agentscope.service.service_toolkit 源代码

# -*- coding: utf-8 -*-
"""Service Toolkit for service function usage."""
import json
from functools import partial
import inspect
from typing import (
    Callable,
    Any,
    Tuple,
    Union,
    Optional,
    Literal,
    get_args,
    get_origin,
    List,
)
from loguru import logger

from ..exception import (
    JsonParsingError,
    FunctionNotFoundError,
    FunctionCallFormatError,
    FunctionCallError,
)
from .service_response import ServiceResponse
from .service_response import ServiceExecStatus
from ..message import Msg

try:
    from docstring_parser import parse
except ImportError:
    parse = None


def _get_type_str(cls: Any) -> Optional[Union[str, list]]:
    """Get the type string."""
    type_str = None
    if hasattr(cls, "__origin__"):
        # Typing class
        if cls.__origin__ is Union:
            type_str = [_get_type_str(_) for _ in get_args(cls)]
            clean_type_str = [_ for _ in type_str if _ != "null"]
            if len(clean_type_str) == 1:
                type_str = clean_type_str[0]
        elif cls.__origin__ in [list, tuple]:
            type_str = "array"
        else:
            type_str = str(cls.__origin__)
    else:
        # Normal class
        if cls is str:
            type_str = "string"
        elif cls in [float, int, complex]:
            type_str = "number"
        elif cls is bool:
            type_str = "boolean"
        elif cls in [list, tuple]:
            type_str = "array"
        elif cls is None.__class__:
            type_str = "null"
        elif cls is Any:
            type_str = "Any"
        else:
            type_str = cls.__name__

    return type_str  # type: ignore[return-value]


[文档] class ServiceFunction: """The service function class.""" name: str """The name of the service function.""" original_func: Callable """The original function before processing.""" processed_func: Callable """The processed function that can be called by the model directly.""" json_schema: dict """The JSON schema description of the service function.""" require_args: bool """Whether calling the service function requests arguments. Some arguments may have default values, so it is not necessary to provide all arguments. """
[文档] def __init__( self, name: str, original_func: Callable, processed_func: Callable, json_schema: dict, ) -> None: """Initialize the service function object.""" self.name = name self.original_func = original_func self.processed_func = processed_func self.json_schema = json_schema self.require_args = ( len( json_schema["function"] .get("parameters", {}) .get("required", []), ) != 0 )
[文档] class ServiceToolkit: """A service toolkit class that turns service function into string prompt format.""" service_funcs: dict[str, ServiceFunction] """The registered functions in the service toolkit.""" _tools_instruction_format: str = ( "## Tool Functions:\n" "The following tool functions are available in the format of\n" "```\n" "{{index}}. {{function name}}: {{function description}}\n" "{{argument1 name}} ({{argument type}}): {{argument description}}\n" "{{argument2 name}} ({{argument type}}): {{argument description}}\n" "...\n" "```\n\n" "{function_prompt}\n" ) """The instruction template for the tool functions.""" _tools_calling_format: str = ( '[{"name": "{function name}", "arguments": {"{argument1 name}": xxx,' ' "{argument2 name}": xxx}}]' ) """The format of the tool function call.""" _tools_execution_format: str = ( "{index}. Execute function {function_name}\n" " [ARGUMENTS]:\n" " {arguments}\n" " [STATUS]: {status}\n" " [RESULT]: {result}\n" ) """The prompt template for the execution results."""
