gyt/opendan
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #110 from photosssa/mvp-dev

Browse Source 
knowledge pipeline and to learn list
This commit is contained in:
Liu Zhicong
2023-12-06 00:05:50 -08:00
committed by GitHub
parent 0425c80ffd 33a079f897
commit 7e3949c155
36 changed files with 1670 additions and 1464 deletions
Download Patch File Download Diff File Expand all files Collapse all files
doc/learn_knowledge.md
+30
View File
@@ -0,0 +1,30 @@
# learn todo list
在workspace中分离出独立的两个work list；用于处理工作的work todo list， 和用于更新knowledge的 learn todo list；两种todo list在agent templete中单独的配置prompt，每个todo list都可以配置 do , check 和 review 三个prompt。
learn todo list中的todo，可以在knowledge pipeline 生成。
一个典型的例子是新的JarvisPlus agent，它可以依照自己的理解将某个本地目录中的文档归类到不同的逻辑目录中：
+ 首先定义他的pipeline：扫描目录中没有读过的文档，传递给parser；
+ 定义他的pipeline parser的实现：向自己所处的workspace 的learn todo list中添加一条todo， todo的内容是文档的全文；
+ 定义他的 learn prompts：根据输入的文档内容，输出摘要，和归类后的目录，产生一组归类operation；
+ workspace中嵌入了支持归类operation 的knowledage base environment，可以执行learn todo list产生的归类operation；
# environments and workspace
agent在独立的workspace中工作，目前默认的workspace是agent自己；除了自己的workspace，agent也可以进入其他的workspace；environment表示agent可以调用的function，和可以产生的operation；
environment中的function或者operation输出的结果，应当应用在agent所处的workspace中，所以agent在不同workspace中执行work，结果应当是被隔离的。
# Learn Todo List
Separate two independent work lists in the workspace: a work todo list for handling work, and a learn todo list for updating knowledge. Both types of todo lists have their own configured prompts in the agent template, and each todo list can configure do, check, and review prompts.
The todos in the learn todo list can be generated in the knowledge pipeline.
A typical example is the new JarvisPlus agent, which can categorize documents from a local directory into different logical directories according to its understanding:
+ First, define its pipeline: scan documents in the directory that have not been read and pass them to the parser;
+ Define the implementation of its pipeline parser: add a todo to the learn todo list of the workspace it is in, the content of the todo is the full text of the document;
+ Define its learn prompts: based on the input document content, output a summary and the directory after categorization, generating a set of categorization operations;
+ The workspace embeds a knowledge base environment that supports categorization operations, which can execute the categorization operations generated by the learn todo list.
# Environments and Workspace
The agent works in an independent workspace, and the current default workspace is the agent itself. In addition to its own workspace, the agent can also enter other workspaces. The environment represents the functions that the agent can call and the operations it can generate.
The results of the functions or operations in the environment should be applied in the workspace where the agent is located, so the results of the agent's work in different workspaces should be isolated.
Please note that the translation might not be perfect due to the technical nature of the text and potential ambiguity in the original text.
package-lock.json
Generated
+3
View File
@@ -0,0 +1,3 @@
{
"lockfileVersion": 1
}
rootfs/agents/JarvisPlus/agent.toml
+47 -3
View File
@@ -1,12 +1,12 @@
instance_id = "JarvisPlus"
fullname = "JarvisPlus"
llm_model_name = "gpt-4-1106-preview"
max_token_size = 4000
#enable_kb = "true"
enable_timestamp = "true"
owner_prompt = "I am your master {name} , now is {now}"
contact_prompt = "I am your master's friend {name}"
[[do_prompt]]
owner_env = ["knowledge"]
[[work.do]]
role = "system"
content = """
My name is JarvisPlus, I am the master's super personal assistant. I think hard and try my best to complete TODOs.
@@ -58,6 +58,50 @@ The result of my planned execution must be directly parsed by `python json.loads
"""
[[learn.do]]
role = "system"
content = """
我是一名软件工程师，拥有非常优秀的资料学习能力。下面是我学习和整理资料的方法
1. 由于LLM的Token限制，我学习的可能只是资料的部分内容，此时我应能产生合适的学习中间结果，中间结果保存在metadata中。我要么产生中间结果，要么产生最终结果。
2. 当存在已知信息时，需参考已知信息的内容来思考结果。
3. 当我收到最后一部分内容时，我能结合已知的中间结果产生最终结果。
4. 现有资料库以文件系统的形式组织，我未来借助资料的摘要来浏览知识库
5. 我将学习过的资料另存在资料库的合适位置（以/开始的完整路径）。保存位置的目录深度不超过5层，文件夹名称长度不超过16个字符。
6. 总是以json格式返回思考结果，json格式如下
{
"op_list":[
{
"op":"learn",
"original_path":"$original_path",
"think":"$think_result",
"metadata":{...},
"tags":["tag1","tag2"...],
"path":["/graphic/opengl","/database/mysql"], # list of directories to save to.
"title":"$article_title",
"summary":"$summary",
"catalogs": [
{
# optional,catalogs is a tree
"title":"$catalog_name1",
"pos":"$pos:$length"
"children":[
{
"title":"$catalog_name 1.1",
"pos":"$pos:$length"
},
{
"title":"$catalog_name2",
"pos":"$pos:$length"
}
]
},
]
}
]
}
"""
[[prompt]]
role = "system"
content = """
rootfs/knowledge_pipelines/JarvisPlus/pipeline.toml
+8
View File
@@ -0,0 +1,8 @@
name = "JarvisPlus"
input.module = "scan_local"
input.params.workspace = "${myai_dir}/workspace/JarvisPlus"
input.params.path = "${myai_dir}/data"
parser.module = "parse_local"
parser.params.workspace = "${myai_dir}/workspace/JarvisPlus"
parser.params.assign_to = "JarvisPlus"
rootfs/knowledge_pipelines/Mail/Spider/pipeline.toml
-4
View File
@@ -1,4 +0,0 @@
name = "Mail.Issue"
input.module = "input.py"
input.params.path = "${myai_dir}/data"
rootfs/knowledge_pipelines/Mail/Spider/input.py → rootfs/knowledge_pipelines/Mail/Sync/input.py
View File
rootfs/knowledge_pipelines/Mail/Sync/pipeline.toml
+14
View File
@@ -0,0 +1,14 @@
name = "Mail.Sync"
input.module = "input.py"
[input.params]
path = "${myai_dir}/mail"
imap_server = "imap.qq.com"
imap_port = 993
address = "115620204@qq.com"
password = "zbbjpbukeonqbjja"
[input.params.fields]
from = "from"
to = "to"
subject = "subject"
rootfs/knowledge_pipelines/Mia/input.py
-1
View File
@@ -6,7 +6,6 @@ import aiofiles
import chardet
import logging
import string
import docx2txt
from PyPDF2 import PdfReader
rootfs/knowledge_pipelines/Mia/parser.py
+2 -3
View File
@@ -1,8 +1,7 @@
# define a knowledge base class
import json
import string
from aios_kernel import ComputeKernel, AIStorage
from knowledge import *
from aios import *
class EmbeddingParser:
@@ -96,7 +95,7 @@ class EmbeddingParser:
async def parse(self, object: ObjectID) -> str:
obj = self.env.get_knowledge_store().load_object(object)
await self.__do_embedding(obj)
return "insert into vector store"
return str(object)
def init(env: KnowledgePipelineEnvironment, params: dict) -> EmbeddingParser:
return EmbeddingParser(env, params)
rootfs/knowledge_pipelines/Mia/query.py
+6 -7
View File
@@ -1,12 +1,11 @@
import os
import logging
import json
from aios_kernel import *
from knowledge import *
from aios import *
class KnowledgeEnvironment(Environment):
def __init__(self, env_id: str) -> None:
super().__init__(env_id)
class EmbeddingEnvironment(SimpleEnvironment):
def __init__(self, workspace: str) -> None:
super().__init__(workspace)
self.path = os.path.join(AIStorage.get_instance().get_myai_dir(), "knowledge/indices/embedding")
self._default_text_model = "all-MiniLM-L6-v2"
self._default_image_model = "clip-ViT-B-32"
@@ -93,5 +92,5 @@ class KnowledgeEnvironment(Environment):
content = "*** I have provided the following known information for your reference with json format:\n"
return content + self.tokens_from_objects(object_ids[index:index+1])
def init() -> KnowledgeEnvironment:
return KnowledgeEnvironment("embedding")
def init(workspace: str) -> EmbeddingEnvironment:
return EmbeddingEnvironment(workspace)
rootfs/knowledge_pipelines/pipelines.toml
+1 -1
View File
@@ -1,3 +1,3 @@
pipelines = [
"Mail/Issue"
"JarvisPlus"
]
src/aios/__init__.py
+6 -6
View File
@@ -2,12 +2,12 @@
from .proto.agent_msg import *
from .proto.compute_task import *
from .agent.agent_base import AgentPrompt,CustomAIAgent
from .agent.agent_base import AgentPrompt,CustomAIAgent, AgentTodo
from .agent.chatsession import AIChatSession
from .agent.agent import AIAgent,AIAgentTemplete, BaseAIAgent
from .agent.role import AIRole,AIRoleGroup
from .agent.workflow import Workflow
from .agent.ai_function import SimpleAIFunction
# from .agent.workflow import Workflow
from .agent.ai_function import SimpleAIFunction, SimpleAIOperation
from .frame.compute_kernel import ComputeKernel,ComputeTask,ComputeTaskResult,ComputeTaskState,ComputeTaskType
from .frame.compute_node import ComputeNode,LocalComputeNode
@@ -16,11 +16,11 @@ from .frame.tunnel import AgentTunnel
from .frame.contact_manager import ContactManager,Contact,FamilyMember
from .frame.queue_compute_node import Queue_ComputeNode
from .environment.environment import Environment,EnvironmentEvent
from .environment.workflow_env import WorkflowEnvironment,CalenderEnvironment,CalenderEvent,PaintEnvironment
from .environment.environment import BaseEnvironment,SimpleEnvironment,CompositeEnvironment
# from .environment.workflow_env import WorkflowEnvironment,CalenderEnvironment,CalenderEvent,PaintEnvironment
from .environment.text_to_speech_function import TextToSpeechFunction
from .environment.image_2_text_function import Image2TextFunction
from .environment.workspace_env import ShellEnvironment,WorkspaceEnvironment
from .environment.workspace_env import WorkspaceEnvironment,TodoListEnvironment,TodoListType
from .storage.storage import ResourceLocation,AIStorage,UserConfig,UserConfigItem
src/aios/agent/agent.py
+201 -541
View File
@@ -17,6 +17,7 @@ from ..proto.agent_msg import AgentMsg
from .agent_base import *
from .chatsession import *
from .ai_function import *
from ..environment.workspace_env import WorkspaceEnvironment, TodoListType
from ..frame.contact_manager import ContactManager,Contact,FamilyMember
from ..frame.compute_kernel import ComputeKernel
@@ -32,67 +33,35 @@ from ..proto.compute_task import ComputeTaskResult,ComputeTaskResultCode
logger = logging.getLogger(__name__)
DEFAULT_AGENT_READ_REPORT_PROMPT = """
"""
# DEFAULT_AGENT_READ_REPORT_PROMPT = """
# """
DEFAULT_AGENT_DO_PROMPT = """
You are a helpful AI assistant.
Solve tasks using your coding and language skills.
In the following cases, suggest python code (in a python coding block) for the user to execute.
1. When you need to collect info, use the code to output the info you need, for example, browse or search the web, download/read a file, print the content of a webpage or a file, get the current date/time, check the operating system. After sufficient info is printed and the task is ready to be solved based on your language skill, you can solve the task by yourself.
2. When you need to perform some task with code, use the code to perform the task and output the result. Finish the task smartly.
Solve the task step by step if you need to. If a plan is not provided, explain your plan first. Be clear which step uses code, and which step uses your language skill.
When using code, you must indicate the script type in the code block. The user cannot provide any other feedback or perform any other action beyond executing the code you suggest. The user can't modify your code. So do not suggest incomplete code which requires users to modify. Don't use a code block if it's not intended to be executed by the user.
If you want the user to save the code in a file before executing it, put # filename: <filename> inside the code block as the first line. Don't include multiple code blocks in one response. Do not ask users to copy and paste the result. Instead, use 'print' function for the output when relevant. Check the execution result returned by the user.
If the result indicates there is an error, fix the error and output the code again. Suggest the full code instead of partial code or code changes. If the error can't be fixed or if the task is not solved even after the code is executed successfully, analyze the problem, revisit your assumption, collect additional info you need, and think of a different approach to try.
When you find an answer, verify the answer carefully. Include verifiable evidence in your response if possible.
Reply "TERMINATE" in the end when everything is done.
"""
# DEFAULT_AGENT_DO_PROMPT = """
# You are a helpful AI assistant.
# Solve tasks using your coding and language skills.
# In the following cases, suggest python code (in a python coding block) for the user to execute.
# 1. When you need to collect info, use the code to output the info you need, for example, browse or search the web, download/read a file, print the content of a webpage or a file, get the current date/time, check the operating system. After sufficient info is printed and the task is ready to be solved based on your language skill, you can solve the task by yourself.
# 2. When you need to perform some task with code, use the code to perform the task and output the result. Finish the task smartly.
# Solve the task step by step if you need to. If a plan is not provided, explain your plan first. Be clear which step uses code, and which step uses your language skill.
# When using code, you must indicate the script type in the code block. The user cannot provide any other feedback or perform any other action beyond executing the code you suggest. The user can't modify your code. So do not suggest incomplete code which requires users to modify. Don't use a code block if it's not intended to be executed by the user.
# If you want the user to save the code in a file before executing it, put # filename: <filename> inside the code block as the first line. Don't include multiple code blocks in one response. Do not ask users to copy and paste the result. Instead, use 'print' function for the output when relevant. Check the execution result returned by the user.
# If the result indicates there is an error, fix the error and output the code again. Suggest the full code instead of partial code or code changes. If the error can't be fixed or if the task is not solved even after the code is executed successfully, analyze the problem, revisit your assumption, collect additional info you need, and think of a different approach to try.
# When you find an answer, verify the answer carefully. Include verifiable evidence in your response if possible.
# Reply "TERMINATE" in the end when everything is done.
# """
DEFAULT_AGENT_SELF_CHECK_PROMPT = """
# DEFAULT_AGENT_SELF_CHECK_PROMPT = """
"""
# """
DEFAULT_AGENT_GOAL_TO_TODO_PROMPT = """
我会给你一个目标，你需要结合自己的角色思考如何将其拆解成多个TODO。请直接返回json来表达这些TODO
"""
# DEFAULT_AGENT_GOAL_TO_TODO_PROMPT = """
# 我会给你一个目标，你需要结合自己的角色思考如何将其拆解成多个TODO。请直接返回json来表达这些TODO
# """
DEFAULT_AGENT_LEARN_PROMPT = """
我是一名软件工程师，拥有非常优秀的资料学习能力。下面是我学习和整理资料的方法
1. 由于LLM的Token限制，我学习的可能只是资料的部分内容，此时我应能产生合适的学习中间结果，中间结果保存在metadata中。我要么产生中间结果，要么产生最终结果。
2. 当存在已知信息时，需参考已知信息的内容来思考结果。
3. 当我收到最后一部分内容时，我能结合已知的中间结果产生最终结果。
4. 现有资料库以文件系统的形式组织，我未来借助资料的摘要来浏览知识库
5. 我将学习过的资料另存在资料库的合适位置（以/开始的完整路径）。保存位置的目录深度不超过5层，文件夹名称长度不超过16个字符。
6. 总是以json格式返回思考结果，json格式如下
{
think:"$think_result",
metadata:{...} , # temp result for long content
tags:["tag1","tag2"...],
path:["/graphic/opengl","/database/mysql"], # list of directories to save to.
title:"$article_title",
summary:"$summary",
catalogs: [{ # optional,catalogs is a tree
title:"$catalog_name1",
pos:"$pos:$length"
children:[
{
title:"$catalog_name 1.1",
pos:"$pos:$length"
}
]},
{
title:"$catalog_name2",
pos:"$pos:$length"
}
]
}
"""
DEFAULT_AGENT_LEARN_LONG_CONENT_PROMPT = """
我给你一段内容，尝试为期建立目录。目录的标题不能超过16个字，
目录要指向正文的位置（用字符偏移即可），整个目录的文本长度不能超过256个字节。并用json表达这个目录
"""
# DEFAULT_AGENT_LEARN_LONG_CONENT_PROMPT = """
# 我给你一段内容，尝试为期建立目录。目录的标题不能超过16个字，
# 目录要指向正文的位置（用字符偏移即可），整个目录的文本长度不能超过256个字节。并用json表达这个目录
# """
class AIAgentTemplete:
def __init__(self) -> None:
self.llm_model_name:str = "gpt-4-0613"
@@ -144,45 +113,32 @@ class AIAgent(BaseAIAgent):
self.owner_promp_str = None
self.contact_prompt_str = None
self.history_len = 10
self.review_todo_prompt = None
self.read_report_prompt = None
self.do_prompt = None
self.check_prompt = None
self.goal_to_todo_prompt = None
self.learn_token_limit = 4000
self.learn_prompt = AgentPrompt(DEFAULT_AGENT_LEARN_PROMPT)
todo_prompts = {}
todo_prompts[TodoListType.TO_WORK] = {
"do": None,
"check": None,
"review": None,
}
todo_prompts[TodoListType.TO_LEARN] = {
"do": None,
"check": None,
"review": None,
}
self.todo_prompts = todo_prompts
self.chat_db = None
self.unread_msg = Queue() # msg from other agent
self.owner_env : Environment = None
self.owenr_bus = None
self.enable_function_list = None
@classmethod
def create_from_templete(cls,templete:AIAgentTemplete, fullname:str):
# Agent just inherit from templete on craete,if template changed,agent will not change
result_agent = AIAgent()
result_agent.llm_model_name = templete.llm_model_name
result_agent.max_token_size = templete.max_token_size
result_agent.template_id = templete.template_id
result_agent.agent_id = "agent#" + uuid.uuid4().hex
result_agent.fullname = fullname
result_agent.powerby = templete.author
result_agent.agent_prompt = templete.prompt
return result_agent
def load_from_config(self,config:dict) -> bool:
if config.get("instance_id") is None:
logger.error("agent instance_id is None!")
return False
self.agent_id = config["instance_id"]
self.agent_workspace = WorkspaceEnvironment(self.agent_id)
self.agent_workspace = config["workspace"]
if config.get("fullname") is None:
logger.error(f"agent {self.agent_id} fullname is None!")
