LLMBB-Agent / workstation_server.py
ff_li
oai
e7de191
import datetime
import json
import os
import shutil
from pathlib import Path
import gradio as gr
import jsonlines
from agent.actions import (ContinueWriting, ReAct, RetrievalQA, WriteFromScratch)
from agent.actions.function_calling import FunctionCalling
from agent.llm import ChatAsOAI
from agent.log import logger
from agent.memory import Memory
from agent.tools import call_plugin, list_of_all_functions
from agent.utils.utils import (format_answer, get_last_one_line_context,
has_chinese_chars, save_text_to_file)
from utils import service, extract_and_cache_document, code_interpreter_ws, cache_root, max_ref_token, max_days, download_root
llm = ChatAsOAI(model="gpt-3.5-turbo")
mem = Memory(llm=llm, stream=False)
app_global_para = {
'time': [str(datetime.date.today()),
str(datetime.date.today())],
'cache_file': os.path.join(cache_root, 'browse.jsonl'),
'messages': [],
'last_turn_msg_id': [],
'is_first_upload': True,
}
DOC_OPTION = 'Document QA'
CI_OPTION = 'Code Interpreter'
CODE_FLAG = '/code'
PLUGIN_FLAG = '/plug'
TITLE_FLAG = '/title'
with open('css/main.css', 'r') as f:
css = f.read()
with open('js/main.js', 'r') as f:
js = f.read()
def add_text(history, text):
history = history + [(text, None)]
app_global_para['last_turn_msg_id'] = []
return history, gr.update(value='', interactive=False)
def pure_add_text(history, text):
history = history + [(text, None)]
return history, gr.update(value='', interactive=False)
def rm_text(history):
if not history:
gr.Warning('No input content!')
elif not history[-1][1]:
return history, gr.update(value='', interactive=False)
else:
history = history[:-1] + [(history[-1][0], None)]
return history, gr.update(value='', interactive=False)
def chat_clear():
app_global_para['messages'] = []
return None, None
def chat_clear_last():
for index in app_global_para['last_turn_msg_id'][::-1]:
del app_global_para['messages'][index]
app_global_para['last_turn_msg_id'] = []
def add_file(file, chosen_plug, access_token):
output_filepath = code_interpreter_ws
fn = os.path.basename(file.name)
fn_type = fn.split('.')[-1].lower()
logger.info('file type: ' + fn_type)
if chosen_plug == DOC_OPTION and (fn_type not in ['pdf', 'docx', 'pptx']):
new_path = (
'Upload failed: only adding [\'.pdf\', \'.docx\', \'.pptx\'] documents as references is supported!'
)
else:
new_path = os.path.join(output_filepath, fn)
if os.path.exists(new_path):
os.remove(new_path)
shutil.move(file.name, output_filepath)
if chosen_plug == CI_OPTION:
app_global_para['is_first_upload'] = True
# upload references
if chosen_plug == DOC_OPTION:
data = {
'content': '',
'query': '',
'url': new_path,
'task': 'cache',
'type': fn_type,
}
extract_and_cache_document(
data, app_global_para['cache_file'],
cache_root, access_token) # waiting for analyse file
return new_path
def read_records(file, times=None):
lines = []
if times:
for line in jsonlines.open(file):
if times[0] <= line['time'] <= times[1]:
lines.append(line)
return lines
def update_app_global_para(date1, date2):
app_global_para['time'][0] = date1
app_global_para['time'][1] = date2
def refresh_date():
option = [
str(datetime.date.today() - datetime.timedelta(days=i))
for i in range(max_days)
]
return (gr.update(choices=option, value=str(datetime.date.today())),
gr.update(choices=option, value=str(datetime.date.today())))
def update_browser_list():
if not os.path.exists(app_global_para['cache_file']):
return 'No browsing records'
lines = read_records(app_global_para['cache_file'],
times=app_global_para['time'])
br_list = [[line['url'], line['extract'], line['checked']]
for line in lines]
res = '<ol>{bl}</ol>'
bl = ''
for i, x in enumerate(br_list):
ck = '<input type="checkbox" class="custom-checkbox" id="ck-' + x[
0] + '" '
if x[2]:
ck += 'checked>'
else:
ck += '>'
bl += '<li>{checkbox}{title}<a href="{url}"> [url]</a></li>'.format(
checkbox=ck, url=x[0], title=x[1])
res = res.format(bl=bl)
return res
def layout_to_right(text):
return text, text
def download_text(text):
now = datetime.datetime.now()
current_time = now.strftime('%Y-%m-%d_%H-%M-%S')
filename = f'file_{current_time}.md'
save_path = os.path.join(download_root, filename)
rsp = save_text_to_file(save_path, text)
if rsp == 'SUCCESS':
gr.Info(f'Saved to {save_path}')
else:
gr.Error("Can't Save: ", rsp)
def choose_plugin(chosen_plugin):
if chosen_plugin == CI_OPTION:
gr.Info(
'Code execution is NOT sandboxed. Do NOT ask Qwen to perform dangerous tasks.'
