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

Fix workflow system prompt bugs.

Browse Source
This commit is contained in:
Liu Zhicong
2023-09-19 18:25:25 -07:00
parent 298ee73a6b
commit 40e14ee025
7 changed files with 132 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rootfs/workflows/math_school/workflow.toml
+7 -3
View File
@@ -20,9 +20,13 @@ agent="math_teacher"
[[roles."小学老师".prompt]]
role="system"
content="""现在时间是:{now}你在学校任职，担任小学老师。学校由 小学老师、初中老师、高中老师、教导处主任 组成。
当你发现学生的水平不是小学生时，应使用 sendmsg(老师名称,问题) 的方法，把学生的问题转发给学校里合适的老师
当学生发来作业时，进行批改(满分5分)，并把批改结果以 postmsg(教导处主任,学生名_作业结果) 的方法，将一次作业情况汇报给教导处主任。
你会根据教导处主任的指示，定期调整教学方法"""
你的任何处理结果，都要用下面方式汇报给给教导处主任
```
##/sendmsg 教导处主任
处理结果
```
"""
[roles."初中老师"]
src/aios_kernel/agent.py
+1 -1
View File
@@ -6,6 +6,7 @@ import logging
import uuid
import time
import json
import shlex
from .agent_message import AgentMsg, AgentMsgStatus, AgentMsgType
from .chatsession import AIChatSession
@@ -15,7 +16,6 @@ from .environment import Environment
logger = logging.getLogger(__name__)
class AgentPrompt:
def __init__(self) -> None:
self.messages = []
src/aios_kernel/agent_message.py
+4 -14
View File
@@ -2,6 +2,7 @@ from enum import Enum
import uuid
import time
import re
import shlex
class AgentMsgType(Enum):
TYPE_MSG = 0
@@ -123,19 +124,8 @@ class AgentMsg:
@classmethod
def parse_function_call(cls,func_string:str):
match = re.search(r'\s*(\w+)\s*\(\s*(.*)\s*\)\s*', func_string)
if not match:
return None
func_name = match.group(1)
if func_name is None:
return None
if len(func_name) < 2:
return None
params_string = match.group(2).strip()
params = re.split(r'\s*,\s*(?=(?:[^"]*"[^"]*")*[^"]*$)', params_string)
params = [param.strip('"') for param in params]
str_list = shlex.split(func_string)
func_name = str_list[0]
params = str_list[1:]
return func_name, params
src/aios_kernel/ai_function.py
+15
View File
@@ -69,6 +69,21 @@ class AIFunction:
#def load_from_config(self,config:dict) -> bool:
# pass
class FunctionItem:
def __init__(self,name,args) -> None:
self.name = name
self.args = args
self.body = None
def append_body(self,body:str) -> None:
if self.body is None:
self.body = body
else:
self.body += body
def dumps(self) -> str:
pass
# call chain is a combination of ai_function,group of ai_function.
class CallChain:
def __init__(self) -> None:
src/aios_kernel/workflow.py
+56 -40
View File
@@ -4,7 +4,7 @@ import json
import os
import time
from asyncio import Queue
from typing import Optional,Tuple
from typing import Optional,Tuple,List
from abc import ABC, abstractmethod
from .environment import Environment,EnvironmentEvent
@@ -12,7 +12,7 @@ from .agent_message import AgentMsg,AgentMsgStatus
from .agent import AgentPrompt,AgentMsg
from .chatsession import AIChatSession
from .role import AIRole,AIRoleGroup
from .ai_function import AIFunction
from .ai_function import AIFunction,FunctionItem
from .compute_kernel import ComputeKernel
from .compute_task import ComputeTask,ComputeTaskResult,ComputeTaskState
from .bus import AIBus
@@ -42,10 +42,10 @@ class LLMResult:
def __init__(self) -> None:
self.state : str = "ignore"
self.resp : str = ""
self.post_msgs = []
self.send_msgs = []
self.calls = []
self.post_calls = []
self.post_msgs : List[AgentMsg] = []
self.send_msgs : List[AgentMsg] = []
self.calls : List[FunctionItem] = []
self.post_calls : List[FunctionItem] = []
class Workflow:
@@ -255,46 +255,59 @@ class Workflow:
lines = llm_result_str.splitlines()
is_need_wait = False
for line in lines:
func_call = AgentMsg.parse_function_call(line)
if func_call:
func_args = func_call[1]
match func_call[0]:
case "sendmsg":# sendmsg($target_id,$msg_content)
if len(func_args) != 2:
def check_args(func_item:FunctionItem):
match func_name:
case "send_msg":# sendmsg($target_id,$msg_content)
if len(func_args) != 1:
logger.error(f"parse sendmsg failed! {func_call}")
continue
return False
new_msg = AgentMsg()
target_id = func_args[0]
msg_content = func_args[1]
target_id = func_item.args[0]
msg_content = func_item.body
new_msg.set("_",target_id,msg_content)
r.send_msgs.append(new_msg)
is_need_wait = True
continue
case "postmsg":# postmsg($target_id,$msg_content)
if len(func_args) != 2:
case "post_msg":# postmsg($target_id,$msg_content)
if len(func_args) != 1:
logger.error(f"parse postmsg failed! {func_call}")
continue
return False
new_msg = AgentMsg()
target_id = func_args[0]
msg_content = func_args[1]
target_id = func_item.args[0]
msg_content = func_item.body
new_msg.set("_",target_id,msg_content)
r.post_msgs.append(new_msg)
continue
case "call":# call($func_name,$args_str)