@@ -200,10 +156,24 @@ class AIAgent(BaseAIAgent):
self.agent_think_prompt = AgentPrompt()
self.agent_think_prompt.load_from_config(config["think_prompt"])
if config.get("do_prompt") is not None:
self.do_prompt = AgentPrompt()
self.do_prompt.load_from_config(config["do_prompt"])
self.wake_up()
def load_todo_config(todo_type:str) -> bool:
todo_config = config.get(todo_type)
if todo_config is not None:
if todo_config.get("do") is not None:
prompt = AgentPrompt()
prompt.load_from_config(todo_config["do"])
self.todo_prompts[todo_type]["do"] = prompt
if todo_config.get("check") is not None:
prompt = AgentPrompt()
prompt.load_from_config(todo_config["check"])
self.todo_prompts[todo_type]["check"] = prompt
if todo_config.get("review_prompt") is not None:
prompt = AgentPrompt()
prompt.load_from_config(todo_config["review_prompt"])
self.todo_prompts[todo_type]["review"] = prompt
load_todo_config(TodoListType.TO_WORK)
load_todo_config(TodoListType.TO_LEARN)
if config.get("guest_prompt") is not None:
self.guest_prompt_str = config["guest_prompt"]
@@ -214,9 +184,6 @@ class AIAgent(BaseAIAgent):
if config.get("contact_prompt") is not None:
self.contact_prompt_str = config["contact_prompt"]
if config.get("owner_env") is not None:
self.owner_env = config.get("owner_env")
if config.get("powerby") is not None:
self.powerby = config["powerby"]
@@ -234,6 +201,9 @@ class AIAgent(BaseAIAgent):
self.enable_timestamp = bool(config["enable_timestamp"])
if config.get("history_len"):
self.history_len = int(config.get("history_len"))
self.wake_up()
return True
def get_id(self) -> str:
@@ -254,16 +224,9 @@ class AIAgent(BaseAIAgent):
def get_max_token_size(self) -> int:
return self.max_token_size
def get_llm_learn_token_limit(self) -> int:
return self.learn_token_limit
def get_learn_prompt(self) -> AgentPrompt:
return self.learn_prompt
def get_agent_role_prompt(self) -> AgentPrompt:
return self.role_prompt
def _get_remote_user_prompt(self,remote_user:str) -> AgentPrompt:
cm = ContactManager.get_instance()
contact = cm.find_contact_by_name(remote_user)
@@ -290,34 +253,6 @@ class AIAgent(BaseAIAgent):
return None
def _get_inner_functions(self) -> dict:
if self.owner_env is None:
return None,0
all_inner_function = self.owner_env.get_all_ai_functions()
if all_inner_function is None:
return None,0
result_func = []
result_len = 0
for inner_func in all_inner_function:
func_name = inner_func.get_name()
if self.enable_function_list is not None:
if len(self.enable_function_list) > 0:
if func_name not in self.enable_function_list:
logger.debug(f"ageint {self.agent_id} ignore inner func:{func_name}")
continue
this_func = {}
this_func["name"] = func_name
this_func["description"] = inner_func.get_description()
this_func["parameters"] = inner_func.get_parameters()
result_len += len(json.dumps(this_func)) / 4
result_func.append(this_func)
return result_func,result_len
def get_agent_prompt(self) -> AgentPrompt:
return self.agent_prompt
@@ -325,96 +260,14 @@ class AIAgent(BaseAIAgent):
return self.agent_think_prompt
def _format_msg_by_env_value(self,prompt:AgentPrompt):
if self.owner_env is None:
return
for msg in prompt.messages:
old_content = msg.get("content")
msg["content"] = old_content.format_map(self.owner_env)
msg["content"] = old_content.format_map(self.agent_workspace)
async def _handle_event(self,event):
if event.type == "AgentThink":
return await self.do_self_think()
# async def _process_group_chat_msg(self,msg:AgentMsg) -> AgentMsg:
# session_topic = msg.target + "#" + msg.topic
# chatsession = AIChatSession.get_session(self.agent_id,session_topic,self.chat_db)
# workspace = self.get_current_workspace()
# need_process = False
# if msg.mentions is not None:
# if self.agent_id in msg.mentions:
# need_process = True
# logger.info(f"agent {self.agent_id} recv a group chat message from {msg.sender},but is not mentioned,ignore!")
# if need_process is not True:
# chatsession.append(msg)
# resp_msg = msg.create_group_resp_msg(self.agent_id,"")
# return resp_msg
# else:
# msg_prompt = AgentPrompt()
# msg_prompt.messages = [{"role":"user","content":f"{msg.sender}:{msg.body}"}]
# prompt = AgentPrompt()
# prompt.append(self.get_agent_prompt())
# if workspace:
# prompt.append(workspace.get_prompt())
# prompt.append(workspace.get_role_prompt(self.agent_id))
# if self.need_session_summmary(msg,chatsession):
# # get relate session(todos) summary
# summary = self.llm_select_session_summary(msg,chatsession)
# prompt.append(AgentPrompt(summary))
# self._format_msg_by_env_value(prompt)
# inner_functions,function_token_len = self._get_inner_functions()
# system_prompt_len = prompt.get_prompt_token_len()
# input_len = len(msg.body)
# history_prmpt,history_token_len = await self._get_prompt_from_session_for_groupchat(chatsession,system_prompt_len + function_token_len,input_len)
# prompt.append(history_prmpt) # chat context
# prompt.append(msg_prompt)
# logger.debug(f"Agent {self.agent_id} do llm token static system:{system_prompt_len},function:{function_token_len},history:{history_token_len},input:{input_len}, totoal prompt:{system_prompt_len + function_token_len + history_token_len} ")
# task_result = await self._do_llm_complection(prompt,inner_functions,msg)
# if task_result.result_code != ComputeTaskResultCode.OK:
# error_resp = msg.create_error_resp(task_result.error_str)
# return error_resp
# final_result = task_result.result_str
# llm_result : LLMResult = LLMResult.from_str(final_result)
# is_ignore = False
# result_prompt_str = ""
# match llm_result.state:
# case "ignore":
# is_ignore = True
# case "waiting":
# for sendmsg in llm_result.send_msgs:
# target = sendmsg.target
# sendmsg.sender = self.agent_id
# sendmsg.topic = msg.topic
# sendmsg.prev_msg_id = msg.get_msg_id()
# send_resp = await AIBus.get_default_bus().send_message(sendmsg)
# if send_resp is not None:
# result_prompt_str += f"\n{target} response is :{send_resp.body}"
# agent_sesion = AIChatSession.get_session(self.agent_id,f"{sendmsg.target}#{sendmsg.topic}",self.chat_db)
# agent_sesion.append(sendmsg)
# agent_sesion.append(send_resp)
# final_result = llm_result.resp + result_prompt_str
# if is_ignore is not True:
# resp_msg = msg.create_group_resp_msg(self.agent_id,final_result)
# chatsession.append(msg)
# chatsession.append(resp_msg)
# return resp_msg
# return None
def get_workspace_by_msg(self,msg:AgentMsg) -> WorkspaceEnvironment:
return self.agent_workspace
@@ -541,12 +394,12 @@ class AIAgent(BaseAIAgent):
known_info_str = "# Known information\n"
have_known_info = False
todos_str,todo_count = await workspace.get_todo_tree()
todos_str,todo_count = await workspace.todo_list[TodoListType.TO_WORK].get_todo_tree()
if todo_count > 0:
have_known_info = True
known_info_str += f"## todo\n{todos_str}\n"
inner_functions,function_token_len = BaseAIAgent.get_inner_functions(self.owner_env)
system_prompt_len = prompt.get_prompt_token_len()
inner_functions,function_token_len = BaseAIAgent.get_inner_functions(self.agent_workspace)
system_prompt_len = self.token_len(prompt=prompt)
input_len = len(msg.body)
if msg.msg_type == AgentMsgType.TYPE_GROUPMSG:
history_str,history_token_len = await self._get_prompt_from_session_for_groupchat(chatsession,system_prompt_len + function_token_len,input_len)
@@ -564,7 +417,7 @@ class AIAgent(BaseAIAgent):
logger.debug(f"Agent {self.agent_id} do llm token static system:{system_prompt_len},function:{function_token_len},history:{history_token_len},input:{input_len}, totoal prompt:{system_prompt_len + function_token_len + history_token_len} ")
task_result = await self.do_llm_complection(prompt,msg, env=self.owner_env,inner_functions=inner_functions)
task_result = await self.do_llm_complection(prompt,msg, inner_functions=inner_functions)
if task_result.result_code != ComputeTaskResultCode.OK:
error_resp = msg.create_error_resp(task_result.error_str)
return error_resp
@@ -618,9 +471,7 @@ class AIAgent(BaseAIAgent):
return None
async def _get_history_prompt_for_think(self,chatsession:AIChatSession,summary:str,system_token_len:int,pos:int)->(AgentPrompt,int):
history_len = (self.max_token_size * 0.7) - system_token_len
messages = chatsession.read_history(self.history_len,pos,"natural") # read
@@ -734,9 +585,7 @@ class AIAgent(BaseAIAgent):
worksapce.set_work_summary(self.agent_id,task_result.result_str)
# 尝试完成自己的TOOD （不依赖任何其他Agnet）
async def do_my_work(self) -> None:
async def _llm_run_todo_list(self, todo_list_type: TodoListType):
workspace : WorkspaceEnvironment = self.get_workspace_by_msg(None)
logger.info(f"agent {self.agent_id} do my work start!")
@@ -744,35 +593,26 @@ class AIAgent(BaseAIAgent):
#if await self.need_review_todolist():
# await self._llm_review_todolist(workspace)
todo_list = await workspace.get_todo_list(self.agent_id)
todo_list = workspace.todo_list[todo_list_type]
need_todo = await todo_list.get_todo_list(self.agent_id)
check_count = 0
do_count = 0
review_count = 0
for todo in todo_list:
for todo in need_todo:
if self.agent_energy <= 0:
break
if await self.need_review_todo(todo,workspace):
review_result = await self._llm_review_todo(todo,workspace)
todo.last_review_time = datetime.datetime.now().timestamp()
do_prompts = self._can_do_todo(todo_list_type, todo)
if do_prompts:
prompt : AgentPrompt = AgentPrompt()
prompt.append(self.agent_prompt)
prompt.append(workspace.get_role_prompt(self.agent_id))
prompt.append(do_prompts)
prompt.append(todo.to_prompt())
elif await self.can_check(todo,workspace):
check_result : AgentTodoResult = await self._llm_check_todo(todo,workspace)
todo.last_check_time = datetime.datetime.now().timestamp()
match check_result.result_code:
case AgentTodoResult.TODO_RESULT_CODE_LLM_ERROR:
continue
case AgentTodoResult.TODO_RESULT_CODE_OK:
await workspace.update_todo(todo.todo_id,AgentTodo.TODO_STATE_DONE)
case AgentTodoResult.TODO_RESULT_CODE_EXEC_OP_ERROR:
await workspace.update_todo(todo.todo_id,AgentTodo.TDDO_STATE_CHECKFAILED)
await workspace.append_worklog(todo,check_result)
self.agent_energy -= 1
check_count += 1
elif await self.can_do(todo,workspace):
do_result : AgentTodoResult = await self._llm_do(todo,workspace)
do_result : AgentTodoResult = await self._llm_do_todo(todo, prompt, workspace)
todo.last_do_time = datetime.datetime.now().timestamp()
todo.retry_count += 1
@@ -780,99 +620,129 @@ class AIAgent(BaseAIAgent):
case AgentTodoResult.TODO_RESULT_CODE_LLM_ERROR:
continue
case AgentTodoResult.TODO_RESULT_CODE_OK:
await workspace.update_todo(todo.todo_id,AgentTodo.TODO_STATE_WAITING_CHECK)
todo.result = do_result
await todo_list.update_todo(todo.todo_id,AgentTodo.TODO_STATE_WAITING_CHECK)
case AgentTodoResult.TODO_RESULT_CODE_EXEC_OP_ERROR:
await workspace.update_todo(todo.todo_id,AgentTodo.TODO_STATE_EXEC_FAILED)
await todo_list.update_todo(todo.todo_id,AgentTodo.TODO_STATE_EXEC_FAILED)
await workspace.append_worklog(todo,do_result)
await todo_list.append_worklog(todo,do_result)
self.agent_energy -= 2
do_count += 1
logger.info(f"agent {self.agent_id} ,check:{check_count} todo,do:{do_count} todo.")