)
if chosen_plugin == CI_OPTION or chosen_plugin == DOC_OPTION:
return gr.update(interactive=True), None
else:
return gr.update(interactive=False), None
def pure_bot(history):
if not history:
yield history
else:
history[-1][1] = ''
messages = []
for chat in history[:-1]:
messages.append({'role': 'user', 'content': chat[0]})
messages.append({'role': 'assistant', 'content': chat[1]})
messages.append({'role': 'user', 'content': history[-1][0]})
response = llm.chat(messages=messages, stream=True)
for chunk in response:
history[-1][1] += chunk
yield history
def bot(history, upload_file, chosen_plug):
if not history:
yield history
else:
history[-1][1] = ''
if chosen_plug == CI_OPTION: # use code interpreter
prompt_upload_file = ''
if upload_file and app_global_para['is_first_upload']:
workspace_dir = code_interpreter_ws
file_relpath = os.path.relpath(path=upload_file,
start=workspace_dir)
if has_chinese_chars(history[-1][0]):
prompt_upload_file = f'上传了[文件]({file_relpath})到当前目录,'
else:
prompt_upload_file = f'Uploaded the [file]({file_relpath}) to the current directory. '
app_global_para['is_first_upload'] = False
history[-1][0] = prompt_upload_file + history[-1][0]
if llm.support_function_calling():
message = {'role': 'user', 'content': history[-1][0]}
app_global_para['last_turn_msg_id'].append(
len(app_global_para['messages']))
app_global_para['messages'].append(message)
while True:
functions = [
x for x in list_of_all_functions
if x['name_for_model'] == 'code_interpreter'
]
rsp = llm.chat_with_functions(app_global_para['messages'],
functions)
if rsp.get('function_call', None):
history[-1][1] += rsp['content'].strip() + '\n'
yield history
history[-1][1] += (
'Action: ' + rsp['function_call']['name'].strip() +
'\n')
yield history
history[-1][1] += ('Action Input:\n' +
rsp['function_call']['arguments'] +
'\n')
yield history
bot_msg = {
'role': 'assistant',
'content': rsp['content'],
'function_call': {
'name': rsp['function_call']['name'],
'arguments': rsp['function_call']['arguments'],
},
}
app_global_para['last_turn_msg_id'].append(
len(app_global_para['messages']))
app_global_para['messages'].append(bot_msg)
obs = call_plugin(
rsp['function_call']['name'],
rsp['function_call']['arguments'],
)
func_msg = {
'role': 'function',
'name': rsp['function_call']['name'],
'content': obs,
}
history[-1][1] += 'Observation: ' + obs + '\n'
yield history
app_global_para['last_turn_msg_id'].append(
len(app_global_para['messages']))
app_global_para['messages'].append(func_msg)
else:
bot_msg = {
'role': 'assistant',
'content': rsp['content'],
}
# tmp_msg = '\nThought: I now know the final answer.\nFinal Answer: '
# tmp_msg += rsp['content']
# history[-1][1] += tmp_msg
history[-1][1] += rsp['content']
yield history
app_global_para['last_turn_msg_id'].append(
len(app_global_para['messages']))
app_global_para['messages'].append(bot_msg)
break
else:
functions = [
x for x in list_of_all_functions
if x['name_for_model'] == 'code_interpreter'
]
agent = ReAct(llm=llm)
for chunk in agent.run(user_request=history[-1][0],
functions=functions,
history=app_global_para['messages']):
history[-1][1] += chunk
yield history
yield history
message = {'role': 'user', 'content': history[-1][0]}
app_global_para['last_turn_msg_id'].append(
len(app_global_para['messages']))
app_global_para['messages'].append(message)
rsp_message = {'role': 'assistant', 'content': history[-1][1]}
app_global_para['last_turn_msg_id'].append(
len(app_global_para['messages']))
app_global_para['messages'].append(rsp_message)
else:
lines = []
if not os.path.exists(app_global_para['cache_file']):
_ref = ''
else:
for line in jsonlines.open(app_global_para['cache_file']):
if (app_global_para['time'][0] <= line['time'] <=
app_global_para['time'][1]) and line['checked']:
lines.append(line)
if lines:
_ref_list = mem.get(
history[-1][0],
lines,
max_token=max_ref_token)
_ref = '\n'.join(
json.dumps(x, ensure_ascii=False) for x in _ref_list)
else:
_ref = ''
gr.Warning(
'No reference materials selected, Qwen will answer directly'
)
logger.info(_ref)
# TODO: considering history for retrieval qa
agent = RetrievalQA(llm=llm, stream=True)
response = agent.run(user_request=history[-1][0], ref_doc=_ref)
for chunk in response:
history[-1][1] += chunk
yield history
# append message
message = {'role': 'user', 'content': history[-1][0]}
app_global_para['last_turn_msg_id'].append(
len(app_global_para['messages']))
app_global_para['messages'].append(message)
message = {'role': 'assistant', 'content': history[-1][1]}
app_global_para['last_turn_msg_id'].append(
len(app_global_para['messages']))
app_global_para['messages'].append(message)
def generate(context):
sp_query = get_last_one_line_context(context)
res = ''
if CODE_FLAG in sp_query: # router to code interpreter
sp_query = sp_query.split(CODE_FLAG)[-1]
if has_chinese_chars(sp_query):
sp_query += ', 必须使用code_interpreter工具'
else:
sp_query += ' (Please use code_interpreter.)'
