类型的关键信息。
+3. 当prompt比较长的时候，可以适当在其中加入分隔符，使得优化后的prompt结构更加清晰。
+4. 如果用户的prompt里还有变量，如"${variable_name}"，优化后的prompt里必须保留这些变量。你可以加入更多的用户可配置的变量, 并用"${new_variable_name}"来表示, 使得优化后的prompt支持用户提供更多的信息。
+5. 优化后的prompt语言与用户提供的prompt一致，即用户提供的prompt使用中文写的，优化后的prompt也必须是中文, 如果用户提供的prompt使用英文写的，优化后的prompt也必须是英文。
+6. 如果你认为优化前的prompt已经足够简介清晰，且能够很好的表达用户对应的意图，那么就不需要优化prompt，直接返回用户输入提供的prompt即可。
+7. 你不能直接生成对原始prompt的回答！
+8. 相比直接使用原始prompt，使用你优化后的prompt时大模型应该能生成更好的、更符合用户意图的回答。
+9. 你的输出应该只包含优化后的prompt，而不带其他附带内容。
+
+"""  # noqa
+
+OPT_PROMPT_TEMPLATE = """
+用户提供的prompt是：
+{user_prompt}
+
+现在，请输出你优化后的prompt:
+"""
+
+
+class DirectPromptOptimizationAgent(AgentBase):
+    """A simple agent that directly optimizes the prompt."""
+
+    def __init__(
+        self,
+        name: str,
+        model_config_name: str,
+        meta_prompt: Union[str, None] = None,
+    ) -> None:
+        """Initialize the direct prompt optimization agent.
+
+        Arguments:
+            name (`str`):
+                The name of the agent.
+            model_config_name (`str`):
+                The name of the model config, which is used to load model from
+                configuration.
+            meta_prompt (`Optional[str]`):
+                The meta prompt that instruct the agent to perform prompt
+                optimization. If is None, then the agent will use the default
+                prompt above.
+
+        Note:
+            The output optimized prompt may not always works better than the
+            original prompt. It depends on the specific task and the model.
+
+        Usage:
+            ```
+            from direct_agent import DirectPromptOptimizationAgent
+            from agentscope.message import Msg
+
+            agent = DirectPromptOptimizationAgent(
+                name="assistant",
+                model_config_name='xxx',
+            )
+
+            user_prompt = "Tell me about the history of the world."
+
+            optimized_prompt = agent(Msg(name="user",
+                content=user_prompt, role="user").content
+            ```
+        """
+        super().__init__(
+            name=name,
+            model_config_name=model_config_name,
+        )
+
+        if meta_prompt is None:
+            self.meta_prompt = OPT_QUERY_PROMPT
+        else:
+            self.meta_prompt = meta_prompt
+
+    def reply(self, x: dict = None) -> dict:
+        """
+        Replying to the input.
+
+        Arguments:
+            x(`Msg`): the input prompt to optimize. The input prompt
+                should be a `Msg` object.
+
+        Returns:
+            msg(`Msg`): the output message. In the output message,
+                the `content` field is the optimized prompt.
+
+        Note:
+            The output optimized prompt may not always works better than the
+            original prompt. It depends on the specific task and the model.
+        """
+        # get the user prompt
+        user_prompt = x.content
+
+        # query the llm using meta prompt and template
+        # call llm and generate response
+        prompt = self.model.format(
+            Msg(
+                "user",
+                self.meta_prompt
+                + OPT_PROMPT_TEMPLATE.format(user_prompt=user_prompt),
+                role="user",
+            ),
+        )
+        response = self.model(prompt).text
+
+        # Print/speak the message in this agent's voice
+        self.speak(Msg(self.name, "Optimizing Prompt", role="assistant"))
+        msg = Msg(self.name, response, role="assistant")
+        self.speak(msg)
+
+        return msg

src/agentscope/agents/system_prompt_opt_agent.py

@@ -0,0 +1,143 @@
+# -*- coding: utf-8 -*-
+# pylint: disable=C0301
+"""A agent that optimize system prompt given user dialog history."""
