erniebot-agent/erniebot_agent/agents/base.py

# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import abc
import inspect
import json
from typing import Any, Dict, List, Literal, Optional, Union

from erniebot_agent.agents.callback.callback_manager import CallbackManager
from erniebot_agent.agents.callback.default import get_default_callbacks
from erniebot_agent.agents.callback.handlers.base import CallbackHandler
from erniebot_agent.agents.schema import (
    AgentFile,
    AgentResponse,
    LLMResponse,
    ToolResponse,
)
from erniebot_agent.chat_models.base import ChatModel
from erniebot_agent.file_io.file_manager import FileManager
from erniebot_agent.file_io.protocol import is_local_file_id, is_remote_file_id
from erniebot_agent.memory.base import Memory
from erniebot_agent.messages import Message, SystemMessage
from erniebot_agent.tools.base import Tool
from erniebot_agent.tools.tool_manager import ToolManager
from erniebot_agent.utils.logging import logger


class BaseAgent(metaclass=abc.ABCMeta):
    llm: ChatModel
    memory: Memory

    @abc.abstractmethod
    async def async_run(self, prompt: str) -> AgentResponse:
        raise NotImplementedError


class Agent(BaseAgent):
    def __init__(
        self,
        llm: ChatModel,
        tools: Union[ToolManager, List[Tool]],
        memory: Memory,
        system_message: Optional[SystemMessage] = None,
        *,
        callbacks: Optional[Union[CallbackManager, List[CallbackHandler]]] = None,
        file_manager: Optional[FileManager] = None,
    ) -> None:
        super().__init__()
        self.llm = llm
        self.memory = memory
        # 1. Get system message exist in memory
        # OR 2. overwrite by the system_message paased in the Agent.
        if system_message:
            self.system_message = system_message
        else:
            self.system_message = memory.get_system_message()
        if isinstance(tools, ToolManager):
            self._tool_manager = tools
        else:
            self._tool_manager = ToolManager(tools)
        if callbacks is None:
            callbacks = get_default_callbacks()
        if isinstance(callbacks, CallbackManager):
            self._callback_manager = callbacks
        else:
            self._callback_manager = CallbackManager(callbacks)
        self.file_manager = file_manager

    async def async_run(self, prompt: str) -> AgentResponse:
        await self._callback_manager.on_run_start(agent=self, prompt=prompt)
        agent_resp = await self._async_run(prompt)
        await self._callback_manager.on_run_end(agent=self, response=agent_resp)
        return agent_resp

    def load_tool(self, tool: Tool) -> None:
        self._tool_manager.add_tool(tool)

    def unload_tool(self, tool: Tool) -> None:
        self._tool_manager.remove_tool(tool)

    def reset_memory(self) -> None:
        self.memory.clear_chat_history()

    def launch_gradio_demo(self, **launch_kwargs: Any):
        # TODO: Unified optional dependencies management
        try:
            import gradio as gr
        except ImportError:
            raise ImportError(
                "Could not import gradio, which is required for `launch_gradio_demo()`."
                " Please run `pip install erniebot-agent[gradio]` to install the optional dependencies."
            ) from None

        raw_messages = []

        def _pre_chat(text, history):
            history.append([text, None])
            return history, gr.update(value="", interactive=False), gr.update(interactive=False)

        async def _chat(history):
            prompt = history[-1][0]
            if len(prompt) == 0:
                raise gr.Error("Prompt should not be empty.")
            response = await self.async_run(prompt)
            history[-1][1] = response.text
            raw_messages.extend(response.chat_history)
            return (
                history,
                _messages_to_dicts(raw_messages),
                _messages_to_dicts(self.memory.get_messages()),
            )

        def _post_chat():
            return gr.update(interactive=True), gr.update(interactive=True)

        def _clear():
            raw_messages.clear()
            self.reset_memory()
            return None, None, None, None

        def _messages_to_dicts(messages):
            return [message.to_dict() for message in messages]

        with gr.Blocks(
            title="ERNIE Bot Agent Demo", theme=gr.themes.Soft(spacing_size="sm", text_size="md")
        ) as demo:
            with gr.Column():
                chatbot = gr.Chatbot(
                    label="Chat history",
                    latex_delimiters=[
                        {"left": "$$", "right": "$$", "display": True},
                        {"left": "$", "right": "$", "display": False},
                    ],
                    bubble_full_width=False,
                )
                prompt_textbox = gr.Textbox(label="Prompt", placeholder="Write a prompt here...")
                with gr.Row():
                    submit_button = gr.Button("Submit")
                    clear_button = gr.Button("Clear")
                with gr.Accordion("Tools", open=False):
                    attached_tools = self._tool_manager.get_tools()
                    tool_descriptions = [tool.function_call_schema() for tool in attached_tools]
                    gr.JSON(value=tool_descriptions)
                with gr.Accordion("Raw messages", open=False):
                    all_messages_json = gr.JSON(label="All messages")
                    agent_memory_json = gr.JSON(label="Messges in memory")
            prompt_textbox.submit(
                _pre_chat,
                inputs=[prompt_textbox, chatbot],
                outputs=[chatbot, prompt_textbox, submit_button],
            ).then(
                _chat,
                inputs=[chatbot],
                outputs=[
                    chatbot,
                    all_messages_json,
                    agent_memory_json,
                ],
            ).then(
                _post_chat, outputs=[prompt_textbox, submit_button]
            )
            submit_button.click(
                _pre_chat,
                inputs=[prompt_textbox, chatbot],
                outputs=[chatbot, prompt_textbox, submit_button],
            ).then(
                _chat,
                inputs=[chatbot],
                outputs=[
                    chatbot,
                    all_messages_json,
                    agent_memory_json,
                ],
            ).then(
                _post_chat, outputs=[prompt_textbox, submit_button]
            )
            clear_button.click(
                _clear,
                outputs=[
                    chatbot,
                    prompt_textbox,
                    all_messages_json,
                    agent_memory_json,
                ],
            )