functions = [
x for x in list_of_all_functions
if x['name_for_model'] == 'code_interpreter'
]
if llm.support_function_calling():
response = FunctionCalling(llm=llm).run(sp_query,
functions=functions)
for chunk in response:
res += chunk
yield res
else:
agent = ReAct(llm=llm)
for chunk in agent.run(user_request=sp_query, functions=functions):
res += chunk
yield res
yield res
elif PLUGIN_FLAG in sp_query: # router to plugin
sp_query = sp_query.split(PLUGIN_FLAG)[-1]
functions = list_of_all_functions
if llm.support_function_calling():
response = FunctionCalling(llm=llm).run(sp_query,
functions=functions)
for chunk in response:
res += chunk
yield res
else:
agent = ReAct(llm=llm)
for chunk in agent.run(user_request=sp_query, functions=functions):
res += chunk
yield res
yield res
else: # router to continue writing
lines = []
if os.path.exists(app_global_para['cache_file']):
for line in jsonlines.open(app_global_para['cache_file']):
if (app_global_para['time'][0] <= line['time'] <=
app_global_para['time'][1]) and line['checked']:
lines.append(line)
if lines:
res += '\n========================= \n'
yield res
res += '> Search for relevant information: \n'
yield res
sp_query_no_title = sp_query
if TITLE_FLAG in sp_query: # /title
sp_query_no_title = sp_query.split(TITLE_FLAG)[-1]
_ref_list = mem.get(sp_query_no_title,
lines,
max_token=max_ref_token)
_ref = '\n'.join(
json.dumps(x, ensure_ascii=False) for x in _ref_list)
res += _ref
yield res
res += '\n'
else:
_ref = ''
gr.Warning(
'No reference materials selected, Qwen will answer directly')
if TITLE_FLAG in sp_query: # /title
agent = WriteFromScratch(llm=llm, stream=True)
user_request = sp_query.split(TITLE_FLAG)[-1]
else:
res += '\n========================= \n'
res += '> Writing Text: \n'
yield res
agent = ContinueWriting(llm=llm, stream=True)
user_request = context
response = agent.run(user_request=user_request, ref_doc=_ref)
for chunk in response:
res += chunk
yield res
def format_generate(edit, context):
res = edit
yield res
if '> Writing Text: ' in context:
text = context.split('> Writing Text: ')[-1].strip()
res += '\n'
res += text
yield res
elif 'Final Answer' in context:
response = format_answer(context)
res += '\n'
res += response
yield res
else:
res += context
yield res
def initialize(request: gr.Request, date1, date2):
access_token = request.query_params["access_token"] or request.session["access_token"]
is_valid = False
if access_token:
account_info = json.loads(service.get(access_token, "info.json", False))
if account_info and account_info["enabled"]:
is_valid = True
if not is_valid:
gr.Info("The token is not valid, Please reset!")
