mirror of
https://github.com/FoundationAgents/MetaGPT.git
synced 2026-10-16 06:56:45 +08:00
352 lines
13 KiB
Python
352 lines
13 KiB
Python
#!/usr/bin/env python
|
|
# -*- coding: utf-8 -*-
|
|
"""
|
|
@Time : 2023/12/11 18:45
|
|
@Author : alexanderwu
|
|
@File : action_node.py
|
|
|
|
NOTE: You should use typing.List instead of list to do type annotation. Because in the markdown extraction process,
|
|
we can use typing to extract the type of the node, but we cannot use built-in list to extract.
|
|
"""
|
|
import json
|
|
from typing import Any, Dict, List, Optional, Tuple, Type
|
|
|
|
from pydantic import BaseModel, create_model, model_validator
|
|
from tenacity import retry, stop_after_attempt, wait_random_exponential
|
|
|
|
from metagpt.config import CONFIG
|
|
from metagpt.llm import BaseLLM
|
|
from metagpt.logs import logger
|
|
from metagpt.provider.postprocess.llm_output_postprocess import llm_output_postprocess
|
|
from metagpt.utils.common import OutputParser, general_after_log
|
|
|
|
TAG = "CONTENT"
|
|
|
|
LANGUAGE_CONSTRAINT = "Language: Please use the same language as Human INPUT."
|
|
FORMAT_CONSTRAINT = f"Format: output wrapped inside [{TAG}][/{TAG}] like format example, nothing else."
|
|
# Delete comments in json
|
|
# If you don't want JSONDecodeError to occur, you can add Delete comments in json after FORMAT_CONSTRAINT
|
|
|
|
|
|
SIMPLE_TEMPLATE = """
|
|
## context
|
|
{context}
|
|
|
|
-----
|
|
|
|
## format example
|
|
{example}
|
|
|
|
## nodes: "<node>: <type> # <instruction>"
|
|
{instruction}
|
|
|
|
## constraint
|
|
{constraint}
|
|
|
|
## action
|
|
Follow instructions of nodes, generate output and make sure it follows the format example.
|
|
"""
|
|
|
|
|
|
def dict_to_markdown(d, prefix="- ", kv_sep="\n", postfix="\n"):
|
|
markdown_str = ""
|
|
for key, value in d.items():
|
|
markdown_str += f"{prefix}{key}{kv_sep}{value}{postfix}"
|
|
return markdown_str
|
|
|
|
|
|
class ActionNode:
|
|
"""ActionNode is a tree of nodes."""
|
|
|
|
schema: str # raw/json/markdown, default: ""
|
|
|
|
# Action Context
|
|
context: str # all the context, including all necessary info
|
|
llm: BaseLLM # LLM with aask interface
|
|
children: dict[str, "ActionNode"]
|
|
|
|
# Action Input
|
|
key: str # Product Requirement / File list / Code
|
|
expected_type: Type # such as str / int / float etc.
|
|
# context: str # everything in the history.
|
|
instruction: str # the instructions should be followed.
|
|
example: Any # example for In Context-Learning.