[文档] def __init__(self) -> None: """Initialize the service toolkit with a list of service functions.""" self.service_funcs = {}
[文档] def add(self, service_func: Callable[..., Any], **kwargs: Any) -> None: """Add a service function to the toolkit, which will be processed into a tool function that can be called by the model directly, and registered in processed_funcs. Args: service_func (`Callable[..., Any]`): The service function to be called. kwargs (`Any`): The arguments to be passed to the service function. Returns: `Tuple(Callable[..., Any], dict)`: A tuple of tool function and a dict in JSON Schema format to describe the function. Note: The description of the function and arguments are extracted from its docstring automatically, which should be well-formatted in **Google style**. Otherwise, their descriptions in the returned dictionary will be empty. Suggestions: 1. The name of the service function should be self-explanatory, so that the agent can understand the function and use it properly. 2. The typing of the arguments should be provided when defining the function (e.g. `def func(a: int, b: str, c: bool)`), so that the agent can specify the arguments properly. 3. The execution results should be a `ServiceResponse` object. Example: .. code-block:: python def bing_search(query: str, api_key: str, num_results=10): \"""Search the query in Bing search engine. Args: query: (`str`): The string query to search. api_key: (`str`): The API key for Bing search. num_results: (`int`, optional): The number of results to return, default to 10. \""" # ... Your implementation here ... return ServiceResponse(status, output) """ # TODO: hotfix for workstation, will be removed in the future if isinstance(service_func, partial): self.add(service_func.func, **service_func.keywords) return processed_func, json_schema = ServiceToolkit.get( service_func, **kwargs, ) # register the service function name = service_func.__name__ if name in self.service_funcs: logger.warning( f"Service function `{name}` already exists, " f"skip adding it.", ) else: self.service_funcs[name] = ServiceFunction( name=name, original_func=service_func, processed_func=processed_func, json_schema=json_schema, )
@property def json_schemas(self) -> dict: """The json schema descriptions of the processed service funcs.""" return {k: v.json_schema for k, v in self.service_funcs.items()} @property def tools_calling_format(self) -> str: """The calling format of the tool functions.""" return self._tools_calling_format @property def tools_instruction(self) -> str: """The instruction of the tool functions.""" tools_prompt = [] for i, (func_name, desc) in enumerate(self.json_schemas.items()): func_desc = desc["function"]["description"] args_desc = desc["function"]["parameters"]["properties"] args_list = [f"{i + 1}. {func_name}: {func_desc}"] for args_name, args_info in args_desc.items(): if "type" in args_info: args_line = ( f'\t{args_name} ({args_info["type"]}): ' f'{args_info.get("description", "")}' ) else: args_line = ( f'\t{args_name}: {args_info.get("description", "")}' ) args_list.append(args_line) func_prompt = "\n".join(args_list) tools_prompt.append(func_prompt) tools_description = "\n".join(tools_prompt) if tools_description == "": # No tools are provided return "" else: return self._tools_instruction_format.format_map( {"function_prompt": tools_description}, ) def _parse_and_check_text( # pylint: disable=too-many-branches self, cmd: Union[list[dict], str], ) -> List[dict]: """Parsing and check the format of the function calling text.""" # Record the error error_info = [] if isinstance(cmd, str): # --- Syntax check: if the input can be loaded by JSON try: processed_text = cmd.strip() # complete "[" and "]" if they are missing index_start = processed_text.find("[") index_end = processed_text.rfind("]") if index_start == -1: index_start = 0 error_info.append('Missing "[" at the beginning.') if index_end == -1: index_end = len(processed_text) error_info.append('Missing "]" at the end.') # remove the unnecessary prefix before "[" and suffix after "]" processed_text = processed_text[ index_start : index_end + 1 # noqa: E203 ] cmds = json.loads(processed_text) except json.JSONDecodeError: # Since we have processed the text, here we can only report # the JSON parsing error raise JsonParsingError( f"Except a list of dictionaries in JSON format, " f"like: {self.tools_calling_format}", ) from None else: cmds = cmd # --- Semantic Check: if the input is a list of dicts with # required fields # Handle the case when the input is a single dictionary if isinstance(cmds, dict): # The error info is already recorded in error_info cmds = [cmds] if not isinstance(cmds, list): # Not list, raise parsing error raise JsonParsingError( f"Except a list of dictionaries in JSON format " f"like: {self.tools_calling_format}", ) # --- Check the format of the command --- for sub_cmd in cmds: if not isinstance(sub_cmd, dict): raise JsonParsingError( f"Except a JSON list of dictionaries, but got" f" {type(sub_cmd)} instead.", ) if "name" not in sub_cmd: raise FunctionCallFormatError( "The field 'name' is required in the dictionary.", ) # Obtain the service function func_name = sub_cmd["name"] # Cannot find the service function if func_name not in self.service_funcs: raise FunctionNotFoundError( f"Cannot find a tool function named `{func_name}`.", ) # If it is json(str) convert to json(dict) if isinstance(sub_cmd["arguments"], str): try: sub_cmd["arguments"] = json.loads(sub_cmd["arguments"]) except json.decoder.JSONDecodeError: logger.debug( f"Fail to parse the argument: {sub_cmd['arguments']}", ) # Type error for the arguments if not isinstance(sub_cmd["arguments"], dict): raise FunctionCallFormatError( "Except a dictionary for the arguments, but got " f"{type(sub_cmd['arguments'])} instead.", ) # Leaving the type checking and required checking to the runtime # error reporting during execution return cmds def _execute_func(self, cmds: List[dict]) -> str: """Execute the function with the arguments. Args: cmds (`List[dict]`): A list of dictionaries, where each dictionary contains the name of the function and its arguments, e.g. {"name": "func1", "arguments": {"arg1": 1, "arg2": 2}}. Returns: `str`: The prompt of the execution results. """ execute_results = [] for i, cmd in enumerate(cmds): service_func = self.service_funcs[cmd["name"]] kwargs = cmd.get("arguments", {}) # Execute the function try: func_res = service_func.processed_func(**kwargs) except Exception as e: func_res = ServiceResponse( status=ServiceExecStatus.ERROR, content=str(e), ) status = ( "SUCCESS" if func_res.status == ServiceExecStatus.SUCCESS else "FAILED" ) arguments = [f"{k}: {v}" for k, v in kwargs.items()] execute_res = self._tools_execution_format.format_map( { "index": i + 1, "function_name": cmd["name"], "arguments": "\n\t\t".join(arguments), "status": status, "result": func_res.content, }, ) execute_results.append(execute_res) execute_results_prompt = "\n".join(execute_results) return execute_results_prompt
[文档] def parse_and_call_func( self, text_cmd: Union[list[dict], str], raise_exception: bool = False, ) -> Msg: """Parse, check the text and call the function.""" try: # --- Step 1: Parse the text according to the tools_call_format cmds = self._parse_and_check_text(text_cmd) # --- Step 2: Call the service function --- execute_results_prompt = self._execute_func(cmds) except FunctionCallError as e: # Catch the function calling error that can be handled by # the model if raise_exception: raise e from None execute_results_prompt = str(e) return Msg("system", execute_results_prompt, "system")
[文档] @classmethod def get( cls, service_func: Callable[..., Any], **kwargs: Any, ) -> Tuple[Callable[..., Any], dict]: """Convert a service function into a tool function that agent can use, and generate a dictionary in JSON Schema format that can be used in OpenAI API directly. While for open-source model, developers should handle the conversation from json dictionary to prompt. Args: service_func (`Callable[..., Any]`): The service function to be called. kwargs (`Any`): The arguments to be passed to the service function. Returns: `Tuple(Callable[..., Any], dict)`: A tuple of tool function and a dict in JSON Schema format to describe the function. Note: The description of the function and arguments are extracted from its docstring automatically, which should be well-formatted in **Google style**. Otherwise, their descriptions in the returned dictionary will be empty. Suggestions: 1. The name of the service function should be self-explanatory, so that the agent can understand the function and use it properly. 2. The typing of the arguments should be provided when defining the function (e.g. `def func(a: int, b: str, c: bool)`), so that the agent can specify the arguments properly. Example: .. code-block:: python def bing_search(query: str, api_key: str, num_results: int=10): '''Search the query in Bing search engine. Args: query (str): The string query to search. api_key (str): The API key for Bing search. num_results (int): The number of results to return, default to 10. ''' pass """ # Get the function for agent to use tool_func = partial(service_func, **kwargs) # Obtain all arguments of the service function argsspec = inspect.getfullargspec(service_func) # Construct the mapping from arguments to their typings if parse is None: raise ImportError( "Missing required package `docstring_parser`" "Please install it by " "`pip install docstring_parser`.", ) docstring = parse(service_func.__doc__) # Function description short_description = docstring.short_description or "" long_description = docstring.long_description or "" func_description = "\n\n".join([short_description, long_description]) # The arguments that requires the agent to specify # to support class method, the self args are deprecated args_agent = set(argsspec.args) - set(kwargs.keys()) - {"self", "cls"} # Check if the arguments from agent have descriptions in docstring args_description = { _.arg_name: _.description for _ in docstring.params } # Prepare default values if argsspec.defaults is None: args_defaults = {} else: args_defaults = dict( zip( reversed(argsspec.args), reversed(argsspec.defaults), # type: ignore ), ) args_required = sorted( list(set(args_agent) - set(args_defaults.keys())), ) # Prepare types of the arguments, remove the return type args_types = { k: v for k, v in argsspec.annotations.items() if k != "return" } # Prepare argument dictionary properties_field = {} for key in args_agent: arg_property = {} # type if key in args_types: try: required_type = _get_type_str(args_types[key]) arg_property["type"] = required_type except Exception: logger.warning( f"Fail and skip to get the type of the " f"argument `{key}`.", ) # For Literal type, add enum field if get_origin(args_types[key]) is Literal: arg_property["enum"] = list(args_types[key].__args__) # description if key in args_description: arg_property["description"] = args_description[key] # default if key in args_defaults and args_defaults[key] is not None: arg_property["default"] = args_defaults[key] properties_field[key] = arg_property # Construct the JSON Schema for the service function func_dict = { "type": "function", "function": { "name": service_func.__name__, "description": func_description.strip(), "parameters": { "type": "object", "properties": properties_field, "required": args_required, }, }, } return tool_func, func_dict
[文档] class ServiceFactory: """A service factory class that turns service function into string prompt format."""