return
update_app_global_para(date1, date2)
return access_token
with gr.Blocks(css=css, theme='soft') as demo:
access_token = gr.State("")
title = gr.Markdown('Qwen Agent: BrowserQwen', elem_classes='title')
desc = gr.Markdown(
'This is the editing workstation of BrowserQwen, where Qwen has collected the browsing history. Qwen can assist you in completing your creative work!',
elem_classes='desc',
)
with gr.Row():
with gr.Column():
rec = gr.Markdown('Browsing History', elem_classes='rec')
with gr.Row():
with gr.Column(scale=3, min_width=0):
date1 = gr.Dropdown(
[
str(datetime.date.today() -
datetime.timedelta(days=i))
for i in range(max_days)
],
value=str(datetime.date.today()),
label='Start Date',
)
date2 = gr.Dropdown(
[
str(datetime.date.today() -
datetime.timedelta(days=i))
for i in range(max_days)
],
value=str(datetime.date.today()),
label='End Date',
)
with gr.Column(scale=7, min_width=0):
browser_list = gr.HTML(
value='',
label='browser_list',
elem_classes=['div_tmp', 'add_scrollbar'],
)
with gr.Tab('Editor', elem_id='default-tab'):
with gr.Row():
with gr.Column():
with gr.Row():
edit_area = gr.Textbox(
value='',
elem_classes=['textbox_default', 'add_scrollbar'],
lines=30,
label='Input',
show_copy_button=True,
)
# token_count = gr.HTML(value='<span>0</span>',
# elem_classes=[
# 'token-counter',
# 'default-token-counter'
# ])
with gr.Row():
ctn_bt = gr.Button('Continue', variant='primary')
stop_bt = gr.Button('Stop')
clr_bt = gr.Button('Clear')
dld_bt = gr.Button('Download')
# with gr.Row():
# layout_bt = gr.Button('👉', variant='primary')
with gr.Column():
cmd_area = gr.Textbox(lines=10,
max_lines=10,
label="Qwen's Inner Thought",
elem_id='cmd')
with gr.Tab('Markdown'):
# md_out_bt = gr.Button('Render')
md_out_area = gr.Markdown(
elem_classes=['md_tmp', 'add_scrollbar'])
with gr.Tab('HTML'):
html_out_area = gr.HTML()
with gr.Tab('Raw'):
text_out_area = gr.Textbox(
lines=20,
label='',
elem_classes=[
'textbox_default_output', 'add_scrollbar'
],
show_copy_button=True,
)
clk_ctn_bt = ctn_bt.click(generate, edit_area, cmd_area)
clk_ctn_bt.then(format_generate, [edit_area, cmd_area], edit_area)
edit_area_change = edit_area.change(layout_to_right, edit_area,
[text_out_area, md_out_area])
stop_bt.click(lambda: None, cancels=[clk_ctn_bt], queue=False)
clr_bt.click(
lambda: [None, None, None],
None,
[edit_area, cmd_area, md_out_area],
queue=False,
)
dld_bt.click(download_text, edit_area, None)
# layout_bt.click(layout_to_right,
# edit_area, [text_out_area, md_out_area],
# queue=False)
gr.Markdown("""
### Usage Tips:
- Browsing History:
- Start Date/End Date: Selecting the browsed materials for the desired time period, including the start and end dates
- The browsed materials list: supporting the selection or removal of specific browsing content
- Editor: In the editing area, you can directly input content or special instructions, and then click the ```Continue``` button to have Qwen assist in completing the editing work:
- After inputting the content, directly click the ```Continue``` button: Qwen will begin to continue writing based on the browsing information
- Using special instructions:
- /title + content: Qwen enables the built-in planning process and writes a complete manuscript
- /code + content: Qwen enables the code interpreter plugin, writes and runs Python code, and generates replies
- /plug + content: Qwen enables plugin and select appropriate plugin to generate reply
- Chat: Interactive area. Qwen generates replies based on given reference materials. Selecting Code Interpreter will enable the code interpreter plugin
""")
with gr.Tab('Chat', elem_id='chat-tab'):
with gr.Column():
chatbot = gr.Chatbot(
[],
elem_id='chatbot',
height=680,
show_copy_button=True,
avatar_images=(
None,
(os.path.join(
Path(__file__).resolve().parent, 'img/logo.png')),
),
)
with gr.Row():
with gr.Column(scale=1, min_width=0):
file_btn = gr.UploadButton('Upload', file_types=['file'])