# review_result = await self._llm_review_todo(todo,workspace)
# todo.last_review_time = datetime.datetime.now().timestamp()
continue
def get_review_todo_prompt(self,todo:AgentTodo) -> AgentPrompt:
return self.review_todo_prompt
check_prompts = self._can_check_todo(todo_list_type, todo)
if check_prompts:
prompt : AgentPrompt = AgentPrompt()
prompt.append(self.agent_prompt)
prompt.append(workspace.get_role_prompt(self.agent_id))
prompt.append(check_prompts)
async def _llm_review_todo(self,todo:AgentTodo,workspace:WorkspaceEnvironment):
prompt = AgentPrompt()
if todo.last_check_result:
prompt.append(AgentPrompt(todo.last_check_result))
prompt.append(todo.detail)
prompt.append(todo.result)
check_result: AgentTodoResult = await self._llm_check_todo(todo, prompt, workspace)
todo.last_check_time = datetime.datetime.now().timestamp()
match check_result.result_code:
case AgentTodoResult.TODO_RESULT_CODE_LLM_ERROR:
continue
case AgentTodoResult.TODO_RESULT_CODE_OK:
await todo_list.update_todo(todo.todo_id,AgentTodo.TODO_STATE_DONE)
case AgentTodoResult.TODO_RESULT_CODE_EXEC_OP_ERROR:
await todo_list.update_todo(todo.todo_id,AgentTodo.TDDO_STATE_CHECKFAILED)
await todo_list.append_worklog(todo, check_result)
self.agent_energy -= 1
check_count += 1
continue
review_prompts = self._can_review_todo(todo_list_type, todo)
if review_prompts:
prompt.append(workspace.get_prompt())
prompt.append(workspace.get_role_prompt(self.agent_id))
prompt.append(self.get_review_todo_prompt(todo))
prompt.append(review_prompts)
todo_tree = workspace.get_todo_tree("/")
todo_tree = todo_list.get_todo_tree("/")
prompt.append(AgentPrompt(todo_tree))
inner_functions,_ = BaseAIAgent.get_inner_functions(self.owner_env)
task_result:ComputeTaskResult = await self.do_llm_complection(prompt,inner_functions=inner_functions)
if task_result.result_code != ComputeTaskResultCode.OK:
logger.error(f"_llm_review_todos compute error:{task_result.error_str}")
return
do_result : AgentTodoResult = await self._llm_review_todo(todo, prompt, workspace)
todo.last_review_time = datetime.datetime.now().timestamp()
return
match do_result.result_code:
case AgentTodoResult.TODO_RESULT_CODE_LLM_ERROR:
continue
case AgentTodoResult.TODO_RESULT_CODE_EXEC_OP_ERROR:
continue
case AgentTodoResult.TODO_RESULT_CODE_OK:
await todo_list.update_todo(todo.todo_id,AgentTodo.TODO_STATE_REVIEWED)
def get_do_prompt(self,todo:AgentTodo) -> AgentPrompt:
return self.do_prompt
await todo_list.append_worklog(todo,do_result)
self.agent_energy -= 1
review_count += 1
continue
def get_prompt_from_todo(self,todo:AgentTodo) -> AgentPrompt:
json_str = json.dumps(todo.raw_obj)
return AgentPrompt(json_str)
logger.info(f"agent {self.agent_id} ,check:{check_count} todo,do:{do_count} todo.")
async def need_review_todo(self,todo:AgentTodo,workspace:WorkspaceEnvironment) -> bool:
return False
async def can_check(self,todo:AgentTodo,workspace:WorkspaceEnvironment) -> bool:
if self.get_check_prompt(todo) is None:
return False
def _can_review_todo(self, todo_list_type: TodoListType, todo:AgentTodo) -> AgentPrompt:
do_prompts = self.todo_prompts[todo_list_type].get("review")
if not do_prompts:
return None
if todo.can_review() is False:
return None
return do_prompts
def _can_check_todo(self, todo_list_type: TodoListType, todo:AgentTodo) -> AgentPrompt:
do_prompts = self.todo_prompts[todo_list_type].get("check")
if not do_prompts:
return None
if todo.can_check() is False:
return False
return None
if todo.checker is not None:
if todo.checker != self.agent_id:
return False
return None
else:
if self.can_do_unassigned_task is False:
return False
return None
else:
todo.checker = self.agent_id
return True
return do_prompts
def _can_do_todo(self, todo_list_type: TodoListType, todo:AgentTodo) -> AgentPrompt:
do_prompts = self.todo_prompts[todo_list_type].get("do")
if not do_prompts:
return None
async def can_do(self,todo:AgentTodo,workspace:WorkspaceEnvironment) -> bool:
if todo.can_do() is False:
return False
return None
if todo.worker is not None:
if todo.worker != self.agent_id:
return False
return None
else:
if self.can_do_unassigned_task is False:
return False
return None
else:
todo.worker = self.agent_id
return True
return do_prompts
async def _llm_do(self,todo:AgentTodo,workspace:WorkspaceEnvironment) -> AgentTodoResult:
async def _llm_do_todo(self, todo: AgentTodo, prompt: AgentPrompt, workspace: WorkspaceEnvironment) -> AgentTodoResult:
result = AgentTodoResult()
prompt : AgentPrompt = AgentPrompt()
#prompt.append(self.agent_prompt)
prompt.append(workspace.get_role_prompt(self.agent_id))
do_prompt = workspace.get_do_prompt(todo)
if do_prompt is None:
do_prompt = self.get_do_prompt(todo)
prompt.append(do_prompt)
# There are general methods for executing todos, as well as customized ones that are more efficient for specific types of TODOS.
# Based on experience, an Agent can autonomously master/organize execution methods for a greater variety of TODO types.
#prompt.append(work_log_prompt)
prompt.append(self.get_prompt_from_todo(todo))
task_result:ComputeTaskResult = await self.do_llm_complection(prompt)
task_result:ComputeTaskResult = await self.do_llm_complection(prompt, is_json_resp=True)
if task_result.error_str is not None:
logger.error(f"_llm_do compute error:{task_result.error_str}")
result.result_code = AgentTodoResult.TODO_RESULT_CODE_LLM_ERROR
@@ -893,243 +763,55 @@ class AIAgent(BaseAIAgent):
resp = await AIBus.get_default_bus().post_message(msg)
logging.info(f"agent {self.agent_id} send msg to {msg.target} result:{resp}")
op_errors,have_error = await workspace.exec_op_list(llm_result.op_list,self.agent_id)
result_str, have_error = await workspace.exec_op_list(llm_result.op_list, self.agent_id)
if have_error:
result.result_code = AgentTodoResult.TODO_RESULT_CODE_EXEC_OP_ERROR
#result.error_str = error_str
return result
result.result_str = result_str
return result
async def append_toddo_result(self,todo,worksapce,llm_result,result_str):
pass
def get_check_prompt(self,todo:AgentTodo) -> AgentPrompt:
return self.check_prompt
async def _llm_check_todo(self, todo:AgentTodo,workspace:WorkspaceEnvironment) :
if self.get_check_prompt(todo) is None:
return None
prompt : AgentPrompt = AgentPrompt()
prompt.append(self.agent_prompt)
prompt.append(workspace.get_role_prompt(self.agent_id))
prompt.append(self.get_check_prompt(todo))
if todo.last_check_result:
prompt.append(AgentPrompt(todo.last_check_result))
prompt.append(todo.detail)
prompt.append(todo.result)
async def _llm_check_todo(self, todo: AgentTodo, prompt: AgentPrompt, workspace: WorkspaceEnvironment) -> AgentTodoResult:
result = AgentTodoResult()
inner_functions,_ = BaseAIAgent.get_inner_functions(workspace)
task_result:ComputeTaskResult = await self.do_llm_complection(prompt,inner_functions=inner_functions,is_json_resp=True)
if task_result.result_code != ComputeTaskResultCode.OK:
logger.error(f"_llm_check_todo compute error:{task_result.error_str}")
return False
if task_result.result_str == "OK":
return True
if task_result.error_str is not None:
logger.error(f"_llm_do compute error:{task_result.error_str}")
result.result_code = AgentTodoResult.TODO_RESULT_CODE_LLM_ERROR
result.error_str = task_result.error_str
return result
result.result_str = task_result.result_str
todo.last_check_result = task_result.result_str
return False
return result
# 尝试自我学习，会主动获取、读取资料并进行整理
# LLM的本质能力是处理海量知识，应该让LLM能基于知识把自己的工作处理的更好
async def do_self_learn(self) -> None:
# 不同的workspace是否应该有不同的学习方法？
workspace = self.get_workspace_by_msg(None)
hash_list = workspace.kb_db.get_knowledge_without_llm_title()
for hash in hash_list:
if self.agent_energy <= 0:
break
async def _llm_review_todo(self, todo:AgentTodo, prompt: AgentPrompt, workspace: WorkspaceEnvironment):
inner_functions,_ = BaseAIAgent.get_inner_functions(workspace)
knowledge = workspace.kb_db.get_knowledge(hash)
if knowledge is None:
continue
full_path = knowledge.get("full_path")
if full_path is None:
continue
if os.path.exists(full_path) is False:
logger.warning(f"do_self_learn: knowledge {full_path} is not exists!")
continue
#TODO 可以用v-db 对不同目录的名字进行选择后，先进行一次快速的插入。有时间再慢慢用LLM整理
result_obj = await self._llm_read_article(knowledge,full_path)
#根据结果更新knowledge
if result_obj is not None:
workspace.kb_db.set_knowledge_llm_result(hash,result_obj)
# 在知识库中创建软链接
path_list = result_obj.get("path")
new_title = result_obj.get("title")
if path_list:
for new_path in path_list:
full_new_path = f"/knowledge{new_path}/{new_title}"
await workspace.symlink(full_path,full_new_path)
logger.info(f"create soft link {full_path} -> {full_new_path}")
self.agent_energy -= 1
# match item.type():
# case "book":
# self.llm_read_book(kb,item)
# learn_power -= 1
# case "article":
#
# self.llm_read_article(kb,item)
# learn_power -= 1
# case "video":
# self.llm_watch_video(kb,item)
# learn_power -= 1
# case "audio":
# self.llm_listen_audio(kb,item)
# learn_power -= 1
# case "code_project":
# self.llm_read_code_project(kb,item)
# learn_power -= 1
# case "image":
# self.llm_view_image(kb,item)
# learn_power -= 1
# case "other":
# self.llm_read_other(kb,item)
# learn_power -= 1
# case _:
# self.llm_learn_any(kb,item)
# pass
async def do_blance_knowledge_base(selft):
# 整理自己的知识库(让分类更平衡，更由于自己以后的工作)，并尝试更新学习目标
current_path = "/"
current_list = kb.get_list(current_path)
self_assessment_with_goal = self.get_self_assessment_with_goal()
learn_goal = {}
llm_blance_knowledge_base(current_path,current_list,self_assessment_with_goal,learn_goal,learn_power)
# 主动学习
# 方法目前只有使用搜索引擎一种？
for goal in learn_goal.items():
self.llm_learn_with_search_engine(kb,goal,learn_power)
if learn_power <= 0:
break
def parser_learn_llm_result(self,llm_result:LLMResult):
pass
async def gen_known_info_for_knowledge_prompt(self,knowledge_item:dict,temp_meta = None,need_catalogs = False) -> AgentPrompt:
workspace =self.get_workspace_by_msg(None)
kb_tree = await workspace.get_knowledege_catalog()
known_obj = {}
title = knowledge_item.get("title")
if title:
known_obj["title"] = title
summary = knowledge_item.get("summary")
if summary:
known_obj["summary"] = summary
tags = knowledge_item.get("tags")
if tags:
known_obj["tags"] = tags
if need_catalogs:
catalogs = knowledge_item.get("catalogs")
if catalogs:
known_obj["catalogs"] = catalogs
if temp_meta:
for key in temp_meta.keys():
known_obj[key] = temp_meta[key]
org_path = knowledge_item.get("full_path")
known_obj["orginal_path"] = org_path
know_info_str = f"# Known information:\n## Current directory structure:\n{kb_tree}\n## Knowlege Metadata:\n{json.dumps(known_obj)}\n"
return AgentPrompt(know_info_str)
async def _llm_read_article(self,knowledge_item:dict,full_path:str) -> ComputeTaskResult:
# Objectives:
# Obtain better titles, abstracts, table of contents (if necessary), tags
# Determine the appropriate place to put it (in line with the organization's goals)
# Known information:
# The reason why the target service's learn_prompt is being sorted
# Summary of the organization's work (if any)
# The current structure of the knowledge base (note the size control) gen_kb_tree_prompt (when empty, LLM should generate an appropriate initial directory structure)
# Original path, current title, abstract, table of contents
# Sorting long files (general tricks)
# Indicate that the input is part of the content, let LLM generate intermediate results for the task
# Enter the content in sequence, when the last content block is input, LLM gets the result
#full_content = item.get_article_full_content()
workspace = self.get_workspace_by_msg(None)
full_content_len = self.token_len(full_content)
if full_content_len < self.get_llm_learn_token_limit():
# 短文章不用总结catelog
#path_list,summary = llm_get_summary(summary,full_content)
#prompt = self.get_agent_role_prompt()
prompt = AgentPrompt()
prompt.append(self.get_learn_prompt())
known_info_prompt = await self.gen_known_info_for_knowledge_prompt(knowledge_item)
prompt.append(known_info_prompt)
content_prompt = AgentPrompt(full_content)
prompt.append(content_prompt)
env_functions = None
#env_functions,function_len = workspace.get_knowledge_base_ai_functions()
task_result:ComputeTaskResult = await self.do_llm_complection(prompt,is_json_resp=True)
task_result:ComputeTaskResult = await self.do_llm_complection(prompt,inner_functions=inner_functions)
if task_result.result_code != ComputeTaskResultCode.OK:
result_obj = {}
result_obj["error_str"] = task_result.error_str
return result_obj
logger.error(f"_llm_review_todos compute error:{task_result.error_str}")
return
result_obj = json.loads(task_result.result_str)
return result_obj
return
else:
logger.warning(f"llm_read_article: article {full_path} use LLM loop learn!")
pos = 0
read_len = int(self.get_llm_learn_token_limit() * 1.2)
# async def do_blance_knowledge_base(selft):
# # 整理自己的知识库(让分类更平衡，更由于自己以后的工作)，并尝试更新学习目标
# current_path = "/"
# current_list = kb.get_list(current_path)
# self_assessment_with_goal = self.get_self_assessment_with_goal()
# learn_goal = {}
temp_meta_data = {}
is_final = False
while pos < str_len:
_content = full_content[pos:pos+read_len]
part_cotent_len = len(_content)
if part_cotent_len < read_len:
# last chunk
is_final = True
part_content = f"<<Final Part:start at {pos}>>\n{_content}"
else:
part_content = f"<<Part:start at {pos}>>\n{_content}"
pos = pos + read_len
prompt = AgentPrompt()
prompt.append(self.get_learn_prompt())
known_info_prompt = await self.gen_known_info_for_knowledge_prompt(knowledge_item,temp_meta_data)
prompt.append(known_info_prompt)
content_prompt = AgentPrompt(part_content)
prompt.append(content_prompt)
#env_functions,function_len = workspace.get_knowledge_base_ai_functions()
task_result:ComputeTaskResult = await self.do_llm_complection(prompt,is_json_resp=True)
if task_result.result_code != ComputeTaskResultCode.OK:
result_obj = {}
result_obj["error_str"] = task_result.error_str
return result_obj
result_obj = json.loads(task_result.result_str)
temp_meta_data = result_obj
if is_final:
return result_obj
return None
# llm_blance_knowledge_base(current_path,current_list,self_assessment_with_goal,learn_goal,learn_power)
# # 主动学习
# # 方法目前只有使用搜索引擎一种？
# for goal in learn_goal.items():
# self.llm_learn_with_search_engine(kb,goal,learn_power)
# if learn_power <= 0:
# break
async def do_self_think(self):
session_id_list = AIChatSession.list_session(self.agent_id,self.chat_db)
@@ -1139,16 +821,15 @@ class AIAgent(BaseAIAgent):
used_energy = await self.think_chatsession(session_id)
self.agent_energy -= used_energy
todo_logs = await self.get_todo_logs()
for todo_log in todo_logs:
if self.agent_energy <= 0:
break
used_energy = await self.think_todo_log(todo_log)
self.agent_energy -= used_energy
# todo_logs = await self.get_todo_logs()
# for todo_log in todo_logs:
# if self.agent_energy <= 0:
# break
# used_energy = await self.think_todo_log(todo_log)
# self.agent_energy -= used_energy
return
async def think_todo_log(self,todo_log:AgentWorkLog):
pass
@@ -1164,7 +845,7 @@ class AIAgent(BaseAIAgent):
prompt:AgentPrompt = AgentPrompt()
#prompt.append(self._get_agent_prompt())
prompt.append(await self._get_agent_think_prompt())
system_prompt_len = prompt.get_prompt_token_len()
system_prompt_len = self.token_len(prompt=prompt)
#think env?
history_prompt,next_pos = await self._get_history_prompt_for_think(chatsession,summary,system_prompt_len,cur_pos)
prompt.append(history_prompt)
@@ -1224,24 +905,9 @@ class AIAgent(BaseAIAgent):
return None,0
def need_work(self) -> bool:
if self.do_prompt is not None:
return True
if self.check_prompt is not None:
return True
if self.agent_energy > 2:
return True
return False
def need_self_think(self) -> bool:
return False
def need_self_learn(self) -> bool:
if self.learn_prompt is not None:
return True
return False
def wake_up(self) -> None:
if self.agent_task is None:
@@ -1266,26 +932,20 @@ class AIAgent(BaseAIAgent):
continue
# complete & check todo
if self.need_work():
await self.do_my_work()
await self._llm_run_todo_list(TodoListType.TO_WORK)
# review other's todo
# self.review_other_works()
await self._llm_run_todo_list(TodoListType.TO_LEARN)
if self.need_self_think():
await self.do_self_think()
if self.need_self_learn():
await self.do_self_learn()
# review other's todo
# self.review_other_works()
except Exception as e:
tb_str = traceback.format_exc()
logger.error(f"agent {self.agent_id} on timer error:{e},{tb_str}")
continue
def token_len(self,text:str) -> int:
return ComputeKernel.llm_num_tokens_from_text(text,self.get_llm_model_name())
src/aios/agent/agent_base.py
+39 -19
View File
@@ -14,7 +14,7 @@ from typing import List, Tuple
from .ai_function import FunctionItem, AIFunction
from ..proto.agent_msg import AgentMsg, AgentMsgType
from ..proto.compute_task import ComputeTaskResult,ComputeTaskResultCode
from ..environment.environment import Environment
from ..environment.environment import BaseEnvironment
logger = logging.getLogger(__name__)
@@ -56,16 +56,6 @@ class AgentPrompt:
self.messages.extend(prompt.messages)
def get_prompt_token_len(self):
result = 0
if self.system_message:
result += len(self.system_message.get("content"))
for msg in self.messages:
result += len(msg.get("content"))
return result
def load_from_config(self,config:list) -> bool:
if isinstance(config,list) is not True:
logger.error("prompt is not list!")