        demo.launch(**launch_kwargs)

    @abc.abstractmethod
    async def _async_run(self, prompt: str) -> AgentResponse:
        raise NotImplementedError

    async def _async_run_tool(self, tool_name: str, tool_args: str) -> ToolResponse:
        tool = self._tool_manager.get_tool(tool_name)
        await self._callback_manager.on_tool_start(agent=self, tool=tool, input_args=tool_args)
        try:
            tool_resp = await self._async_run_tool_without_hooks(tool, tool_args)
        except (Exception, KeyboardInterrupt) as e:
            await self._callback_manager.on_tool_error(agent=self, tool=tool, error=e)
            raise
        await self._callback_manager.on_tool_end(agent=self, tool=tool, response=tool_resp)
        return tool_resp

    async def _async_run_llm(self, messages: List[Message], **opts: Any) -> LLMResponse:
        await self._callback_manager.on_llm_start(agent=self, llm=self.llm, messages=messages)
        try:
            llm_resp = await self._async_run_llm_without_hooks(messages, **opts)
        except (Exception, KeyboardInterrupt) as e:
            await self._callback_manager.on_llm_error(agent=self, llm=self.llm, error=e)
            raise
        await self._callback_manager.on_llm_end(agent=self, llm=self.llm, response=llm_resp)
        return llm_resp

    async def _async_run_tool_without_hooks(self, tool: Tool, tool_args: str) -> ToolResponse:
        bnd_args = self._parse_tool_args(tool, tool_args)
        # XXX: Sniffing is less efficient and probably unnecessary.
        # Can we make a protocol to statically recognize file inputs and outputs
        # or can we have the tools introspect about this?
        input_files = await self._sniff_and_extract_files_from_args(bnd_args.arguments, tool, "input")
        tool_ret = await tool(*bnd_args.args, **bnd_args.kwargs)
        output_files = await self._sniff_and_extract_files_from_args(tool_ret, tool, "output")
        tool_ret_json = json.dumps(tool_ret, ensure_ascii=False)
        return ToolResponse(json=tool_ret_json, files=input_files + output_files)

    async def _async_run_llm_without_hooks(
        self, messages: List[Message], functions=None, **opts: Any
    ) -> LLMResponse:
        llm_ret = await self.llm.async_chat(messages, functions=functions, stream=False, **opts)
        return LLMResponse(message=llm_ret)

    def _parse_tool_args(self, tool: Tool, tool_args: str) -> inspect.BoundArguments:
        args_dict = json.loads(tool_args)
        if not isinstance(args_dict, dict):
            raise ValueError("`tool_args` cannot be interpreted as a dict.")
        # TODO: Check types
        sig = inspect.signature(tool.__call__)
        bnd_args = sig.bind(**args_dict)
        bnd_args.apply_defaults()
        return bnd_args

    async def _sniff_and_extract_files_from_args(
        self, args: Dict[str, Any], tool: Tool, file_type: Literal["input", "output"]
    ) -> List[AgentFile]:
        agent_files: List[AgentFile] = []
        for val in args.values():
            if isinstance(val, str):
                if is_local_file_id(val):
                    if self.file_manager is None:
                        logger.warning(
                            f"A file is used by {repr(tool)}, but the agent has no file manager to fetch it."
                        )
                        continue
                    file = self.file_manager.look_up_file_by_id(val)
                    if file is None:
                        raise RuntimeError(f"Unregistered ID {repr(val)} is used by {repr(tool)}.")
                elif is_remote_file_id(val):
                    if self.file_manager is None:
                        logger.warning(
                            f"A file is used by {repr(tool)}, but the agent has no file manager to fetch it."
                        )
                        continue
                    file = self.file_manager.look_up_file_by_id(val)
                    if file is None:
                        file = await self.file_manager.retrieve_remote_file_by_id(val)
                else:
                    continue
                agent_files.append(AgentFile(file=file, type=file_type, used_by=tool.tool_name))
        return agent_files