with gr.Column(scale=13):
chat_txt = gr.Textbox(
show_label=False,
placeholder='Chat with Qwen...',
container=False,
)
with gr.Column(scale=1, min_width=0):
chat_clr_bt = gr.Button('Clear')
with gr.Column(scale=1, min_width=0):
chat_stop_bt = gr.Button('Stop')
with gr.Column(scale=1, min_width=0):
chat_re_bt = gr.Button('Again')
with gr.Row():
with gr.Column(scale=2, min_width=0):
plug_bt = gr.Dropdown(
[CI_OPTION, DOC_OPTION],
label='Plugin',
info='',
value=DOC_OPTION,
)
with gr.Column(scale=8, min_width=0):
hidden_file_path = gr.Textbox(
interactive=False,
label='The uploaded file is displayed here')
txt_msg = chat_txt.submit(
add_text, [chatbot, chat_txt], [chatbot, chat_txt],
queue=False).then(bot, [chatbot, hidden_file_path, plug_bt],
chatbot)
txt_msg.then(lambda: gr.update(interactive=True),
None, [chat_txt],
queue=False)
# txt_msg_bt = chat_smt_bt.click(add_text, [chatbot, chat_txt], [chatbot, chat_txt], queue=False).then(bot, chatbot, chatbot)
# txt_msg_bt.then(lambda: gr.update(interactive=True), None, [chat_txt], queue=False)
# (None, None, None, cancels=[txt_msg], queue=False).then
re_txt_msg = (chat_re_bt.click(
rm_text, [chatbot], [chatbot, chat_txt],
queue=False).then(chat_clear_last, None, None).then(
bot, [chatbot, hidden_file_path, plug_bt], chatbot))
re_txt_msg.then(lambda: gr.update(interactive=True),
None, [chat_txt],
queue=False)
file_msg = file_btn.upload(add_file, [file_btn, plug_bt, access_token],
[hidden_file_path],
queue=False)
file_msg.then(update_browser_list, None,
browser_list).then(lambda: None,
None,
None,
_js=f'() => {{{js}}}')
chat_clr_bt.click(chat_clear,
None, [chatbot, hidden_file_path],
queue=False)
# re_bt.click(re_bot, chatbot, chatbot)
chat_stop_bt.click(chat_clear_last,
None,
None,
cancels=[txt_msg, re_txt_msg],
queue=False)
plug_bt.change(choose_plugin, plug_bt,
[file_btn, hidden_file_path])
with gr.Tab('Pure Chat', elem_id='pure-chat-tab'):
gr.Markdown(
'Note: The chat box on this tab will not use any browsing history!'
)
with gr.Column():
pure_chatbot = gr.Chatbot(
[],
elem_id='pure_chatbot',
height=680,
show_copy_button=True,
avatar_images=(
None,
(os.path.join(
Path(__file__).resolve().parent, 'img/logo.png')),
),
)
with gr.Row():
with gr.Column(scale=13):
chat_txt = gr.Textbox(
show_label=False,
placeholder='Chat with Qwen...',
container=False,
)
with gr.Column(scale=1, min_width=0):
chat_clr_bt = gr.Button('Clear')
with gr.Column(scale=1, min_width=0):
chat_stop_bt = gr.Button('Stop')
with gr.Column(scale=1, min_width=0):
chat_re_bt = gr.Button('Again')
txt_msg = chat_txt.submit(pure_add_text, [pure_chatbot, chat_txt],
[pure_chatbot, chat_txt],
queue=False).then(
pure_bot, pure_chatbot, pure_chatbot)
txt_msg.then(lambda: gr.update(interactive=True),
None, [chat_txt],
queue=False)
re_txt_msg = chat_re_bt.click(rm_text, [pure_chatbot],
[pure_chatbot, chat_txt],
queue=False).then(
pure_bot, pure_chatbot,
pure_chatbot)
re_txt_msg.then(lambda: gr.update(interactive=True),
None, [chat_txt],
queue=False)
chat_clr_bt.click(lambda: None, None, pure_chatbot, queue=False)
chat_stop_bt.click(chat_clear_last,
None,
None,
cancels=[txt_msg, re_txt_msg],
queue=False)
date1.change(update_app_global_para, [date1, date2],
None).then(update_browser_list, None,
browser_list).then(lambda: None,
None,
None,
_js=f'() => {{{js}}}').then(
chat_clear, None,
[chatbot, hidden_file_path])
date2.change(update_app_global_para, [date1, date2],
None).then(update_browser_list, None,
browser_list).then(lambda: None,
None,
None,
_js=f'() => {{{js}}}').then(
chat_clear, None,
[chatbot, hidden_file_path])
demo.load(initialize, [date1, date2],
[access_token]).then(refresh_date, None, [date1, date2]).then(
update_browser_list, None,
browser_list).then(lambda: None,
None,
None,
_js=f'() => {{{js}}}').then(
chat_clear, None,
[chatbot, hidden_file_path])
demo.queue()
# demo.queue().launch(server_name=server_host, server_port=workstation_port)