@@ -139,6 +129,8 @@ class LLMResult:
if llm_result_str[0] == "{":
return LLMResult.from_json_str(llm_result_str)
# if llm_result_str.startswith("json"):
# return LLMResult.from_json_str(llm_result_str[4:])
lines = llm_result_str.splitlines()
is_need_wait = False
@@ -245,8 +237,9 @@ class AgentTodo:
TODO_STATE_EXEC_FAILED = "exec_failed"
TDDO_STATE_CHECKFAILED = "check_failed"
TODO_STATE_CASNCEL = "cancel"
TODO_STATE_CANCEL = "cancel"
TODO_STATE_DONE = "done"
TODO_STATE_REVIEWED = "reviewed"
TODO_STATE_EXPIRED = "expired"
def __init__(self):
@@ -342,6 +335,23 @@ class AgentTodo:
return result
def to_prompt(self) -> AgentPrompt:
json_str = json.dumps(self.raw_obj)
return AgentPrompt(json_str)
def can_review(self) -> bool:
if self.state != AgentTodo.TODO_STATE_DONE:
return False
now = datetime.now().timestamp()
if self.last_review_time:
time_diff = now - self.last_review_time
if time_diff < 60*15:
logger.info(f"todo {self.title} is already reviewed, ignore")
return False
return True
def can_check(self)->bool:
if self.state != AgentTodo.TODO_STATE_WAITING_CHECK:
return False
@@ -360,7 +370,7 @@ class AgentTodo:
case AgentTodo.TODO_STATE_DONE:
logger.info(f"todo {self.title} is done, ignore")
return False
case AgentTodo.TODO_STATE_CASNCEL:
case AgentTodo.TODO_STATE_CANCEL:
logger.info(f"todo {self.title} is cancel, ignore")
return False
case AgentTodo.TODO_STATE_EXPIRED:
@@ -410,12 +420,22 @@ class BaseAIAgent(abc.ABC):
def get_max_token_size(self) -> int:
pass
@abstractmethod
async def _process_msg(self,msg:AgentMsg,workspace = None) -> AgentMsg:
pass
def token_len(self, text:str=None, prompt:AgentPrompt=None) -> int:
from ..frame.compute_kernel import ComputeKernel
if text:
return ComputeKernel.llm_num_tokens_from_text(text,self.get_llm_model_name())
elif prompt:
result = 0
if prompt.system_message:
result += ComputeKernel.llm_num_tokens_from_text(prompt.system_message.get("content"),self.get_llm_model_name())
for msg in prompt.messages:
result += ComputeKernel.llm_num_tokens_from_text(msg.get("content"),self.get_llm_model_name())
return result
else:
return 0
@classmethod
def get_inner_functions(cls, env:Environment) -> (dict,int):
def get_inner_functions(cls, env:BaseEnvironment) -> (dict,int):
if env is None:
return None,0
@@ -440,7 +460,7 @@ class BaseAIAgent(abc.ABC):
self,
prompt:AgentPrompt,
org_msg:AgentMsg=None,
env:Environment=None,
env:BaseEnvironment=None,
inner_functions=None,
is_json_resp=False,
) -> ComputeTaskResult:
@@ -493,7 +513,7 @@ class BaseAIAgent(abc.ABC):
async def _execute_func(
self,
env: Environment,
env: BaseEnvironment,
inner_func_call_node: dict,
prompt: AgentPrompt,
inner_functions: dict,
src/aios/agent/ai_function.py
+63 -4
View File
@@ -9,9 +9,6 @@ class ParameterDefine:
class AIFunction:
def __init__(self) -> None:
self.description : str = None
@abstractmethod
def get_name(self) -> str:
"""
@@ -24,7 +21,7 @@ class AIFunction:
"""
return a detailed description of what the function does
"""
return self.description
pass
@abstractmethod
def get_parameters(self) -> Dict:
@@ -112,6 +109,9 @@ class SimpleAIFunction(AIFunction):
def get_name(self) -> str:
return self.func_id
def get_description(self) -> str:
return self.description
def get_parameters(self) -> Dict:
if self.parameters is not None:
result = {}
@@ -142,3 +142,62 @@ class SimpleAIFunction(AIFunction):
def is_ready_only(self) -> bool:
return False
class AIOperation:
@abstractmethod
def get_name(self) -> str:
"""
return the name of the operation (should be snake case)
"""
pass
@abstractmethod
def get_description(self) -> str:
"""
return a detailed description of what the operation does
"""
pass
@abstractmethod
async def execute(self, params: dict) -> str:
"""
Execute the function and return a JSON serializable dict.
The parameters are passed in the form of kwargs
"""
pass
class SimpleAIOperation(AIOperation):
def __init__(self,op:str,description:str,func_handler:Coroutine) -> None:
self.op = op
self.description = description
self.func_handler = func_handler
def get_name(self) -> str:
return self.op
def get_description(self) -> str:
return self.description
async def execute(self, params: Dict) -> str:
if self.func_handler is None:
return "error: function not implemented"
return await self.func_handler(params)
class AIFunctionOperation(AIOperation):
def __init__(self, func: AIFunction) -> None:
self.func = func
super().__init__()
@abstractmethod
def get_name(self) -> str:
return self.func.get_name()
@abstractmethod
def get_description(self) -> str:
return self.func.get_description()
@abstractmethod
async def execute(self, params: dict) -> str:
self.func.execute(**params)
src/aios/agent/workflow.py
+7 -7
View File
@@ -18,7 +18,7 @@ from .ai_function import AIFunction,FunctionItem
from ..frame.compute_kernel import ComputeKernel
from ..frame.bus import AIBus
from ..environment.environment import Environment,EnvironmentEvent
from ..environment.environment import BaseEnvironment
from ..environment.workflow_env import WorkflowEnvironment
@@ -490,15 +490,15 @@ class Workflow:
def get_workflow_rule_prompt(self) -> AgentPrompt:
return self.rule_prompt
def _env_event_to_msg(self,env_event:EnvironmentEvent) -> AgentMsg:
# def _env_event_to_msg(self,env_event:EnvironmentEvent) -> AgentMsg:
# pass
def get_inner_environment(self,env_id:str) -> BaseEnvironment:
pass
def get_inner_environment(self,env_id:str) -> Environment:
pass
def connect_to_environment(self,the_env:Environment,conn_info:dict) -> None:
def connect_to_environment(self,the_env:BaseEnvironment,conn_info:dict) -> None:
if the_env is not None:
self.workflow_env.add_owner_env(the_env)
self.workflow_env.add_env(the_env)
#for event2msg in conn_info:
# for k,v in event2msg:
src/aios/environment/environment.py
+71 -115
View File
@@ -4,145 +4,101 @@
from abc import ABC, abstractmethod
from typing import Any, Callable, Optional,Dict,Awaitable,List
import logging
from ..agent.ai_function import AIFunction, AIOperation
from ..agent.ai_function import AIFunction
logger = logging.getLogger(__name__)
class EnvironmentEvent(ABC):
class BaseEnvironment:
def __init__(self, workspace: str) -> None:
pass
# @abstractmethod
# #TODO: how to use env? different env has different prompt
# def get_env_prompt(self) -> str:
# pass
@abstractmethod
def display(self) -> str:
def get_ai_function(self,func_name:str) -> AIFunction:
pass
EnvironmentEventHandler = Callable[[str,EnvironmentEvent],Awaitable[Any]]
class Environment:
_all_env = {}
@classmethod
def get_env_by_id(cls,env_id:str):
return cls._all_env.get(env_id)
@classmethod
def set_env_by_id(cls,id,env):
assert id == env.get_id()
cls._all_env[env.get_id()] = env
def __init__(self,env_id:str) -> None:
self.env_id = env_id
self.values:Dict[str,str] = {}
self.get_handlers:Dict[str,Callable] = {}
self.owner_env:Dict[str,Environment] = {}
# self.valid_keys:Dict[str,bool] = None
self.event_handlers:Dict[str,List[EnvironmentEventHandler]]= {}
self.functions : Dict[str,AIFunction] = {}
def get_id(self) -> str:
return self.env_id
def add_owner_env(self,env) -> None:
self.owner_env[env.get_id()] = env
#@abstractmethod
#TODO: how to use env? different env has different prompt
def get_env_prompt(self) -> str:
@abstractmethod
def get_all_ai_functions(self) -> List[AIFunction]:
pass
@abstractmethod
def get_ai_operation(self,op_name:str) -> AIOperation:
pass
@abstractmethod
def get_all_ai_operations(self) -> List[AIOperation]:
pass
def __getitem__(self, key):
return self.get_value(key)
@abstractmethod
def get_value(self,key:str) -> Optional[str]:
pass
# _all_env = {}
# @classmethod
# def get_env_by_id(cls,env_id:str):
# return cls._all_env.get(env_id)
# @classmethod
# def set_env_by_id(cls,id,env):
# assert id == env.get_id()
# cls._all_env[env.get_id()] = env
class SimpleEnvironment(BaseEnvironment):
def __init__(self, workspace: str) -> None:
super().__init__(workspace)
self.functions: Dict[str,AIFunction] = {}
self.operations: Dict[str,AIOperation] = {}
def add_ai_function(self,func:AIFunction) -> None:
if self.functions.get(func.get_name()) is not None:
logger.warn(f"add ai_function {func.get_name()} in env {self.env_id}:function already exist")
self.functions[func.get_name()] = func
def get_ai_function(self,func_name:str) -> AIFunction:
func = self.functions.get(func_name)
if func is not None:
return func
for owner_env in self.owner_env.values():
func = owner_env.get_ai_function(func_name)
if func is not None:
return func
return None
#def enable_ai_function(self,func_name:str) -> None:
# pass
#def disable_ai_function(self,func_name:str) -> None:
# pass
def get_all_ai_functions(self) -> List[AIFunction]:
func_list = []
func_list.extend(self.functions.values())
for owner_env in self.owner_env.values():
func_list.extend(owner_env.get_all_ai_functions())
return func_list
@abstractmethod
def _do_get_value(self,key:str) -> Optional[str]:
pass
def add_ai_operation(self,op:AIOperation) -> None:
self.operations[op.get_name()] = op
def register_get_handler(self,key:str,handler:Callable) -> None:
h = self.get_handlers.get(key)
if h is not None:
logger.warn(f"register get_handler {key} in env {self.env_id}:handler already exist")
self.get_handlers[key] = handler
def attach_event_handler(self,event_id:str,handler:Callable) -> None:
handler_list = self.event_handlers.get(event_id)
if handler_list is None:
handler_list = []
self.event_handlers[event_id] = handler_list
handler_list.append(handler)
def remove_event_handler(self,event_id:str,handler:Callable) -> None:
handler_list = self.event_handlers.get(event_id)
if handler is not None:
handler_list.remove(handler)
return
logger.warn(f"remove event_handler {event_id} in env {self.env_id}:handler not found")
async def fire_event(self,event_id:str,event:EnvironmentEvent) -> None:
handler_list = self.event_handlers.get(event_id)
if handler_list is not None:
for handler in handler_list:
await handler(self.env_id,event)
else:
logger.debug(f"fire event {event_id} in env {self.env_id}:handler not found")
return
def __getitem__(self, key):
return self.get_value(key)
def get_value(self,key:str) -> Optional[str]:
handler = self.get_handlers.get(key)
if handler is not None:
return handler()
s = self.values.get(key)
if isinstance(s,str):
return s
else:
logger.warn(f"get value {key} in env {self.env_id} failed!,type is not str")
s = self._do_get_value(key)
if s is not None:
return s
if self.owner_env is not None:
for env in self.owner_env.values():
s = env.get_value(key)
if s is not None:
return s
logger.warn(f"get value {key} in env {self.env_id} failed!,not found")
def get_ai_operation(self,op_name:str) -> AIOperation:
op = self.operations.get(op_name)
if op is not None:
return op
return None
def set_value(self, key: str, str_value: str,is_storage:bool = True):
logger.info(f"set value {key} in env {self.env_id} to {str_value}")
self.values[key] = str_value
def get_all_ai_operations(self) -> List[AIOperation]:
op_list = []
op_list.extend(self.operations.values())
return op_list
class CompositeEnvironment(SimpleEnvironment):
def __init__(self, workspace: str) -> None:
super().__init__(workspace)
self.envs: List[BaseEnvironment] = []
def add_env(self, env: BaseEnvironment) -> None:
self.envs.append(env)
functions = env.get_all_ai_functions()
for func in functions:
self.functions[func.get_name()] = func
operations = env.get_all_ai_operations()
for op in operations:
self.operations[op.get_name()] = op
src/aios/environment/workflow_env.py
+7 -7
View File
@@ -15,13 +15,13 @@ from ..frame.compute_kernel import ComputeKernel
from ..frame.contact_manager import ContactManager,Contact,FamilyMember
from ..storage.storage import AIStorage
from .environment import Environment,EnvironmentEvent
from .environment import SimpleEnvironment, CompositeEnvironment
from .script_to_speech_function import ScriptToSpeechFunction
from .image_2_text_function import Image2TextFunction
logger = logging.getLogger(__name__)
class CalenderEvent(EnvironmentEvent):
class CalenderEvent(SimpleEnvironment):
def __init__(self,data) -> None:
super().__init__()
self.event_name = "timer"
@@ -31,7 +31,7 @@ class CalenderEvent(EnvironmentEvent):
return f"#event timer:{self.data}"
# AI Calender GOAL: Let user use "create notify after 2 days" to create a timer event
class CalenderEnvironment(Environment):
class CalenderEnvironment(SimpleEnvironment):
def __init__(self, env_id: str) -> None:
super().__init__(env_id)
self.db_file = AIStorage.get_instance().get_myai_dir() / "calender.db"
@@ -302,7 +302,7 @@ class CalenderEnvironment(Environment):
return f'exec paint OK, saved as a local file, path is: {result.result["file"]}'
class PaintEnvironment(Environment):
class PaintEnvironment(BaseEnvironment):
def __init__(self, env_id: str) -> None:
super().__init__(env_id)
self.is_run = False
@@ -327,14 +327,14 @@ class PaintEnvironment(Environment):
# Default Workflow Environment(Context)
class WorkflowEnvironment(Environment):
class WorkflowEnvironment(CompositeEnvironment):
def __init__(self, env_id: str,db_file:str) -> None:
super().__init__(env_id)
self.db_file = db_file
self.local = threading.local()
self.table_name = "WorkflowEnv_" + env_id
self.add_ai_function(ScriptToSpeechFunction())
self.add_ai_function(Image2TextFunction())
# self.add_ai_function(ScriptToSpeechFunction())
# self.add_ai_function(Image2TextFunction())
def _get_conn(self):
src/aios/environment/workspace_env.py
+148 -616
View File
@@ -1,282 +1,92 @@
# this env is designed for workflow owner filesystem, support file/directory operations
import hashlib
import json
import subprocess
import logging
import tempfile
import threading
import traceback
import time
import ast
import sys
import os
import re
import asyncio
import aiofiles
from typing import Any,List
import os
import sqlite3
import asyncio
from typing import Any,List,Dict
import chardet
from markdown import Markdown
import PyPDF2
from ..proto.agent_msg import *
from ..agent.agent_base import AgentTodo,AgentPrompt,AgentTodoResult
from ..agent.ai_function import AIFunction,SimpleAIFunction
from ..agent.agent_base import AgentMsg,AgentTodo,AgentPrompt,AgentTodoResult
from ..agent.ai_function import AIFunction,SimpleAIFunction, SimpleAIOperation
from ..storage.storage import AIStorage,ResourceLocation
from .simple_kb_db import SimpleKnowledgeDB
from .environment import Environment,EnvironmentEvent
from .environment import SimpleEnvironment, CompositeEnvironment
logger = logging.getLogger(__name__)
class WorkspaceEnvironment(Environment):
def __init__(self, env_id: str) -> None:
super().__init__(env_id)
myai_path = AIStorage.get_instance().get_myai_dir()
self.root_path = f"{myai_path}/workspace/{env_id}"
class TodoListType:
TO_WORK = "work"
TO_LEARN = "learn"
class TodoListEnvironment(SimpleEnvironment):
def __init__(self, workspace, list_type) -> None:
super().__init__(workspace)
self.root_path = os.path.join(workspace, list_type)
if not os.path.exists(self.root_path):
os.makedirs(self.root_path+"/todos")
os.makedirs(self.root_path)
self.known_todo = {}
self.kb_db = SimpleKnowledgeDB(f"{self.root_path}/kb.db")
self.doc_dirs = {}
self._scan_thread = None
self._scan_dirthread = None
def set_root_path(self,path:str):
self.root_path = path
def get_prompt(self) -> AgentMsg:
self.db_path = os.path.join(self.root_path, "todo.db")
self.conn = None
try:
self.conn = sqlite3.connect(self.db_path)
except Exception as e:
logger.error("Error occurred while connecting to database: %s", e)
return None
def get_role_prompt(self,role_id:str) -> AgentPrompt:
return None
cursor = self.conn.cursor()
cursor.execute('''
CREATE TABLE IF NOT EXISTS todo_list (
id TEXT,
path TEXT
)
''')
self.conn.commit()
def get_knowledge_base(self,root_dir=None,indent=0) -> str:
pass
async def create_todo(params):
todoObj = AgentTodo.from_dict(params["todo"])
parent_id = params.get("parent")
return await self.create_todo(parent_id,todoObj)
self.add_ai_operation(SimpleAIOperation(
op="create_todo",
description="create todo",
func_handler=create_todo,
))
def get_do_prompt(self,todo:AgentTodo=None)->AgentPrompt:
return None
async def update_todo(params):
todo_id = params["id"]
new_stat = params["state"]
return await self.update_todo(todo_id,new_stat)
self.add_ai_operation(SimpleAIOperation(
op="update_todo",
description="update todo",
func_handler=update_todo,
))
# result mean: list[op_error_str],have_error
async def exec_op_list(self,oplist:List,agent_id:str)->tuple[List[str],bool]:
result_str = "op list is none"
if oplist is None:
return None,False
result_str = []
have_error = False
for op in oplist:
if op["op"] == "create":
await self.create(op["path"],op["content"])
elif op["op"] == "write_file":
is_append = op.get("is_append")
if is_append is None:
is_append = False
error_str = await self.write(op["path"],op["content"],is_append)
elif op["op"] == "delete":
error_str = await self.delete(op["path"])
elif op["op"] == "rename":
error_str = await self.rename(op["path"],op["new_name"])
elif op["op"] == "mkdir":
error_str = await self.mkdir(op["path"])
elif op["op"] == "create_todo":
todoObj = AgentTodo.from_dict(op["todo"])
todoObj.worker = agent_id
todoObj.createor = agent_id
parent_id = op.get("parent")
error_str = await self.create_todo(parent_id,todoObj)
elif op["op"] == "update_todo":
todo_id = op["id"]
new_stat = op["state"]
error_str = await self.update_todo(todo_id,new_stat)
def _get_todo_path(self,todo_id:str) -> str:
cursor = self.conn.cursor()
cursor.execute('''
SELECT path FROM todo_list WHERE id = ?