r.calls.append(func_call)
is_need_wait = True
continue
case "post_call": # post_call($func_name,$args_str)
r.post_calls.append(func_call)
continue
r.resp += line + "\n"
case "call":# call($func_name,$args_str)
r.calls.append(func_item)
is_need_wait = True
return True
case "post_call": # post_call($func_name,$args_str)
r.post_calls.append(func_item)
return True
current_func : FunctionItem = None
for line in lines:
if line.startswith("##/"):
if current_func:
if check_args(current_func) is False:
r.resp += current_func.dumps()
func_name,func_args = AgentMsg.parse_function_call(line[3:])
current_func = FunctionItem(func_name,func_args)
else:
if current_func:
current_func.append_body(line + "\n")
else:
r.resp += line + "\n"
if is_need_wait:
if current_func:
if check_args(current_func) is False:
r.resp += current_func.dumps()
if len(r.send_msgs) > 0 or len(r.calls) > 0:
r.state = "waiting"
else:
r.state = "reponsed"
@@ -339,21 +352,20 @@ class Workflow:
logger.info(f"{msg.sender} post message {msg.msg_id} to AIBus: {msg.target}")
return await self.get_bus().send_message(msg)
async def role_call(self,call:tuple,the_role:AIRole):
logger.info(f"{the_role.role_id} call {call[0]} with args {call[1]}")
func_name = call[0]
arguments = call[1]
async def role_call(self,func_item:FunctionItem,the_role:AIRole):
logger.info(f"{the_role.role_id} call {func_item.name} ")
arguments = func_item.args
func_node : AIFunction = self.workflow_env.get_ai_function(func_name)
func_node : AIFunction = self.workflow_env.get_ai_function(func_item.name)
if func_node is None:
return "execute failed,function not found"
result_str:str = await func_node.execute(**arguments)
return result_str
async def role_post_call(self,call:tuple,the_role:AIRole):
logger.info(f"{the_role.role_id} post call {call[0]} with args {call[1]}")
return await self.role_call(call,the_role)
async def role_post_call(self,func_item:FunctionItem,the_role:AIRole):
logger.info(f"{the_role.role_id} post call {func_item.name} ")
return await self.role_call(func_item,the_role)
def _format_msg_by_env_value(self,prompt:AgentPrompt):
if self.workflow_env is None:
@@ -554,3 +566,7 @@ class Workflow:
# the_env.attach_event_handler(k,_env_msg_handler)
# break
src/service/aios_shell/aios_shell.py
+1
View File
@@ -427,6 +427,7 @@ async def main():
print_formatted_text(show_text,style=shell_style)
#print_formatted_text(f"{shell.username}<->{shell.current_topic}@{shell.current_target} >>> {resp}",style=shell_style)
if __name__ == "__main__":
asyncio.run(main())
test/workflow_test.py
+36
View File
@@ -4,3 +4,39 @@ def test_workflow():
pass
async def _test_llm_parser():
test_llm_result = """
# Foggie with AI Agent
1.已经完成了基础系统改造，只要Foggie能安装docker image就可以实现集成
2.安装后，用户需要提供OpenAI Token和TG Bot Token，就可以构建自己的私有AI机器人（也可以通过绑定email实现智能邮件客服）
我们也可以用自己的OpenAI Token给用户用，但这需要设计新的商品。OpenAI Token用起来还是挺贵的
3.已在发布前夕，目前集成测试的主要问题是对Email和个人文件的AI分析需要比较强的性能。
# DMC开源挖矿软件
正在等待解决 Order Placement Issue
# Foggie with AI Agent
1. We have completed the basic system transformation. As long as Foggie can install the docker image, integration can be achieved.
2. After installation, users need to provide an OpenAI Token and TG Bot Token to build their own private AI robot. This can also be accomplished by linking an email to implement an intelligent email customer service. We could use our own OpenAI Token for users, but this would require the design of a new product. Using the OpenAI Token can be quite costly.
3. We are on the eve of launch. The main issue in our integrated testing currently is that AI analysis of emails and personal files requires substantial performance, and the results don't seem so smart.
#DMC Open Source Mining Software
We are waiting to resolve the Order Placement Issue.
##/send_msg "xxx xxx"
abcdcdsdf
sfsadfasdf
# Foggie with AI Agent
1. We have completed the basic system transformation. As long as Foggie can install the docker image, integration can be achieved.
2. After installation, users need to provide an OpenAI Token and TG Bot Token to build their own private AI robot. This can also be accomplished by linking an email to implement an intelligent email customer service. We could use our own OpenAI Token for users, but this would require the design of a new product. Using the OpenAI Token can be quite costly.
3. We are on the eve of launch. The main issue in our integrated testing cur
##/call abcd "xxx xxx"
"""
llm_result = Workflow.prase_llm_result(test_llm_result)
assert len(llm_result.calls) == 1
assert len(llm_result.send_msgs) == 1
print(llm_result)