[文档] @classmethod def get( cls, service_func: Callable[..., Any], **kwargs: Any, ) -> Tuple[Callable[..., Any], dict]: """Convert a service function into a tool function that agent can use, and generate a dictionary in JSON Schema format that can be used in OpenAI API directly. While for open-source model, developers should handle the conversation from json dictionary to prompt. Args: service_func (`Callable[..., Any]`): The service function to be called. kwargs (`Any`): The arguments to be passed to the service function. Returns: `Tuple(Callable[..., Any], dict)`: A tuple of tool function and a dict in JSON Schema format to describe the function. Note: The description of the function and arguments are extracted from its docstring automatically, which should be well-formatted in **Google style**. Otherwise, their descriptions in the returned dictionary will be empty. Suggestions: 1. The name of the service function should be self-explanatory, so that the agent can understand the function and use it properly. 2. The typing of the arguments should be provided when defining the function (e.g. `def func(a: int, b: str, c: bool)`), so that the agent can specify the arguments properly. Example: .. code-block:: python def bing_search(query: str, api_key: str, num_results: int=10): '''Search the query in Bing search engine. Args: query (str): The string query to search. api_key (str): The API key for Bing search. num_results (int): The number of results to return, default to 10. ''' pass """ logger.warning( "The service factory will be deprecated in the future." " Try to use the `ServiceToolkit` class instead.", ) # Get the function for agent to use tool_func = partial(service_func, **kwargs) # Obtain all arguments of the service function argsspec = inspect.getfullargspec(service_func) # Construct the mapping from arguments to their typings if parse is None: raise ImportError( "Missing required package `docstring_parser`" "Please install it by " "`pip install docstring_parser`.", ) docstring = parse(service_func.__doc__) # Function description short_description = docstring.short_description or "" long_description = docstring.long_description or "" func_description = "\n".join([short_description, long_description]) # The arguments that requires the agent to specify # we remove the self argument, for class methods args_agent = set(argsspec.args) - set(kwargs.keys()) - {"self", "cls"} # Check if the arguments from agent have descriptions in docstring args_description = { _.arg_name: _.description for _ in docstring.params } # Prepare default values if argsspec.defaults is None: args_defaults = {} else: args_defaults = dict( zip( reversed(argsspec.args), reversed(argsspec.defaults), # type: ignore ), ) args_required = sorted( list(set(args_agent) - set(args_defaults.keys())), ) # Prepare types of the arguments, remove the return type args_types = { k: v for k, v in argsspec.annotations.items() if k != "return" } # Prepare argument dictionary properties_field = {} for key in args_agent: arg_property = {} # type if key in args_types: try: required_type = _get_type_str(args_types[key]) arg_property["type"] = required_type except Exception: logger.warning( f"Fail and skip to get the type of the " f"argument `{key}`.", ) # For Literal type, add enum field if get_origin(args_types[key]) is Literal: arg_property["enum"] = list(args_types[key].__args__) # description if key in args_description: arg_property["description"] = args_description[key] # default if key in args_defaults and args_defaults[key] is not None: arg_property["default"] = args_defaults[key] properties_field[key] = arg_property # Construct the JSON Schema for the service function func_dict = { "type": "function", "function": { "name": service_func.__name__, "description": func_description, "parameters": { "type": "object", "properties": properties_field, "required": args_required, }, }, } return tool_func, func_dict