|
|
|
|
# Action Output
|
|
content: str
|
|
instruct_content: BaseModel
|
|
|
|
def __init__(
|
|
self,
|
|
key: str,
|
|
expected_type: Type,
|
|
instruction: str,
|
|
example: Any,
|
|
content: str = "",
|
|
children: dict[str, "ActionNode"] = None,
|
|
schema: str = "",
|
|
):
|
|
self.key = key
|
|
self.expected_type = expected_type
|
|
self.instruction = instruction
|
|
self.example = example
|
|
self.content = content
|
|
self.children = children if children is not None else {}
|
|
self.schema = schema
|
|
|
|
def __str__(self):
|
|
return (
|
|
f"{self.key}, {repr(self.expected_type)}, {self.instruction}, {self.example}"
|
|
f", {self.content}, {self.children}"
|
|
)
|
|
|
|
def __repr__(self):
|
|
return self.__str__()
|
|
|
|
def add_child(self, node: "ActionNode"):
|
|
"""增加子ActionNode"""
|
|
self.children[node.key] = node
|
|
|
|
def add_children(self, nodes: List["ActionNode"]):
|
|
"""批量增加子ActionNode"""
|
|
for node in nodes:
|
|
self.add_child(node)
|
|
|
|
@classmethod
|
|
def from_children(cls, key, nodes: List["ActionNode"]):
|
|
"""直接从一系列的子nodes初始化"""
|
|
obj = cls(key, str, "", "")
|
|
obj.add_children(nodes)
|
|
return obj
|
|
|
|
def get_children_mapping(self, exclude=None) -> Dict[str, Tuple[Type, Any]]:
|
|
"""获得子ActionNode的字典，以key索引"""
|
|
exclude = exclude or []
|
|
return {k: (v.expected_type, ...) for k, v in self.children.items() if k not in exclude}
|
|
|
|
def get_self_mapping(self) -> Dict[str, Tuple[Type, Any]]:
|
|
"""get self key: type mapping"""
|
|
return {self.key: (self.expected_type, ...)}
|
|
|
|
def get_mapping(self, mode="children", exclude=None) -> Dict[str, Tuple[Type, Any]]:
|
|
"""get key: type mapping under mode"""
|
|
if mode == "children" or (mode == "auto" and self.children):
|
|
return self.get_children_mapping(exclude=exclude)
|
|
return {} if exclude and self.key in exclude else self.get_self_mapping()
|
|
|
|
@classmethod
|
|
def create_model_class(cls, class_name: str, mapping: Dict[str, Tuple[Type, Any]]):
|
|
"""基于pydantic v1的模型动态生成，用来检验结果类型正确性"""
|
|
|
|
def check_fields(cls, values):
|
|
required_fields = set(mapping.keys())
|
|
missing_fields = required_fields - set(values.keys())
|
|
if missing_fields:
|
|
raise ValueError(f"Missing fields: {missing_fields}")
|
|
|
|
unrecognized_fields = set(values.keys()) - required_fields
|
|
if unrecognized_fields:
|
|
logger.warning(f"Unrecognized fields: {unrecognized_fields}")
|
|
return values
|
|
|
|
validators = {"check_missing_fields_validator": model_validator(mode="before")(check_fields)}
|
|
|
|
new_class = create_model(class_name, __validators__=validators, **mapping)
|
|
return new_class
|
|
|
|
def create_children_class(self, exclude=None):
|
|
"""使用object内有的字段直接生成model_class"""
|
|
class_name = f"{self.key}_AN"
|
|
mapping = self.get_children_mapping(exclude=exclude)
|
|
return self.create_model_class(class_name, mapping)
|
|
|
|
def to_dict(self, format_func=None, mode="auto", exclude=None) -> Dict:
|
|
"""将当前节点与子节点都按照node: format的格式组织成字典"""
|
|
|
|
# 如果没有提供格式化函数，使用默认的格式化方式
|
|
if format_func is None:
|
|
format_func = lambda node: f"{node.instruction}"
|
|
|
|
# 使用提供的格式化函数来格式化当前节点的值
|
|
formatted_value = format_func(self)
|
|
|
|
# 创建当前节点的键值对
|
|
if mode == "children" or (mode == "auto" and self.children):
|
|
node_dict = {}
|
|
else:
|
|
node_dict = {self.key: formatted_value}