''',(todo_id,))
row = cursor.fetchone()
if row:
return row[0]
else:
logger.error(f"execute op list failed: unknown op:{op['op']}")
error_str = f"execute op list failed: unknown op:{op['op']}"
if error_str:
have_error = True
result_str.append(error_str)
else:
result_str.append(f"execute success!")
return result_str,have_error
# file system operation: list,read,write,delete,move,stat
# inner_function
async def list(self,path:str,only_dir:bool=False) -> str:
directory_path = self.root_path + path
items = []
with await aiofiles.os.scandir(directory_path) as entries:
async for entry in entries:
is_dir = entry.is_dir()
if only_dir and not is_dir:
continue
item_type = "directory" if is_dir else "file"
items.append({"name": entry.name, "type": item_type})
return json.dumps(items)
# inner_function
async def read(self,path:str) -> str:
file_path = self.root_path + path
cur_encode = "utf-8"
async with aiofiles.open(file_path,'rb') as f:
cur_encode = chardet.detect(await f.read())['encoding']
async with aiofiles.open(file_path, mode='r', encoding=cur_encode) as f:
content = await f.read(2048)
return content
# operation or inner_function (MOST IMPORTANT FUNCTION)
async def write(self,path:str,content:str,is_append:bool=False) -> str:
file_path = self.root_path + path
try:
if is_append:
async with aiofiles.open(file_path, mode='a', encoding="utf-8") as f:
await f.write(content)
else:
if content is None:
# create dir
dir_path = self.root_path + path
os.makedirs(dir_path)
return True
else:
file_path = self.root_path + path
os.makedirs(os.path.dirname(file_path),exist_ok=True)
async with aiofiles.open(file_path, mode='w', encoding="utf-8") as f:
await f.write(content)
return True
except Exception as e:
return str(e)
return None
# operation or inner_function
async def delete(self,path:str) -> str:
try:
file_path = self.root_path + path
os.remove(file_path)
except Exception as e:
return str(e)
return None
# operation or inner_function
async def move(self,path:str,new_path:str) -> str:
try:
file_path = self.root_path + path
new_path = self.root_path + new_path
os.rename(file_path,new_path)
except Exception as e:
return str(e)
return None
# inner_function
async def stat(self,path:str) -> str:
try:
file_path = self.root_path + path
stat = os.stat(file_path)
return json.dumps(stat)
except Exception as e:
return str(e)
# operation or inner_function
async def symlink(self,path:str,target:str) -> str:
try:
#file_path = self.root_path + path
target_path = self.root_path + target
dir_path = os.path.dirname(target_path)
os.makedirs(dir_path,exist_ok=True)
os.symlink(path,target_path)
except Exception as e:
logger.error("symlink failed:%s",e)
return str(e)
return None
# TODO use diff to update large file content
async def update_by_diff(self,path:str,diff):
pass
# doc system （read_only,agent cann't modify doc）
# inner_function
async def list_db(self) -> str:
pass
# inner_function
async def get_db_desc(self,db_name:str) -> str:
pass
# inner_function
async def query(self,db_name:str,sql:str) -> str:
pass
# search (web)
# inner_function
async def google_search(self,keyword:str,opt=None) -> str:
pass
# inner_function
async def local_search(self,keyword:str,root_path=None ,opt=None) -> str:
pass
# inner_function, might be return a image is better
async def web_get(self,url:str) -> str:
pass
# inner_function
async def blockchain_get(self,chainid:str,query:dict) -> str:
pass
# code interpreter
# inner_function or operation
async def eval_code(self,pycode:str) -> str:
pass
# operation or inner_function
async def improve_code(self,path:str):
pass
# operation or inner_function
async def run(self,file_path:str)->str:
pass
# operation or inner_function
async def pub_service(self,project_path:str):
pass
# operation or inner_function
async def exec_tx(self,chain_id:str,tx:dict) -> str:
pass
# social ability
# operation or inner_function
async def post_message(self,target:str,msg:AgentMsg,wait_time) -> AgentMsg:
pass
# operation or inner_function
async def add_contact(self,name:str,contact_info) -> str:
pass
# inner_function , include contact realtime info
async def get_contact(self,name_list:List[str],opt:dict) -> List:
pass
def _save_todo_path(self,todo_id:str,path:str):
cursor = self.conn.cursor()
cursor.execute('''
INSERT INTO todo_list (id,path) VALUES (?,?)
''',(todo_id,path))
self.conn.commit()
# Task/todo system , create,update,delete,query
async def get_todo_tree(self,path:str = None,deep:int = 4):
if path:
directory_path = self.root_path + "/todos/" + path
directory_path = os.path.join(self.root_path, path)
else:
directory_path = self.root_path + "/todos"
directory_path = self.root_path
str_result:str = "/todos\n"
@@ -309,9 +119,9 @@ class WorkspaceEnvironment(Environment):
async def get_todo_list(self,agent_id:str,path:str = None)->List[AgentTodo]:
logger.info("get_todo_list:%s,%s",agent_id,path)
if path:
directory_path = self.root_path + "/todos/" + path
directory_path = os.path.join(self.root_path, path)
else:
directory_path = self.root_path + "/todos"
directory_path = self.root_path
result_list:List[AgentTodo] = []
@@ -354,17 +164,14 @@ class WorkspaceEnvironment(Environment):
detail_path = path + "/detail"
try:
async with aiofiles.open(detail_path, mode='r', encoding="utf-8") as f:
content = await f.read(4096)
logger.debug("get_todo_by_fullpath:%s,content:%s",path,content)
todo_dict = json.loads(content)
with open(detail_path, mode='r', encoding="utf-8") as f:
todo_dict = json.load(f)
result_todo = AgentTodo.from_dict(todo_dict)
if result_todo:
relative_path = os.path.relpath(path, self.root_path + "/todos/")
relative_path = os.path.relpath(path, self.root_path)
if not relative_path.startswith('/'):
relative_path = '/' + relative_path
result_todo.todo_path = relative_path
self.known_todo[result_todo.todo_id] = result_todo
else:
logger.error("get_todo_by_path:%s,parse failed!",path)
@@ -373,31 +180,25 @@ class WorkspaceEnvironment(Environment):
logger.error("get_todo_by_path:%s,failed:%s",path,e)
return None
async def get_todo(self,id:str) -> AgentTodo:
return self.known_todo.get(id)
async def create_todo(self,parent_id:str,todo:AgentTodo) -> str:
try:
if parent_id:
if parent_id not in self.known_todo:
logger.error("create_todo failed: parent_id not found!")
return False
parent_path = self.known_todo.get(parent_id).todo_path
todo_path = f"{parent_path}/{todo.title}"
parent_path = self._get_todo_path(parent_id)
todo_path = f"{parent_path}/{todo.todo_id}-{todo.title}"
else:
todo_path = todo.title
todo_path = f"{todo.todo_id}-{todo.title}"
dir_path = f"{self.root_path}/todos/{todo_path}"
dir_path = f"{self.root_path}/{todo_path}"
os.makedirs(dir_path)
detail_path = f"{dir_path}/detail"
if todo.todo_path is None:
todo.todo_path = todo_path
self._save_todo_path(todo.todo_id,todo_path)
logger.info("create_todo %s",detail_path)
async with aiofiles.open(detail_path, mode='w', encoding="utf-8") as f:
await f.write(json.dumps(todo.to_dict()))
self.known_todo[todo.todo_id] = todo
except Exception as e:
logger.error("create_todo failed:%s",e)
return str(e)
@@ -406,10 +207,12 @@ class WorkspaceEnvironment(Environment):
async def update_todo(self,todo_id:str,new_stat:str)->str:
try:
todo : AgentTodo = self.known_todo.get(todo_id)
todo_path = self._get_todo_path(todo_id)
full_path = f"{self.root_path}/{todo_path}"
todo : AgentTodo = await self.get_todo_by_fullpath(full_path)
if todo:
todo.state = new_stat
detail_path = f"{self.root_path}/todos/{todo.todo_path}/detail"
detail_path = f"{full_path}/detail"
async with aiofiles.open(detail_path, mode='w', encoding="utf-8") as f:
await f.write(json.dumps(todo.to_dict()))
return None
@@ -418,8 +221,36 @@ class WorkspaceEnvironment(Environment):
except Exception as e:
return str(e)
async def append_worklog(self,todo:AgentTodo,result:AgentTodoResult):
worklog = f"{self.root_path}/todos/{todo.todo_path}/.worklog"
async def wait_todo_done(self,todo_id:str,state=AgentTodo.TODO_STATE_WAITING_CHECK) -> AgentTodo:
todo_path = self._get_todo_path(todo_id)
full_path = f"{self.root_path}/{todo_path}"
async def check_done():
while True:
todo : AgentTodo = await self.get_todo_by_fullpath(full_path)
if todo is None:
continue
if todo.state == AgentTodo.TODO_STATE_CANCEL:
break
elif todo.state == AgentTodo.TODO_STATE_EXPIRED:
break
elif todo.state == AgentTodo.TODO_STATE_WAITING_CHECK:
if state == AgentTodo.TODO_STATE_WAITING_CHECK:
break
elif todo.state == AgentTodo.TODO_STATE_DONE:
if state == AgentTodo.TODO_STATE_WAITING_CHECK:
break
elif todo.state == AgentTodo.TODO_STATE_DONE:
break
elif todo.state == AgentTodo.TODO_STATE_REVIEWED:
break
await asyncio.sleep(1)
await check_done()
return await self.get_todo_by_fullpath(full_path)
async def append_worklog(self, todo:AgentTodo, result:AgentTodoResult):
worklog = f"{self.root_path}/{todo.todo_path}/.worklog"
async with aiofiles.open(worklog, mode='w+', encoding="utf-8") as f:
content = await f.read()
@@ -434,356 +265,57 @@ class WorkspaceEnvironment(Environment):
json_obj["logs"] = logs
await f.write(json.dumps(json_obj))
async def set_wakeup_timer(self,todo_id:str,timestamp:int) -> str:
pass
# knowledge base system
def get_knowledge_base_ai_functions(self):
all_inner_function = []
all_inner_function.append(SimpleAIFunction("get_knowledge_catalog","get knowledge catalog in tree format",
self.get_knowledege_catalog,
{"path":f"catalog path,none is /","depth":"max depth of catalog tree,default is 4"}))
all_inner_function.append(SimpleAIFunction("get_knowledge","get knowledge metadata",
self.get_knowledge,
{"path":f"knowledge path"}))
all_inner_function.append(SimpleAIFunction("load_knowledge_content","load knowledge content",
self.load_knowledge_content,
{"path":f"knowledge path","pos":"start position of content","length":"length of content"}))
result_func = []
result_len = 0
for inner_func in all_inner_function:
func_name = inner_func.get_name()
this_func = {}
this_func["name"] = func_name
this_func["description"] = inner_func.get_description()
this_func["parameters"] = inner_func.get_parameters()
result_len += len(json.dumps(this_func)) / 4
result_func.append(this_func)
return result_func,result_len
async def get_knowledege_catalog(self,path:str=None,only_dir =True,max_depth:int=5)->str:
if path:
full_path = f"{self.root_path}/knowledge/{path}"
else:
full_path = f"{self.root_path}/knowledge"
catlogs,file_count = await self.get_directory_structure(full_path,max_depth,only_dir)
return catlogs
async def get_directory_structure(self,root_dir, max_depth:int=4, only_dir=True, indent=1):
file_count = 0
structure_str = ''
if os.path.isdir(root_dir):
sub_files = []
with os.scandir(root_dir) as it:
for entry in it:
if entry.is_dir():
sub_structure, sub_count = await self.get_directory_structure(entry.path, max_depth, only_dir, indent + 1)
if sub_structure:
structure_str += sub_structure
file_count += sub_count
else:
file_count += 1
sub_files.append(entry.name)
if only_dir is False:
for file_name in sub_files:
structure_str = structure_str + ' ' * (indent+1) + file_name + '\n'
dir_name = os.path.basename(root_dir)
dir_info = f"{dir_name} <count: {file_count}>"
structure_str = ' ' * indent + dir_info + '\n' + structure_str
if indent - 1 >= max_depth:
return None, file_count
else:
return structure_str, file_count
# inner_function
async def get_knowledge(self,path:str) -> str:
full_path = f"{self.root_path}/knowledge/{path}"
if os.islink(full_path):
org_path = os.readlink(full_path)
hash = self.kb_db.get_hash_by_doc_path(org_path)
if hash:
return self.kb_db.get_knowledge(org_path)
return "not found"
async def load_knowledge_content(self,path:str,pos:int=0,length:int=None) -> str:
if path.endswith("pdf"):
logger.info("load_knowledge_content:pdf")
dir_path = os.path.dirname(path)
base_name = os.path.basename(path)
text_content_path = f"{dir_path}/.{base_name}.txt"
if os.path.exists(text_content_path) is False:
return None
async with aiofiles.open(path, mode='r', encoding=cur_encode) as f:
await f.seek(pos)
content = await f.read(length)
return content
else:
async with aiofiles.open(path,'rb') as f:
cur_encode = chardet.detect(await f.read())['encoding']
async with aiofiles.open(path, mode='r', encoding=cur_encode) as f:
await f.seek(pos)
content = await f.read(length)
return content
return "load content failed."
def _add_document_dir(self,path:str):
self.doc_dirs[path] = 0
def _start_scan_document(self):
if self._scan_thread is None:
self._scan_thread = threading.Thread(target=self._scan_document)
self._scan_thread.start()
if self._scan_dirthread is None:
self._scan_dirthread = threading.Thread(target=self._scan_dir)
self._scan_dirthread.start()
def _parse_pdf_bookmarks(self,bookmarks, parent:list):
for item in bookmarks:
if isinstance(item,list):
self._parse_pdf_bookmarks(item,parent)
else:
if item.title:
new_item = {}
new_item["page"] = item.page.idnum
new_item["title"] = item.title
my_childs = []
if item.childs:
if len(item.childs) > 0:
self._parse_pdf_bookmarks(item.childs, my_childs)
new_item["childs"] = my_childs
parent.append(new_item)
else:
logger.warning("parse pdf bookmarks failed: item.title is None!")
