dify/api/services/completion_service.py

507 lines
21 KiB
Python
Raw Normal View History

2023-05-15 08:51:32 +08:00
import json
import logging
import threading
import time
import uuid
from typing import Generator, Union, Any
from flask import current_app, Flask
from redis.client import PubSub
from sqlalchemy import and_
from core.completion import Completion
from core.conversation_message_task import PubHandler, ConversationTaskStoppedException
from core.llm.error import LLMBadRequestError, LLMAPIConnectionError, LLMAPIUnavailableError, LLMRateLimitError, \
LLMAuthorizationError, ProviderTokenNotInitError, QuotaExceededError, ModelCurrentlyNotSupportError
from extensions.ext_database import db
from extensions.ext_redis import redis_client
from models.model import Conversation, AppModelConfig, App, Account, EndUser, Message
from services.app_model_config_service import AppModelConfigService
from services.errors.app import MoreLikeThisDisabledError
from services.errors.app_model_config import AppModelConfigBrokenError
from services.errors.completion import CompletionStoppedError
from services.errors.conversation import ConversationNotExistsError, ConversationCompletedError
from services.errors.message import MessageNotExistsError
class CompletionService:
@classmethod
def completion(cls, app_model: App, user: Union[Account | EndUser], args: Any,
from_source: str, streaming: bool = True,
is_model_config_override: bool = False) -> Union[dict | Generator]:
# is streaming mode
inputs = args['inputs']
query = args['query']
conversation_id = args['conversation_id'] if 'conversation_id' in args else None
conversation = None
if conversation_id:
conversation_filter = [
Conversation.id == args['conversation_id'],
Conversation.app_id == app_model.id,
Conversation.status == 'normal'
]
if from_source == 'console':
conversation_filter.append(Conversation.from_account_id == user.id)
else:
conversation_filter.append(Conversation.from_end_user_id == user.id if user else None)
conversation = db.session.query(Conversation).filter(and_(*conversation_filter)).first()
if not conversation:
raise ConversationNotExistsError()
if conversation.status != 'normal':
raise ConversationCompletedError()
if not conversation.override_model_configs:
app_model_config = db.session.query(AppModelConfig).get(conversation.app_model_config_id)
if not app_model_config:
raise AppModelConfigBrokenError()
else:
conversation_override_model_configs = json.loads(conversation.override_model_configs)
app_model_config = AppModelConfig(
id=conversation.app_model_config_id,
app_id=app_model.id,
provider="",
model_id="",
configs="",
opening_statement=conversation_override_model_configs['opening_statement'],
suggested_questions=json.dumps(conversation_override_model_configs['suggested_questions']),
model=json.dumps(conversation_override_model_configs['model']),
user_input_form=json.dumps(conversation_override_model_configs['user_input_form']),
pre_prompt=conversation_override_model_configs['pre_prompt'],
agent_mode=json.dumps(conversation_override_model_configs['agent_mode']),
)
if is_model_config_override:
# build new app model config
if 'model' not in args['model_config']:
raise ValueError('model_config.model is required')
if 'completion_params' not in args['model_config']['model']:
raise ValueError('model_config.model.completion_params is required')
completion_params = AppModelConfigService.validate_model_completion_params(
cp=args['model_config']['model']['completion_params'],
model_name=app_model_config.model_dict["name"]
)
app_model_config_model = app_model_config.model_dict
app_model_config_model['completion_params'] = completion_params
app_model_config = AppModelConfig(
id=app_model_config.id,
app_id=app_model.id,
provider="",
model_id="",
configs="",
opening_statement=app_model_config.opening_statement,
suggested_questions=app_model_config.suggested_questions,
model=json.dumps(app_model_config_model),
user_input_form=app_model_config.user_input_form,
pre_prompt=app_model_config.pre_prompt,
agent_mode=app_model_config.agent_mode,
)
else:
if app_model.app_model_config_id is None:
raise AppModelConfigBrokenError()
app_model_config = app_model.app_model_config
if not app_model_config:
raise AppModelConfigBrokenError()
if is_model_config_override:
if not isinstance(user, Account):
raise Exception("Only account can override model config")
# validate config
model_config = AppModelConfigService.validate_configuration(
account=user,
config=args['model_config'],
mode=app_model.mode
)
app_model_config = AppModelConfig(
id=app_model_config.id,
app_id=app_model.id,
provider="",
model_id="",
configs="",
opening_statement=model_config['opening_statement'],
suggested_questions=json.dumps(model_config['suggested_questions']),
suggested_questions_after_answer=json.dumps(model_config['suggested_questions_after_answer']),
more_like_this=json.dumps(model_config['more_like_this']),
model=json.dumps(model_config['model']),
user_input_form=json.dumps(model_config['user_input_form']),
pre_prompt=model_config['pre_prompt'],
agent_mode=json.dumps(model_config['agent_mode']),
)
# clean input by app_model_config form rules
inputs = cls.get_cleaned_inputs(inputs, app_model_config)
generate_task_id = str(uuid.uuid4())
pubsub = redis_client.pubsub()
pubsub.subscribe(PubHandler.generate_channel_name(user, generate_task_id))
user = cls.get_real_user_instead_of_proxy_obj(user)