|
|
|
|
if mode == "root":
|
|
return node_dict
|
|
|
|
# 遍历子节点并递归调用 to_dict 方法
|
|
exclude = exclude or []
|
|
for _, child_node in self.children.items():
|
|
if child_node.key in exclude:
|
|
continue
|
|
node_dict.update(child_node.to_dict(format_func))
|
|
|
|
return node_dict
|
|
|
|
def compile_to(self, i: Dict, schema, kv_sep) -> str:
|
|
if schema == "json":
|
|
return json.dumps(i, indent=4)
|
|
elif schema == "markdown":
|
|
return dict_to_markdown(i, kv_sep=kv_sep)
|
|
else:
|
|
return str(i)
|
|
|
|
def tagging(self, text, schema, tag="") -> str:
|
|
if not tag:
|
|
return text
|
|
if schema == "json":
|
|
return f"[{tag}]\n" + text + f"\n[/{tag}]"
|
|
else: # markdown
|
|
return f"[{tag}]\n" + text + f"\n[/{tag}]"
|
|
|
|
def _compile_f(self, schema, mode, tag, format_func, kv_sep, exclude=None) -> str:
|
|
nodes = self.to_dict(format_func=format_func, mode=mode, exclude=exclude)
|
|
text = self.compile_to(nodes, schema, kv_sep)
|
|
return self.tagging(text, schema, tag)
|
|
|
|
def compile_instruction(self, schema="markdown", mode="children", tag="", exclude=None) -> str:
|
|
"""compile to raw/json/markdown template with all/root/children nodes"""
|
|
format_func = lambda i: f"{i.expected_type} # {i.instruction}"
|
|
return self._compile_f(schema, mode, tag, format_func, kv_sep=": ", exclude=exclude)
|
|
|
|
def compile_example(self, schema="json", mode="children", tag="", exclude=None) -> str:
|
|
"""compile to raw/json/markdown examples with all/root/children nodes"""
|
|
|
|
# 这里不能使用f-string，因为转译为str后再json.dumps会额外加上引号，无法作为有效的example
|
|
# 错误示例："File list": "['main.py', 'const.py', 'game.py']", 注意这里值不是list，而是str
|
|
format_func = lambda i: i.example
|
|
return self._compile_f(schema, mode, tag, format_func, kv_sep="\n", exclude=exclude)
|
|
|
|
def compile(self, context, schema="json", mode="children", template=SIMPLE_TEMPLATE, exclude=[]) -> str:
|
|
"""
|
|
mode: all/root/children
|
|
mode="children": 编译所有子节点为一个统一模板，包括instruction与example
|
|
mode="all": NotImplemented
|
|
mode="root": NotImplemented
|
|
schmea: raw/json/markdown
|
|
schema="raw": 不编译，context, lang_constaint, instruction
|
|
schema="json"：编译context, example(json), instruction(markdown), constraint, action
|
|
schema="markdown": 编译context, example(markdown), instruction(markdown), constraint, action
|
|
"""
|
|
if schema == "raw":
|
|
return context + "\n\n## Actions\n" + LANGUAGE_CONSTRAINT + "\n" + self.instruction
|
|
|
|
# FIXME: json instruction会带来格式问题，如："Project name": "web_2048 # 项目名称使用下划线",
|
|
# compile example暂时不支持markdown
|
|
instruction = self.compile_instruction(schema="markdown", mode=mode, exclude=exclude)
|
|
example = self.compile_example(schema=schema, tag=TAG, mode=mode, exclude=exclude)
|
|
# nodes = ", ".join(self.to_dict(mode=mode).keys())
|
|
constraints = [LANGUAGE_CONSTRAINT, FORMAT_CONSTRAINT]
|
|
constraint = "\n".join(constraints)
|
|
|
|
prompt = template.format(
|
|
context=context,
|
|
example=example,
|
|
instruction=instruction,
|
|
constraint=constraint,
|
|
)
|
|
return prompt
|
|
|
|
@retry(
|
|
wait=wait_random_exponential(min=1, max=20),
|
|
stop=stop_after_attempt(6),
|
|
after=general_after_log(logger),
|
|
)
|
|
async def _aask_v1(
|
|
self,
|
|
prompt: str,
|
|
output_class_name: str,
|
|
output_data_mapping: dict,
|
|