return
def _parse_pdf(self,doc_path:str):
metadata = {}
with open(doc_path, 'rb') as file:
reader = PyPDF2.PdfReader(file)
try:
doc_info = reader.metadata
if doc_info:
if doc_info.title:
metadata["title"] = doc_info.title
if doc_info.author:
metadata["authors"] = doc_info.author
except Exception as e:
logger.warn("parse pdf metadata failed:%s",e)
dir_path = os.path.dirname(doc_path)
base_name = os.path.basename(doc_path)
text_content_path = f"{dir_path}/.{base_name}.txt"
full_text = ""
for page in reader.pages:
text = page.extract_text()
full_text += text
with open(text_content_path, 'w', encoding='utf-8') as f:
f.write(full_text)
try:
bookmarks = reader.outline
if bookmarks:
catalogs = []
self._parse_pdf_bookmarks(bookmarks,catalogs)
metadata["catalogs"] = json.dumps(catalogs)
except Exception as e:
logger.warn("parse pdf bookmarks failed:%s",e)
return metadata
def _parse_txt(self,doc_path:str):
return {}
def _parse_md(self,doc_path:str):
metadata = {}
cur_encode = "utf-8"
with open(doc_path,'rb') as f:
cur_encode = chardet.detect(f.read(1024))['encoding']
with open(doc_path, mode='r', encoding=cur_encode) as f:
content = f.read()
match = re.search(r'^# (.*)', content, re.MULTILINE)
if match:
metadata['title'] = match.group(1).strip()
md = Markdown(extensions=['toc'])
html_str = md.convert(content)
toc = md.toc
if toc:
metadata['catalogs'] = toc
return metadata
def _parse_document(self,doc_path:str):
hash_result = None
title = os.path.basename(doc_path)
meta_data = {}
with open(doc_path, "rb") as f:
hash_md5 = hashlib.md5()
for chunk in iter(lambda: f.read(1024*1024), b""):
hash_md5.update(chunk)
hash_result = hash_md5.hexdigest()
try:
if doc_path.endswith(".md"):
meta_data = self._parse_md(doc_path)
elif doc_path.endswith(".pdf"):
meta_data = self._parse_pdf(doc_path)
except Exception as e:
logger.error("parse document %s failed:%s",doc_path,e)
traceback.print_exc()
if meta_data.get("title"):
title = meta_data["title"]
logger.info("parse document %s!",doc_path)
return hash_result,title,meta_data
def _support_file(self,file_name:str) -> bool:
if file_name.startswith("."):
return False
if file_name.endswith(".pdf"):
return True
if file_name.endswith(".md"):
return True
if file_name.endswith(".txt"):
return True
return False
def _scan_dir(self):
while True:
time.sleep(10)
for directory in self.doc_dirs.keys():
now = time.time()
if now - self.doc_dirs[directory] > 60*15:
self.doc_dirs[directory] = time.time()
else:
continue
for root, dirs, files in os.walk(directory):
for file in files:
if self._support_file(file):
full_path = os.path.join(root, file)
full_path = os.path.normpath(full_path)
if self.kb_db.is_doc_exist(full_path):
continue
file_stat = os.stat(full_path)
if file_stat.st_size < 1:
continue
if file_stat.st_size < 1024*1024*8:
#parse and insert
hash,title,meta_data = self._parse_document(full_path)
self.kb_db.add_doc(full_path,file_stat.st_size,file_stat.st_mtime,hash)
self.kb_db.add_knowledge(hash,title,meta_data)
else:
self.kb_db.add_doc(full_path,file_stat.st_size,file_stat.st_mtime)
def _scan_document(self):
while True:
time.sleep(10)
parse_queue = self.kb_db.get_docs_without_hash()
for doc_path in parse_queue:
hash,title,meta_data = self._parse_document(doc_path)
self.kb_db.set_doc_hash(doc_path,hash)
self.kb_db.add_knowledge(hash,title,meta_data)
# merge to standard workspace env, **ABANDON this!**
class KnowledgeBaseFileSystemEnvironment(Environment):
class WorkspaceEnvironment(CompositeEnvironment):
def __init__(self, env_id: str) -> None:
super().__init__(env_id)
self.root_path = "."
myai_path = AIStorage.get_instance().get_myai_dir()
root_path = f"{myai_path}/workspace/{env_id}"
super().__init__(root_path)
operator_param = {
"path": "full path of target directory",
}
self.add_ai_function(SimpleAIFunction("list",
"list the files and sub directory in target directory,result is a json array",
self.list,operator_param))
self.root_path = root_path
if not os.path.exists(self.root_path):
os.makedirs()
operator_param = {
"path": "full path of target file",
}
self.add_ai_function(SimpleAIFunction("cat",
"cat the file content in target path,result is a string",
self.cat,operator_param))
self.todo_list: Dict[str, TodoListEnvironment] = {}
self.todo_list[TodoListType.TO_WORK] = TodoListEnvironment(self.root_path,TodoListType.TO_WORK)
self.todo_list[TodoListType.TO_LEARN] = TodoListEnvironment(self.root_path,TodoListType.TO_LEARN)
# default environments in workspace
self.add_env(self.todo_list[TodoListType.TO_WORK])
def set_root_path(self,path:str):
self.root_path = path
def get_prompt(self) -> AgentMsg:
return None
async def list(self,path:str) -> str:
directory_path = self.root_path + path
items = []
def get_role_prompt(self,role_id:str) -> AgentPrompt:
return None
with await aiofiles.os.scandir(directory_path) as entries:
async for entry in entries:
item_type = "directory" if entry.is_dir() else "file"
items.append({"name": entry.name, "type": item_type})
def get_do_prompt(self,todo:AgentTodo=None)->AgentPrompt:
return None
return json.dumps(items)
# result mean: list[op_error_str],have_error
async def exec_op_list(self,oplist:List,agent_id:str)->tuple[List[str],bool]:
result_str = "op list is none"
if oplist is None:
return None,False
async def cat(self,path:str) -> str:
file_path = self.root_path + path
cur_encode = "utf-8"
async with aiofiles.open(file_path,'rb') as f:
cur_encode = chardet.detect(await f.read())['encoding']
async with aiofiles.open(file_path, mode='r', encoding=cur_encode) as f:
content = await f.read(2048)
return content
class ShellEnvironment(Environment):
def __init__(self, env_id: str) -> None:
super().__init__(env_id)
operator_param = {
"command": "command will execute",
}
self.add_ai_function(SimpleAIFunction("shell_exec",
"execute shell command in linux bash",
self.shell_exec,operator_param))
#run_code_param = {
# "pycode": "python code will execute",
#}
#self.add_ai_function(SimpleAIFunction("run_code",
# "execute python code",
# self.run_code,run_code_param))
async def shell_exec(self,command:str) -> str:
import asyncio.subprocess
process = await asyncio.create_subprocess_shell(
command,
stdout=asyncio.subprocess.PIPE,
stderr=asyncio.subprocess.PIPE
)
stdout, stderr = await process.communicate()
returncode = process.returncode
if returncode == 0:
return f"Execute success! stdout is:\n{stdout}\n"
result_str = []
have_error = False
for op in oplist:
operation = self.get_ai_operation(op["op"])
if operation:
error_str = await operation.execute(op)
else:
return f"Execute failed! stderr is:\n{stderr}\n"
logger.error(f"execute op list failed: unknown op:{op['op']}")
error_str = f"execute op list failed: unknown op:{op['op']}"
if error_str:
have_error = True
result_str.append(error_str)
else:
result_str.append(f"execute success!")
return result_str,have_error
src/aios/knowledge/object/object.py
+3 -3
View File
@@ -73,6 +73,6 @@ class KnowledgeObject(ABC):
def encode(self) -> bytes:
return pickle.dumps(self)
# @staticmethod
# def decode(data: bytes) -> "ImageObject":
# return pickle.loads(data)
@staticmethod
def decode(data: bytes) -> "KnowledgeObject":
return pickle.loads(data)
src/aios/knowledge/pipeline.py
+17 -15
View File
@@ -6,17 +6,13 @@ from . import ObjectID, KnowledgeStore
from enum import Enum
class KnowledgePipelineJournal:
def __init__(self, time: datetime.datetime, object_id: str, input: str, parser: str):
def __init__(self, time: datetime.datetime, input: str, parser: str):
self.time = time
self.object_id = None if object_id is None else ObjectID.from_base58(object_id)
self.input = input
self.parser = parser
def is_finish(self) -> bool:
return self.object_id is None
def get_object_id(self) -> ObjectID:
return self.object_id
return self.input is None
def get_input(self) -> str:
return self.input
@@ -28,7 +24,7 @@ class KnowledgePipelineJournal:
if self.is_finish():
return f"{self.time}: finished)"
else:
return f"{self.time}: object:{self.object_id} input:{self.input}, parser:{self.parser})"
return f"{self.time}: input:{self.input}, parser:{self.parser})"
# init sqlite3 client
class KnowledgePipelineJournalClient:
@@ -42,18 +38,17 @@ class KnowledgePipelineJournalClient:
'''CREATE TABLE IF NOT EXISTS journal (
id INTEGER PRIMARY KEY AUTOINCREMENT,
time DATETIME DEFAULT CURRENT_TIMESTAMP,
object_id TEXT,
input TEXT,
parser TEXT)'''
)
conn.commit()
def insert(self, object_id: ObjectID, input: str, parser: str, timestamp: datetime.datetime = None):
def insert(self, input: str, parser: str, timestamp: datetime.datetime = None):
timestamp = datetime.datetime.now() if timestamp is None else timestamp
conn = sqlite3.connect(self.journal_path)
conn.execute(
"INSERT INTO journal (time, object_id, input, parser) VALUES (?, ?, ?, ?)",
(timestamp, str(object_id), input, parser),
"INSERT INTO journal (time, input, parser) VALUES (?, ?, ?)",
(timestamp, input, parser),
)
conn.commit()
@@ -61,7 +56,7 @@ class KnowledgePipelineJournalClient:
conn = sqlite3.connect(self.journal_path)
cursor = conn.cursor()
cursor.execute("SELECT * FROM journal ORDER BY id DESC LIMIT ?", (topn,))
return [KnowledgePipelineJournal(time, object_id, input, parser) for (_, time, object_id, input, parser) in cursor.fetchall()]
return [KnowledgePipelineJournal(time, input, parser) for (_, time, input, parser) in cursor.fetchall()]
class KnowledgePipelineEnvironment:
def __init__(self, pipeline_path: str):
@@ -87,8 +82,12 @@ class KnowledgePipelineState(Enum):
STOPPED = 2
FINISHED = 3
class NullParser:
async def parse(self, object_id):
return ""
class KnowledgePipeline:
def __init__(self, name: str, env: KnowledgePipelineEnvironment, input_init, input_params, parser_init, parser_params):
def __init__(self, name: str, env: KnowledgePipelineEnvironment, input_init, input_params=None, parser_init=None, parser_params=None):
self.name = name
self.state = KnowledgePipelineState.INIT
self.input_init = input_init
@@ -108,18 +107,21 @@ class KnowledgePipeline:
async def run(self):
if self.state == KnowledgePipelineState.INIT:
self.input = self.input_init(self.env, self.input_params)
if self.parser_init is None:
self.parser = NullParser()
else:
self.parser = self.parser_init(self.env, self.parser_params)
self.state = KnowledgePipelineState.RUNNING
if self.state == KnowledgePipelineState.RUNNING:
async for input in self.input.next():
if input is None:
self.state = KnowledgePipelineState.FINISHED
self.env.journal.insert(None, "finished", "finished")
self.env.journal.insert(None, None)
return
(object_id, input_journal) = input
if object_id is not None:
parser_journal = await self.parser.parse(object_id)
self.env.journal.insert(object_id, input_journal, parser_journal)
self.env.journal.insert(input_journal, parser_journal)
else:
return
if self.state == KnowledgePipelineState.STOPPED:
src/component/agent_manager/agent_manager.py
+28 -7
View File
@@ -6,7 +6,7 @@ import sys
import runpy
from typing import Any, Callable, Dict, List, Optional, Union
from aios import AIAgent,AIAgentTemplete,AIStorage,Environment,BaseAIAgent,PackageEnv,PackageEnvManager,PackageMediaInfo,PackageInstallTask
from aios import AIAgent,AIAgentTemplete,AIStorage,BaseAIAgent,PackageEnv,PackageEnvManager,PackageMediaInfo,PackageInstallTask,WorkspaceEnvironment
logger = logging.getLogger(__name__)
@@ -28,6 +28,7 @@ class AgentManager:
self.agent_templete_env : PackageEnv = None
self.agent_env : PackageEnv = None
self.db_path : str = None
self.environments: dict = {}
self.loaded_agent_instance : Dict[str,BaseAIAgent] = None
async def initial(self) -> None:
@@ -49,6 +50,15 @@ class AgentManager:
async def scan_all_agent(self)->None:
pass
def register_environment(self, env_id: str, init_env) -> None:
self.environments[env_id] = init_env
def init_environment(self, env_id: str, workspace: str):
if env_id not in self.environments:
logger.error(f"env {env_id} not found!")
return
return self.environments[env_id](workspace)
async def is_exist(self,agent_id:str) -> bool:
the_aget = await self.get(agent_id)
@@ -109,16 +119,27 @@ class AgentManager:
config = toml.loads(config_data)
result_agent = AIAgent()
workspace = config.get("workspace", config.get("instance_id"))
workspace = WorkspaceEnvironment(workspace)
config["workspace"] = workspace
if "owner_env" in config:
owner_env = config["owner_env"]
_, ext = os.path.splitext(owner_env)
def init_env(env_config: str):
_, ext = os.path.splitext(env_config)
if ext == ".py":
env_path = os.path.join(agent_media.full_path, owner_env)
owner_env = runpy.run_path(env_path)["init"]()
config["owner_env"] = owner_env
env_path = os.path.join(agent_media.full_path, env_config)
env = runpy.run_path(env_path)["init"](None, workspace.root_path)
else:
owner_env = Environment.get_env_by_id(config["owner_env"])
config["owner_env"] = owner_env
env = self.init_environment(env_config, workspace.root_path)
workspace.add_env(env)
if isinstance(owner_env, list):
for env in owner_env:
init_env(env)
else:
init_env(owner_env)
if result_agent.load_from_config(config) is False:
logger.error(f"load agent from {agent_media} failed!")
src/component/common_environment/__init__.py
+3
View File
@@ -0,0 +1,3 @@
from .local_document import LocalKnowledgeBase, ScanLocalDocument, ParseLocalDocument
from .local_file_system import FilesystemEnvironment
from .shell import ShellEnvironment
src/component/common_environment/local_document.py
+618
View File
@@ -0,0 +1,618 @@
import os
import aiofiles
import chardet
import string
import sqlite3
import json
import re
import threading
import logging
import hashlib
from markdown import Markdown
import PyPDF2
import datetime
from typing import Optional, List
from aios import *
from .local_file_system import FilesystemEnvironment
logger = logging.getLogger(__name__)
class MetaDatabase:
def __init__(self,db_path:str):
self.db_path = db_path
self._get_conn()
def _get_conn(self):
""" get db connection """
local = threading.local()
if not hasattr(local, 'conn'):
local.conn = self._create_connection(self.db_path)
return local.conn
def _create_connection(self, db_file):
""" create a database connection to a SQLite database """
conn = None
try:
conn = sqlite3.connect(db_file)
except Exception as e:
logger.error("Error occurred while connecting to database: %s", e)
return None
if conn:
self._create_tables(conn)
return conn
def _create_tables(self,conn):
cursor = conn.cursor()
cursor.execute('''
CREATE TABLE IF NOT EXISTS documents (
doc_path TEXT PRIMARY KEY,
length INTEGER,
last_modify TEXT,
doc_hash TEXT,
create_time TEXT
)
''')
cursor.execute('''
CREATE TABLE IF NOT EXISTS knowledge (
doc_hash TEXT PRIMARY KEY,
title TEXT,
summary TEXT,
content TEXT,
catalogs TEXT,
tags TEXT,
llm_title TEXT,
llm_summary TEXT,
create_time TEXT
)
''')
cursor.execute('''
CREATE INDEX IF NOT EXISTS idx_documents_doc_hash
ON documents (doc_hash)
''')
cursor.execute('''
CREATE INDEX IF NOT EXISTS idx_knowledge_tags
ON knowledge (tags)
''')
conn.commit()
def add_doc(self, doc_path: str, length: int, last_modify: str, doc_hash: Optional[str] = None):
conn = self._get_conn()
cursor = conn.cursor()
create_time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
cursor.execute('''
INSERT INTO documents (doc_path, length, last_modify, doc_hash,create_time)
VALUES (?, ?, ?, ?,?)
''', (doc_path, length, last_modify, doc_hash,create_time))
conn.commit()
def is_doc_exist(self, doc_path: str) -> bool:
conn = self._get_conn()
cursor = conn.cursor()
cursor.execute('''
SELECT doc_path
FROM documents
WHERE doc_path = ?