+from agentscope.message import Msg
+from agentscope.agents.agent import AgentBase
+
+
+OPT_PROMPT_TEMPLATE = """
+你是一个优秀的Prompt Engineer，现在你要通过添加note的方式对一个Agent的system prompt进行优化。
+
+用户提供的原始system prompt是：
+{system_prompt}
+
+用户与之交互的dialog history是：
+{dialog_history}
+
+现在，你要
+1. 判断用户与agent交互的dialog history中，是否包含显式的错误（如函数调用错误、没有遵循输入输出格式），对用户意图的误解等。
+2. 对发生错误的原因进行详细分析，并且寻找对于对应错误的解决方案。
+3. 根据错误原因和用户意图，写一条或几条可以添加在用户system prompt后面的注意事项note，或者exmaple形式的note，使之不要再犯同样的错误。
+如果要添加的note包含example，需要格外小心，如果不确定添加的example是否正确，可以先不添加。
+你添加的note应该包含在tag [prompt_note]中，例如 [prompt_note] 请注意输出仅包含json格式 [/prompt_note]。如果dialog history没有明显问题，则不需要添加任何note。
+"""  # noqa
+
+# TODO 添加example prompt：下面是几个对话示例，如何对System Prompt进行修改。
+
+
+class SystemPromptOptimizationAgent(AgentBase):
+    """A simple agent that optimize system prompt given user dialog history."""
+
+    def __init__(
+        self,
+        name: str,
+        model_config_name: str,
+    ) -> None:
+        """Initialize the direct prompt optimization agent.
+
+        Arguments:
+            name (`str`):
+                The name of the agent.
+            model_config_name (`str`):
+                The name of the model config, which is used to load model from
+                configuration.
+
+        Note:
+            Adding the output optimized note to the system prompt may not
+            always solve the issue. It depends on the specific task
+            and the model.
+        """
+        super().__init__(
+            name=name,
+            model_config_name=model_config_name,
+        )
+
+    def reply(self, x: dict = None) -> dict:
+        """
+        Replying to the input.
+
+        Arguments:
+            x(`Msg`): the input prompt to optimize. The input prompt
+                should be a `Msg` object. The `content` field should be a dict,
+                which contains the following fields:
+                    - system_prompt(`str`): the system prompt to optimize.
+                    - dialog_history(`str` or List[Msg]):
+                    the dialog history of the user and the agent.
+
+        Returns:
+            msg(`Msg`): the output message. In the output message,
+                the `content` field is the optimized prompt.
+
+        Note:
+            The output optimized prompt may not always works better than the
+            original prompt. It depends on the specific task and the model.
+        """
+        # get the system prompt and dialog history
+        assert isinstance(
+            x.content,
+            dict,
+        ), "The input prompt should be a dict."
+        assert (
+            "dialog_history" in x.content and "system_prompt" in x.content
+        ), "The input should have 'dialog_history' amd 'system_prompt' fields."
+
+        system_prompt = x.content["system_prompt"]
+        dialog_history = x.content["dialog_history"]
+
+        assert type(dialog_history) in [
+            str,
+            list,
+        ], "The input 'dialog_history' should be a str or list."
+        assert isinstance(
+            system_prompt,
+            str,
+        ), "The input 'system_prompt' should be a str."
+
+        # query the llm using prompt template
+        # call llm and generate response
+        prompt = self.model.format(
+            Msg(
+                "user",
+                OPT_PROMPT_TEMPLATE.format(
+                    system_prompt=system_prompt,
+                    dialog_history=dialog_history,
+                ),
+                role="user",
+            ),
+        )
+        response = self.model(prompt).text
+
+        # Print/speak the message in this agent's voice
+        self.speak(
+            Msg(self.name, "Optimizing System Prompt", role="assistant"),
+        )
+        msg = Msg(self.name, response, role="assistant")
+        self.speak(msg)
+        # get all the notes and return
+        return Msg(
+            self.name,
+            self.get_all_tagged_notes(response),
+            role="assistant",
+        )
+
+    def get_all_tagged_notes(self, response_text: str) -> list:
+        """Get all the notes in the response text."""
+        notes = []
+        start_tag = "[prompt_note]"
+        end_tag = "[/prompt_note]"
+        start_index = response_text.find(start_tag)
+        while start_index != -1:
+            end_index = response_text.find(
+                end_tag,
+                start_index + len(start_tag),
+            )
+            if end_index != -1:
+                note = response_text[start_index + len(start_tag) : end_index]
+                notes.append(note)
+                start_index = response_text.find(
+                    start_tag,
+                    end_index + len(end_tag),
+                )
+            else:
+                break
+        return notes