system_msgs: Optional[list[str]] = None,
|
|
schema="markdown", # compatible to original format
|
|
timeout=CONFIG.timeout,
|
|
) -> (str, BaseModel):
|
|
"""Use ActionOutput to wrap the output of aask"""
|
|
content = await self.llm.aask(prompt, system_msgs, timeout=timeout)
|
|
logger.debug(f"llm raw output:\n{content}")
|
|
output_class = self.create_model_class(output_class_name, output_data_mapping)
|
|
|
|
if schema == "json":
|
|
parsed_data = llm_output_postprocess(
|
|
output=content, schema=output_class.model_json_schema(), req_key=f"[/{TAG}]"
|
|
)
|
|
else: # using markdown parser
|
|
parsed_data = OutputParser.parse_data_with_mapping(content, output_data_mapping)
|
|
|
|
logger.debug(f"parsed_data:\n{parsed_data}")
|
|
instruct_content = output_class(**parsed_data)
|
|
return content, instruct_content
|
|
|
|
def get(self, key):
|
|
return self.instruct_content.model_dump()[key]
|
|
|
|
def set_recursive(self, name, value):
|
|
setattr(self, name, value)
|
|
for _, i in self.children.items():
|
|
i.set_recursive(name, value)
|
|
|
|
def set_llm(self, llm):
|
|
self.set_recursive("llm", llm)
|
|
|
|
def set_context(self, context):
|
|
self.set_recursive("context", context)
|
|
|
|
async def simple_fill(self, schema, mode, timeout=CONFIG.timeout, exclude=None):
|
|
prompt = self.compile(context=self.context, schema=schema, mode=mode, exclude=exclude)
|
|
|
|
if schema != "raw":
|
|
mapping = self.get_mapping(mode, exclude=exclude)
|
|
class_name = f"{self.key}_AN"
|
|
content, scontent = await self._aask_v1(prompt, class_name, mapping, schema=schema, timeout=timeout)
|
|
self.content = content
|
|
self.instruct_content = scontent
|
|
else:
|
|
self.content = await self.llm.aask(prompt)
|
|
self.instruct_content = None
|
|
|
|
return self
|
|
|
|
async def fill(self, context, llm, schema="json", mode="auto", strgy="simple", timeout=CONFIG.timeout, exclude=[]):
|
|
"""Fill the node(s) with mode.
|
|
|
|
:param context: Everything we should know when filling node.
|
|
:param llm: Large Language Model with pre-defined system message.
|
|
:param schema: json/markdown, determine example and output format.
|
|
- raw: free form text
|
|
- json: it's easy to open source LLM with json format
|
|
- markdown: when generating code, markdown is always better
|
|
:param mode: auto/children/root
|
|
- auto: automated fill children's nodes and gather outputs, if no children, fill itself
|
|
- children: fill children's nodes and gather outputs
|
|
- root: fill root's node and gather output
|
|
:param strgy: simple/complex
|
|
- simple: run only once
|
|
- complex: run each node
|
|
:param timeout: Timeout for llm invocation.
|
|
:param exclude: The keys of ActionNode to exclude.
|
|
:return: self
|
|
"""
|
|
self.set_llm(llm)
|
|
self.set_context(context)
|
|
if self.schema:
|
|
schema = self.schema
|
|
|
|
if strgy == "simple":
|
|
return await self.simple_fill(schema=schema, mode=mode, timeout=timeout, exclude=exclude)
|
|
elif strgy == "complex":
|
|
# 这里隐式假设了拥有children
|
|
tmp = {}
|
|
for _, i in self.children.items():
|
|
if exclude and i.key in exclude:
|
|
continue
|
|
child = await i.simple_fill(schema=schema, mode=mode, timeout=timeout, exclude=exclude)
|
|
tmp.update(child.instruct_content.dict())
|
|
cls = self.create_children_class()
|
|
self.instruct_content = cls(**tmp)
|
|
return self
|