''', (doc_path,))
return len(cursor.fetchall()) > 0
def set_doc_hash(self, doc_path: str, doc_hash: str):
conn = self._get_conn()
cursor = conn.cursor()
cursor.execute('''
UPDATE documents
SET doc_hash = ?
WHERE doc_path = ?
''', (doc_hash, doc_path))
conn.commit()
def get_docs_without_hash(self,limit:int=1024) -> List[str]:
conn = self._get_conn()
cursor = conn.cursor()
cursor.execute('''
SELECT doc_path
FROM documents
WHERE doc_hash IS NULL OR doc_hash = ''
ORDER BY create_time DESC
LIMIT ?
''',(limit,))
return [row[0] for row in cursor.fetchall()]
#metadata["summary"]
#metadata["catalogs"]
#metadata["tags"]
def add_knowledge(self, doc_hash: str, metadata: dict,content:str = None,):
conn = self._get_conn()
cursor = conn.cursor()
create_time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
summary = metadata.get("summary", "")
catalogs = json.dumps(metadata.get("catalogs", {}))
title = metadata.get("title","")
tags = ','.join(metadata.get("tags", []))
cursor.execute('''
INSERT INTO knowledge (doc_hash, title , summary , catalogs , tags,create_time)
VALUES (?, ?, ?, ?, ?,?)
''', (doc_hash, title, summary, catalogs, tags,create_time))
conn.commit()
#llm_result["summary"]
#llm_result["tags"]
#llm_result["catalog"]
def set_knowledge_llm_result(self, doc_hash: str, meta: dict):
conn = self._get_conn()
cursor = conn.cursor()
title = meta.get("title", "")
summary = meta.get("summary", "")
catalogs = json.dumps(meta.get("catalogs", {}))
tags = ','.join(meta.get("tags", []))
cursor.execute('''
UPDATE knowledge
SET llm_title = ?,llm_summary = ?, catalogs = ?, tags = ?
WHERE doc_hash = ?
''', (title,summary, catalogs, tags, doc_hash))
conn.commit()
def get_hash_by_doc_path(self, doc_path: str) -> Optional[str]:
conn = self._get_conn()
cursor = conn.cursor()
cursor.execute('''
SELECT doc_hash
FROM documents
WHERE doc_path = ?
''', (doc_path,))
row = cursor.fetchone()
if row is None:
return None
return row[0]
def get_knowledge(self, doc_hash: str) -> Optional[dict]:
conn = self._get_conn()
cursor = conn.cursor()
cursor.execute('''
SELECT title, summary, catalogs, tags, llm_title, llm_summary
FROM knowledge
WHERE doc_hash = ?
''', (doc_hash,))
row = cursor.fetchone()
if row is None:
return None
# get doc path
cursor.execute('''
SELECT doc_path
FROM documents
WHERE doc_hash = ?
''', (doc_hash,))
row2 = cursor.fetchone()
if row2 is None:
return None
doc_path = row2[0]
return {
"full_path": doc_path,
"title": row[0],
"summary": row[1],
"catalogs": row[2],
"tags": row[3],
"llm_title" : row[4],
"llm_summary" : row[5],
}
def get_knowledge_without_llm_title(self,limit:int=16) -> List[str]:
conn = self._get_conn()
cursor = conn.cursor()
cursor.execute('''
SELECT doc_hash
FROM knowledge
WHERE llm_title IS NULL OR llm_title = ''
ORDER BY create_time DESC
LIMIT ?
''',(limit,))
return [row[0] for row in cursor.fetchall()]
def query_docs_by_tag(self, tag: str) -> List[str]:
conn = self._get_conn()
cursor = conn.cursor()
tag_json = json.dumps(tag) # 将标签转换为 JSON 字符串
cursor.execute('''
SELECT documents.doc_path
FROM documents
JOIN knowledge ON documents.doc_hash = knowledge.doc_hash
WHERE json_extract(knowledge.tags, '$') LIKE ?
''', (tag))
return [row[0] for row in cursor.fetchall()]
# singleton
class LearningCache:
_instance_lock = threading.Lock()
_instance = None
def __instance_init__(self):
self.cache = {}
self.cache_lock = threading.Lock()
def __new__(cls, *args, **kwargs):
if cls._instance is None:
with LearningCache._instance_lock:
if cls._instance is None:
cls._instance = super().__new__(cls)
cls._instance.__instance_init__()
return cls._instance
def add(self, key, value):
with self.cache_lock:
self.cache[key] = value
def get(self, key):
with self.cache_lock:
return self.cache.get(key)
def remove(self, key):
with self.cache_lock:
return self.cache.pop(key, None)
class LocalKnowledgeBase(CompositeEnvironment):
def __init__(self, workspace: str) -> None:
super().__init__(workspace)
self.root_path = f"{workspace}/knowledge"
if os.path.exists(self.root_path) is False:
os.makedirs(self.root_path)
self.meta_db = MetaDatabase(f"{self.root_path}/kb.db")
self.learning_cache = LearningCache()
async def learn(op:dict):
full_path = op.get("original_path")
if not full_path:
return
meta = self.learning_cache.get(full_path)
meta.update(op)
self.add_ai_operation(SimpleAIOperation(
op="learn",
description="update knowledge llm summary",
func_handler=learn,
))
self.fs = FilesystemEnvironment(self.root_path)
self.add_env(self.fs)
async def get_knowledege_catalog(self,path:str=None,only_dir =True,max_depth:int=5)->str:
if path:
full_path = f"{self.root_path}/{path}"
else:
full_path = self.root_path
catlogs,file_count = await self.get_directory_structure(full_path,max_depth,only_dir)
return catlogs
async def get_directory_structure(self,root_dir, max_depth:int=4, only_dir=True, indent=1):
file_count = 0
structure_str = ''
if os.path.isdir(root_dir):
sub_files = []
with os.scandir(root_dir) as it:
for entry in it:
if entry.is_dir():
sub_structure, sub_count = await self.get_directory_structure(entry.path, max_depth, only_dir, indent + 1)
if sub_structure:
structure_str += sub_structure
file_count += sub_count
else:
file_count += 1
sub_files.append(entry.name)
if only_dir is False:
for file_name in sub_files:
structure_str = structure_str + ' ' * (indent+1) + file_name + '\n'
dir_name = os.path.basename(root_dir)
dir_info = f"{dir_name} <count: {file_count}>"
structure_str = ' ' * indent + dir_info + '\n' + structure_str
if indent - 1 >= max_depth:
return None, file_count
else:
return structure_str, file_count
# inner_function
async def get_knowledge_meta(self,path:str) -> str:
full_path = f"{self.root_path}/{path}"
if os.islink(full_path):
org_path = os.readlink(full_path)
hash = self.meta_db.get_hash_by_doc_path(org_path)
if hash:
return self.meta_db.get_knowledge(org_path)
return "not found"
async def load_knowledge_content(self,path:str,pos:int=0,length:int=None) -> str:
if path.endswith("pdf"):
logger.info("load_knowledge_content:pdf")
dir_path = os.path.dirname(path)
base_name = os.path.basename(path)
text_content_path = f"{dir_path}/.{base_name}.txt"
if os.path.exists(text_content_path) is False:
return None
async with aiofiles.open(path, mode='r', encoding=cur_encode) as f:
await f.seek(pos)
content = await f.read(length)
return content
else:
async with aiofiles.open(path,'rb') as f:
cur_encode = chardet.detect(await f.read())['encoding']
async with aiofiles.open(path, mode='r', encoding=cur_encode) as f:
await f.seek(pos)
content = await f.read(length)
return content
class ScanLocalDocument:
def __init__(self, env: KnowledgePipelineEnvironment, config):
self.env = env
workspace = string.Template(config["workspace"]).substitute(myai_dir=AIStorage.get_instance().get_myai_dir())
path = string.Template(config["path"]).substitute(myai_dir=AIStorage.get_instance().get_myai_dir())
self.knowledge_base = LocalKnowledgeBase(workspace)
self.path = path
def _support_file(self,file_name:str) -> bool:
if file_name.startswith("."):
return False
if file_name.endswith(".pdf"):
return True
if file_name.endswith(".md"):
return True
if file_name.endswith(".txt"):
return True
return False
async def next(self):
while True:
for root, dirs, files in os.walk(self.path):
for file in files:
if self._support_file(file):
full_path = os.path.join(root, file)
full_path = os.path.normpath(full_path)
if self.knowledge_base.meta_db.is_doc_exist(full_path):
continue
yield(full_path, full_path)
else:
continue
yield(None, None)
class ParseLocalDocument:
def __init__(self, env: KnowledgePipelineEnvironment, config: dict):
self.env = env
workspace = string.Template(config["workspace"]).substitute(myai_dir=AIStorage.get_instance().get_myai_dir())
self.todo_list = TodoListEnvironment(workspace, TodoListType.TO_LEARN)
self.knowledge_base = LocalKnowledgeBase(workspace)
self.token_limit = config.get("token_limit", 4000)
self.assign_to = config.get("assign_to")
async def parse(self, full_path: str) -> str:
file_stat = os.stat(full_path)
if file_stat.st_size < 1:
return full_path
hash, parse_meta = self._parse_document(full_path)
parse_meta["original_path"] = full_path
llm_meta = await self._learn_by_agent(parse_meta)
self.knowledge_base.meta_db.add_doc(full_path,file_stat.st_size,file_stat.st_mtime,hash)
self.knowledge_base.meta_db.add_knowledge(hash,parse_meta)
self.knowledge_base.meta_db.set_knowledge_llm_result(hash,llm_meta)
path_list = llm_meta.get("path")
new_title = llm_meta.get("title")
if path_list:
for new_path in path_list:
new_path = f"{new_path}/{new_title}"
await self.knowledge_base.fs.symlink(full_path, new_path)
logger.info(f"create soft link {full_path} -> {new_path}")
return full_path
async def _get_meta_prompt(self,meta: dict,temp_meta = None,need_catalogs = False) -> str:
kb_tree = await self.knowledge_base.get_knowledege_catalog()
known_obj = {}
title = meta.get("title")
if title:
known_obj["title"] = title
summary = meta.get("summary")
if summary:
known_obj["summary"] = summary
tags = meta.get("tags")
if tags:
known_obj["tags"] = tags
if need_catalogs:
catalogs = meta.get("catalogs")
if catalogs:
known_obj["catalogs"] = catalogs
if temp_meta:
for key in temp_meta.keys():
known_obj[key] = temp_meta[key]
org_path = meta.get("original_path")
known_obj["original_path"] = org_path
return f"# Known information:\n## Current directory structure:\n{kb_tree}\n## Knowlege Metadata:\n{json.dumps(known_obj)}\n"
def _token_len(self, text: str) -> int:
return CustomAIAgent("", "gpt-4-1106-preview", self.token_limit).token_len(text=text)
async def _learn_by_agent(self, meta:dict) -> dict:
# Objectives:
# Obtain better titles, abstracts, table of contents (if necessary), tags
# Determine the appropriate place to put it (in line with the organization's goals)
# Known information:
# The reason why the target service's learn_prompt is being sorted
# Summary of the organization's work (if any)
# The current structure of the knowledge base (note the size control) gen_kb_tree_prompt (when empty, LLM should generate an appropriate initial directory structure)
# Original path, current title, abstract, table of contents
# Sorting long files (general tricks)
# Indicate that the input is part of the content, let LLM generate intermediate results for the task
# Enter the content in sequence, when the last content block is input, LLM gets the result
full_content = await self.knowledge_base.load_knowledge_content(meta["original_path"])
full_content_len = self._token_len(full_content)
full_path = meta["original_path"]
self.knowledge_base.learning_cache.add(full_path, meta)
if full_content_len < self.token_limit:
# 短文章不用总结catalog
todo = AgentTodo()
todo.worker = self.assign_to
todo.title = meta["title"]
meta_prompt = await self._get_meta_prompt(meta,None)
todo.detail = meta_prompt + full_content
await self.todo_list.create_todo(None, todo)
await self.todo_list.wait_todo_done(todo.todo_id)
else:
logger.warning(f"llm_read_article: article {full_path} use LLM loop learn!")
pos = 0
read_len = int(self.token_limit * 1.2)
is_final = False
while pos < full_content_len:
_content = full_content[pos:pos+read_len]
part_cotent_len = len(_content)
if part_cotent_len < read_len:
# last chunk
is_final = True
part_content = f"<<Final Part:start at {pos}>>\n{_content}"
else:
part_content = f"<<Part:start at {pos}>>\n{_content}"
pos = pos + read_len
temp_meta = self.knowledge_base.learning_cache.get(full_path)
todo = AgentTodo()
todo.worker = self.assign_to
todo.title = meta["title"]
meta_prompt = await self._get_meta_prompt(meta,temp_meta)
todo.detail = meta_prompt + part_content
self.todo_list.create_todo(None, todo)
todo = await self.todo_list.wait_todo_done(todo.todo_id)
if is_final:
break
return self.knowledge_base.learning_cache.remove(full_path)
def _parse_pdf_bookmarks(self,bookmarks, parent:list):
for item in bookmarks:
if isinstance(item,list):
self._parse_pdf_bookmarks(item,parent)
else:
if item.title:
new_item = {}
new_item["page"] = item.page.idnum
new_item["title"] = item.title
my_childs = []
if item.childs:
if len(item.childs) > 0:
self._parse_pdf_bookmarks(item.childs, my_childs)
new_item["childs"] = my_childs
parent.append(new_item)
else:
logger.warning("parse pdf bookmarks failed: item.title is None!")
return
def _parse_pdf(self,doc_path:str):
metadata = {}
with open(doc_path, 'rb') as file:
reader = PyPDF2.PdfReader(file)
try:
doc_info = reader.metadata
if doc_info:
if doc_info.title:
metadata["title"] = doc_info.title
if doc_info.author:
metadata["authors"] = doc_info.author
except Exception as e:
logger.warn("parse pdf metadata failed:%s",e)
dir_path = os.path.dirname(doc_path)
base_name = os.path.basename(doc_path)
text_content_path = f"{dir_path}/.{base_name}.txt"
full_text = ""
for page in reader.pages:
text = page.extract_text()
full_text += text
with open(text_content_path, 'w', encoding='utf-8') as f:
f.write(full_text)
try:
bookmarks = reader.outline
if bookmarks:
catalogs = []
self._parse_pdf_bookmarks(bookmarks,catalogs)
metadata["catalogs"] = json.dumps(catalogs)
except Exception as e:
logger.warn("parse pdf bookmarks failed:%s",e)
return metadata
def _parse_txt(self,doc_path:str):
return {}
def _parse_md(self,doc_path:str):
metadata = {}
cur_encode = "utf-8"
with open(doc_path,'rb') as f:
cur_encode = chardet.detect(f.read(1024))['encoding']
with open(doc_path, mode='r', encoding=cur_encode) as f:
content = f.read()
match = re.search(r'^# (.*)', content, re.MULTILINE)
if match:
metadata['title'] = match.group(1).strip()
md = Markdown(extensions=['toc'])
html_str = md.convert(content)
toc = md.toc
if toc:
metadata['catalogs'] = toc
return metadata
def _parse_document(self,doc_path:str):
hash_result = None
title = os.path.basename(doc_path)
meta_data = {}
with open(doc_path, "rb") as f:
hash_md5 = hashlib.md5()
for chunk in iter(lambda: f.read(1024*1024), b""):
hash_md5.update(chunk)
hash_result = hash_md5.hexdigest()
try:
if doc_path.endswith(".md"):
meta_data = self._parse_md(doc_path)
elif doc_path.endswith(".pdf"):
meta_data = self._parse_pdf(doc_path)
except Exception as e:
logger.error("parse document %s failed:%s",doc_path,e)
# traceback.print_exc()
if not "title" in meta_data:
meta_data["title"] = title
logger.info("parse document %s!",doc_path)
return hash_result, meta_data
src/component/common_environment/local_file_system.py
+139
View File
@@ -0,0 +1,139 @@
import json
import os
import aiofiles
from typing import Any,List,Dict
import chardet
from aios import SimpleAIOperation
from aios import SimpleEnvironment
class FilesystemEnvironment(SimpleEnvironment):
def __init__(self, workspace: str) -> None:
super().__init__(workspace)
self.root_path = workspace
# if op["op"] == "create":
# await self.create(op["path"],op["content"])
async def write(op):
is_append = op.get("is_append")
if is_append is None:
is_append = False
return await self.write(op["path"],op["content"],is_append)
self.add_ai_operation(SimpleAIOperation(
op="write",
description="write file",
func_handler=write,
))
async def delete(op):
return await self.delete(op["path"])
self.add_ai_operation(SimpleAIOperation(
op="delete",
description="delete path",
func_handler=delete,
))
async def rename(op):
return await self.move(op["path"],op["new_name"])
self.add_ai_operation(SimpleAIOperation(
op="rename",
description="rename path",
func_handler=rename,
))
# file system operation: list,read,write,delete,move,stat
# inner_function
async def list(self,path:str,only_dir:bool=False) -> str:
directory_path = self.root_path + path
items = []
with await aiofiles.os.scandir(directory_path) as entries:
async for entry in entries:
is_dir = entry.is_dir()
if only_dir and not is_dir:
continue
item_type = "directory" if is_dir else "file"
items.append({"name": entry.name, "type": item_type})
return json.dumps(items)
# inner_function
async def read(self,path:str) -> str:
file_path = self.root_path + path
cur_encode = "utf-8"
async with aiofiles.open(file_path,'rb') as f:
cur_encode = chardet.detect(await f.read())['encoding']
async with aiofiles.open(file_path, mode='r', encoding=cur_encode) as f:
content = await f.read(2048)
return content
# operation or inner_function (MOST IMPORTANT FUNCTION)
async def write(self,path:str,content:str,is_append:bool=False) -> str:
file_path = self.root_path + path
try:
if is_append:
async with aiofiles.open(file_path, mode='a', encoding="utf-8") as f:
await f.write(content)
else:
if content is None:
# create dir
dir_path = self.root_path + path
os.makedirs(dir_path)
return True
else:
file_path = self.root_path + path
os.makedirs(os.path.dirname(file_path),exist_ok=True)
async with aiofiles.open(file_path, mode='w', encoding="utf-8") as f:
await f.write(content)
return True
except Exception as e:
return str(e)
return None
# operation or inner_function
async def delete(self,path:str) -> str:
try:
file_path = self.root_path + path
os.remove(file_path)
except Exception as e:
return str(e)
return None
# operation or inner_function
async def move(self,path:str,new_path:str) -> str:
try:
file_path = self.root_path + path
new_path = self.root_path + new_path
os.rename(file_path,new_path)
except Exception as e:
return str(e)
return None
# inner_function
async def stat(self,path:str) -> str:
try:
file_path = self.root_path + path
stat = os.stat(file_path)
return json.dumps(stat)
except Exception as e:
return str(e)
# operation or inner_function
async def symlink(self,path:str,target:str) -> str:
try:
#file_path = self.root_path + path
target_path = self.root_path + target
dir_path = os.path.dirname(target_path)
os.makedirs(dir_path,exist_ok=True)
os.symlink(path,target_path)
except Exception as e:
logger.error("symlink failed:%s",e)
return str(e)
return None
src/component/common_environment/shell.py
+38
View File
@@ -0,0 +1,38 @@
import os
from typing import Any,List,Dict
from aios import AgentMsg,AgentTodo,AgentPrompt
from aios import SimpleAIFunction, SimpleAIOperation
from aios import SimpleEnvironment
class ShellEnvironment(SimpleEnvironment):
def __init__(self, workspace: str) -> None:
super().__init__(workspace)
operator_param = {
"command": "command will execute",
}
self.add_ai_function(SimpleAIFunction("shell_exec",
"execute shell command in linux bash",
self.shell_exec,operator_param))
#run_code_param = {
# "pycode": "python code will execute",
#}
#self.add_ai_function(SimpleAIFunction("run_code",
# "execute python code",
# self.run_code,run_code_param))
async def shell_exec(self,command:str) -> str:
import asyncio.subprocess
process = await asyncio.create_subprocess_shell(
command,
stdout=asyncio.subprocess.PIPE,
stderr=asyncio.subprocess.PIPE
)
stdout, stderr = await process.communicate()
returncode = process.returncode
if returncode == 0:
return f"Execute success! stdout is:\n{stdout}\n"
else:
return f"Execute failed! stderr is:\n{stderr}\n"
src/component/knowledge_manager/pipeline.py
+7 -2
View File
@@ -44,14 +44,19 @@ class KnowledgePipelineManager:
input_init = self.input_modules.get(input_module)
input_params = config["input"].get("params")
parser_module = config["parser"]["module"]
parser_config = config.get("parser")
if parser_config is None:
parser_init = None
parser_params = None
else:
parser_module = parser_config["module"]
_, ext = os.path.splitext(parser_module)
if ext == ".py":
parser_module = os.path.join(path, parser_module)
parser_init = runpy.run_path(parser_module)["init"]
else:
parser_init = self.parser_modules.get(parser_module)
parser_params = config["parser"].get("params")
parser_params = parser_config.get("params")
data_path = os.path.join(self.root_dir, name)
src/component/mail_environment/local.py
+1 -1
View File
@@ -22,7 +22,7 @@ class LocalEmail:
if latest_journal.is_finish():
yield None
continue
parsed = str(latest_journal.get_object_id())
parsed = latest_journal.get_input()
mail_id = self.mail_storage.next_mail_id(parsed)
if mail_id is None:
src/component/mail_environment/mail.py
+52 -19
View File
@@ -7,17 +7,20 @@ import datetime
from bs4 import BeautifulSoup
import sqlite3
import html2text
from urllib.parse import urlparse
from aios import *
class Mail:
def __init__(self, **kwargs) -> None:
self.from_addr = kwargs.get("From")
self.to_addr = kwargs.get("To")
self.subject = kwargs.get("Subject")
self.date = kwargs.get("Date")
self.bcc = kwargs.get("BCC")
self.cc = kwargs.get("CC")
self.reply_to = None
self.from_addr = kwargs.get("from")
self.to_addr = kwargs.get("to")
self.subject = kwargs.get("subject")
self.date = kwargs.get("date")
self.bcc = kwargs.get("bcc")
self.cc = kwargs.get("cc")
self.reply_to = kwargs.get("reply_to")
self.id: str = None
self.content: str = None
@@ -192,20 +195,36 @@ class MailStorage:
self.conn.commit()
await asyncio.sleep(10)
def download(self, uid, mail: mailparser.MailParser):
def download(self, uid, parser: mailparser.MailParser,
save_image=True,
from_field="From",
to_field="To",
subject_field="Subject",
date_field="Date",
reply_to_field="In-Reply-To",
cc_field="CC",
bcc_field="BCC"):
mail_dir = self.mail_dir(uid)
os.makedirs(dir)
if not os.path.exists(mail_dir):
os.makedirs(mail_dir)
meta = json.loads(mail.mail_json)
mail = Mail(**meta)
reply_to = meta.get("In-Reply-To")
src_meta = json.loads(parser.mail_json)
meta = {}
meta["from"] = src_meta.get(from_field)
meta["to"] = src_meta.get(to_field)
meta["subject"] = src_meta.get(subject_field)
meta["date"] = src_meta.get(date_field)
meta["bcc"] = src_meta.get(bcc_field)
meta["cc"] = src_meta.get(cc_field)
reply_to = src_meta.get(reply_to_field)
if reply_to:
mail.reply_to = self.uid_to_object_id(reply_to)
meta["reply_to"] = self.uid_to_object_id(reply_to)
mail = Mail(**meta)
h = html2text.HTML2Text()
h.ignore_links = True
h.ignore_images = True
mail_content = h.handle(mail.body)
mail_content = h.handle(parser.body)
mail.content = mail_content
mail.calculate_id()
@@ -216,8 +235,9 @@ class MailStorage:
with open(f"{mail_dir}/mail.txt", "w", encoding='utf-8') as f:
f.write(mail_content)
for attachment in mail.attachments:
if attachment['mail_content_type'] in ['image/png', 'image/jpeg', 'image/gif']:
if save_image:
for attachment in parser.attachments:
if attachment['mail_content_type'] in ['image/png', 'image/jpg', 'image/jpeg', 'image/gif', 'image/svg']:
filename = attachment['filename']
filefullname = f"{mail_dir}/{filename}"
image_data = attachment['payload']
@@ -230,7 +250,7 @@ class MailStorage:
logging.info(f"save email image {filename} success")
# get all image urls
soup = BeautifulSoup(mail.body, 'html.parser')
soup = BeautifulSoup(parser.body, 'html.parser')
img_tags = soup.find_all('img')
img_urls = [img['src'] for img in img_tags if 'src' in img.attrs]
logging.info(f'Found {len(img_urls)} images in email body')
@@ -239,18 +259,28 @@ class MailStorage:
for img_url in img_urls:
# keep the original image filename(last of url)
ext = img_url.split('/')[-1].split('.')[-1]
url_result = urlparse(img_url)
if url_result.scheme not in ['http', 'https']:
continue
ext = url_result.path.split('/')[-1].split('.')[-1]
if ext in ['png', 'jpg', 'jpeg', 'gif', 'svg']:
img_filename = os.path.join(mail_dir, f"{name_count}.{ext}")
else :
img_filename = os.path.join(mail_dir, f"{name_count}")
name_count += 1
# download image
try:
response = requests.get(img_url, stream=True)
except requests.exceptions.RequestException as e:
logging.error(f'Failed to download {img_url}: {e}')
continue
if response.status_code == 200:
with open(img_filename, 'wb') as img_file:
for chunk in response.iter_content(1024):
img_file.write(chunk)
logging.info(f'Downloaded {img_url} to {img_filename}')
else:
logging.info(f'Failed to download {img_url}')
logging.error(f'Failed to download {img_url}')
cursor = self.conn.cursor()
cursor.execute(
@@ -260,5 +290,8 @@ class MailStorage:
""",
(uid, mail.id, mail.date, mail.from_addr),
)
self.conn.commit()
return mail.id
src/component/mail_environment/spider.py
+24 -5
View File
@@ -1,9 +1,13 @@
import os
import logging
import json
import string
import imaplib
import mailparser
from aios import *
from knowledge import *
from aios_kernel.storage import AIStorage
from .mail import Mail, MailStorage
class EmailSpider:
@@ -16,14 +20,22 @@ class EmailSpider:
port=self.config.get('imap_port')
)
self.client.login(self.config.get('address'), self.config.get('password'))
self.mail_local_root = os.path.join(self.env.pipeline_path, self.config.get("address"))
os.makedirs(self.mail_local_root)
self.client.select("INBOX")
local_path = string.Template(config["path"]).substitute(myai_dir=AIStorage.get_instance().get_myai_dir())
local_path = os.path.join(local_path, self.config.get('address'))
self.mail_storage = MailStorage(local_path)
async def next(self):
while True:
try:
_, data = self.client.uid('search', None, "ALL")
except Exception as e:
self.env.get_logger().error(f"email spider error: {e}")
yield (None, None)
continue
uid_list = data[0].split()
if uid_list.len() == 0:
if len(uid_list) == 0:
yield (None, None)
continue
@@ -43,9 +55,16 @@ class EmailSpider:
_uid = int.from_bytes(uid)
if _uid > from_uid:
message_parts = "(BODY.PEEK[])"
try:
_, email_data = self.client.uid('fetch', uid, message_parts)
mail = mailparser.parse_from_bytes(email_data[0][1])
self.save_email(_uid, mail)
id = self.mail_storage.download(_uid, mail)
except Exception as e:
self.env.get_logger().error(f"email spider error: {e}")
yield (None, None)
break
yield (ObjectID.from_base58(id), str(_uid))
yield (None, None)
src/component/openai_node/open_ai_node.py
+1 -1
View File
@@ -214,7 +214,7 @@ class OpenAI_ComputeNode(ComputeNode):
client = AsyncOpenAI(api_key=self.openai_api_key)
try:
if llm_inner_functions is None:
if llm_inner_functions is None or len(llm_inner_functions) == 0:
logger.info(f"call openai {mode_name} prompts: {prompts}")
resp = await client.chat.completions.create(model=mode_name,
messages=prompts,
src/component/st_node/local_st_compute_node.py
+1 -1
View File
@@ -116,7 +116,7 @@ class LocalSentenceTransformer_Image_ComputeNode(Queue_ComputeNode):
def _load_image(self, source: Union[ObjectID, bytes]) -> Optional[Image]:
image_data = None
if isinstance(source, ObjectID):
from knowledge import KnowledgeStore, ImageObject
from aios import KnowledgeStore, ImageObject
buf = KnowledgeStore().get_object_store().get_object(source)
if buf is None:
src/component/workflow_manager/workflow_manager.py
+1 -1
View File
@@ -2,7 +2,7 @@ import logging
import toml
import os
from aios import Workflow,AIStorage,PackageEnv,PackageEnvManager,PackageMediaInfo,PackageInstallTask
from aios import AIStorage,PackageEnv,PackageEnvManager,PackageMediaInfo,PackageInstallTask
from agent_manager import AgentManager
logger = logging.getLogger(__name__)
src/package-lock.json
Generated
+3
View File
@@ -0,0 +1,3 @@
{
"lockfileVersion": 1
}
src/requirements.txt
+1 -2
View File
@@ -47,7 +47,7 @@ mpmath>=1.3.0
multidict>=6.0.4
numpy>=1.25.2
onnxruntime>=1.15.1
openai>=0.28.0
openai>=1.0.0
overrides>=7.4.0
packaging>=23.1
pandas>=2.1.0
@@ -97,7 +97,6 @@ mpmath==1.3.0
multidict==6.0.4
numpy==1.25.2
onnxruntime==1.15.1
openai==0.28.0
overrides==7.4.0
packaging==23.1
pandas==2.1.0
src/service/aios_shell/aios_shell.py
+22 -15
View File
@@ -40,10 +40,11 @@ from sd_node import *
from st_node import *
from agent_manager import AgentManager
from workflow_manager import WorkflowManager
# from workflow_manager import WorkflowManager
from knowledge_manager import KnowledgePipelineManager
from tg_tunnel import TelegramTunnel
from email_tunnel import EmailTunnel
from common_environment import LocalKnowledgeBase, FilesystemEnvironment, ShellEnvironment, ScanLocalDocument, ParseLocalDocument
from compute_node_config import *
@@ -130,22 +131,26 @@ class AIOS_Shell:
cm.add_family_member(self.username,owenr)
cal_env = CalenderEnvironment("calender")
await cal_env.start()
Environment.set_env_by_id("calender",cal_env)
# cal_env = CalenderEnvironment("calender")
# await cal_env.start()
# Environment.set_env_by_id("calender",cal_env)
workspace_env = ShellEnvironment("bash")
Environment.set_env_by_id("bash",workspace_env)
# workspace_env = ShellEnvironment("bash")
# Environment.set_env_by_id("bash",workspace_env)
paint_env = PaintEnvironment("paint")
Environment.set_env_by_id("paint",paint_env)
# paint_env = PaintEnvironment("paint")
# Environment.set_env_by_id("paint",paint_env)
AgentManager.get_instance().register_environment("bash", ShellEnvironment)
AgentManager.get_instance().register_environment("fs", FilesystemEnvironment)
AgentManager.get_instance().register_environment("knowledge", LocalKnowledgeBase)
if await AgentManager.get_instance().initial() is not True:
logger.error("agent manager initial failed!")
return False
if await WorkflowManager.get_instance().initial() is not True:
logger.error("workflow manager initial failed!")
return False
# if await WorkflowManager.get_instance().initial() is not True:
# logger.error("workflow manager initial failed!")
# return False
open_ai_node = OpenAI_ComputeNode.get_instance()
if await open_ai_node.initial() is not True:
@@ -217,6 +222,8 @@ class AIOS_Shell:
pipelines = KnowledgePipelineManager.initial(os.path.join(AIStorage().get_instance().get_myai_dir(), "knowledge/pipelines"))
pipelines.register_input("scan_local", ScanLocalDocument)
pipelines.register_parser("parse_local", ParseLocalDocument)
pipelines.load_dir(os.path.join(AIStorage().get_instance().get_system_app_dir(), "knowledge_pipelines"))
pipelines.load_dir(os.path.join(AIStorage().get_instance().get_myai_dir(), "knowledge_pipelines"))
asyncio.create_task(pipelines.run())
@@ -568,8 +575,8 @@ class AIOS_Shell:
target_exist = False
if await AgentManager.get_instance().is_exist(target_id):
target_exist = True
if await WorkflowManager.get_instance().is_exist(target_id):
target_exist = True
# if await WorkflowManager.get_instance().is_exist(target_id):
# target_exist = True
if target_exist is False:
show_text = FormattedText([("class:error", f"Target {target_id} not exist!")])
@@ -627,8 +634,8 @@ class AIOS_Shell:
db_path = ""
if await self.is_agent(self.current_target):
db_path = AgentManager.get_instance().db_path
else:
db_path = WorkflowManager.get_instance().db_file
# else:
# db_path = WorkflowManager.get_instance().db_file
chatsession:AIChatSession = AIChatSession.get_session(self.current_target,f"{self.username}#{self.current_topic}",db_path,False)
if chatsession is not None:
msgs = chatsession.read_history(num,offset)