generate_worker_thread = threading.Thread(target=cls.generate_worker, kwargs={
'flask_app': current_app._get_current_object(),
'generate_task_id': generate_task_id,
'app_model': app_model,
'app_model_config': app_model_config,
'query': query,
'inputs': inputs,
'user': user,
'conversation': conversation,
'streaming': streaming,
'is_model_config_override': is_model_config_override
})
generate_worker_thread.start()
# wait for 5 minutes to close the thread
cls.countdown_and_close(generate_worker_thread, pubsub, user, generate_task_id)
return cls.compact_response(pubsub, streaming)
@classmethod
def get_real_user_instead_of_proxy_obj(cls, user: Union[Account, EndUser]):
if isinstance(user, Account):
user = db.session.query(Account).get(user.id)
elif isinstance(user, EndUser):
user = db.session.query(EndUser).get(user.id)
else:
raise Exception("Unknown user type")
return user
@classmethod
def generate_worker(cls, flask_app: Flask, generate_task_id: str, app_model: App, app_model_config: AppModelConfig,
query: str, inputs: dict, user: Union[Account, EndUser],
conversation: Conversation, streaming: bool, is_model_config_override: bool):
with flask_app.app_context():
try:
if conversation:
# fixed the state of the conversation object when it detached from the original session
conversation = db.session.query(Conversation).filter_by(id=conversation.id).first()
# run
Completion.generate(
task_id=generate_task_id,
app=app_model,
app_model_config=app_model_config,
query=query,
inputs=inputs,
user=user,
conversation=conversation,
streaming=streaming,
is_override=is_model_config_override,
)
except ConversationTaskStoppedException:
pass
except (LLMBadRequestError, LLMAPIConnectionError, LLMAPIUnavailableError,
LLMRateLimitError, ProviderTokenNotInitError, QuotaExceededError,
ModelCurrentlyNotSupportError) as e:
db.session.rollback()
PubHandler.pub_error(user, generate_task_id, e)
except LLMAuthorizationError:
db.session.rollback()
PubHandler.pub_error(user, generate_task_id, LLMAuthorizationError('Incorrect API key provided'))
except Exception as e:
db.session.rollback()
logging.exception("Unknown Error in completion")
PubHandler.pub_error(user, generate_task_id, e)
@classmethod
def countdown_and_close(cls, worker_thread, pubsub, user, generate_task_id) -> threading.Thread:
# wait for 5 minutes to close the thread
timeout = 300
def close_pubsub():
sleep_iterations = 0
while sleep_iterations < timeout and worker_thread.is_alive():
time.sleep(1)
sleep_iterations += 1
if worker_thread.is_alive():
PubHandler.stop(user, generate_task_id)
try:
pubsub.close()
except:
pass
countdown_thread = threading.Thread(target=close_pubsub)
countdown_thread.start()
return countdown_thread
@classmethod
def generate_more_like_this(cls, app_model: App, user: Union[Account | EndUser],
message_id: str, streaming: bool = True) -> Union[dict | Generator]:
if not user:
raise ValueError('user cannot be None')
message = db.session.query(Message).filter(
Message.id == message_id,
Message.app_id == app_model.id,
Message.from_source == ('api' if isinstance(user, EndUser) else 'console'),
Message.from_end_user_id == (user.id if isinstance(user, EndUser) else None),
Message.from_account_id == (user.id if isinstance(user, Account) else None),
).first()
if not message:
raise MessageNotExistsError()
current_app_model_config = app_model.app_model_config
more_like_this = current_app_model_config.more_like_this_dict
if not current_app_model_config.more_like_this or more_like_this.get("enabled", False) is False:
raise MoreLikeThisDisabledError()
app_model_config = message.app_model_config
if message.override_model_configs:
override_model_configs = json.loads(message.override_model_configs)
pre_prompt = override_model_configs.get("pre_prompt", '')
elif app_model_config:
pre_prompt = app_model_config.pre_prompt
else:
raise AppModelConfigBrokenError()
generate_task_id = str(uuid.uuid4())
pubsub = redis_client.pubsub()
pubsub.subscribe(PubHandler.generate_channel_name(user, generate_task_id))
user = cls.get_real_user_instead_of_proxy_obj(user)
generate_worker_thread = threading.Thread(target=cls.generate_more_like_this_worker, kwargs={
'flask_app': current_app._get_current_object(),
'generate_task_id': generate_task_id,
'app_model': app_model,
'app_model_config': app_model_config,
'message': message,
'pre_prompt': pre_prompt,
'user': user,
'streaming': streaming
})
generate_worker_thread.start()
cls.countdown_and_close(generate_worker_thread, pubsub, user, generate_task_id)
return cls.compact_response(pubsub, streaming)
@classmethod
def generate_more_like_this_worker(cls, flask_app: Flask, generate_task_id: str, app_model: App,
app_model_config: AppModelConfig, message: Message, pre_prompt: str,
user: Union[Account, EndUser], streaming: bool):
with flask_app.app_context():
try:
# run
Completion.generate_more_like_this(
task_id=generate_task_id,
app=app_model,
user=user,
message=message,
pre_prompt=pre_prompt,
app_model_config=app_model_config,
streaming=streaming
)
except ConversationTaskStoppedException:
pass
except (LLMBadRequestError, LLMAPIConnectionError, LLMAPIUnavailableError,
LLMRateLimitError, ProviderTokenNotInitError, QuotaExceededError,
ModelCurrentlyNotSupportError) as e:
db.session.rollback()
PubHandler.pub_error(user, generate_task_id, e)
except LLMAuthorizationError:
db.session.rollback()
PubHandler.pub_error(user, generate_task_id, LLMAuthorizationError('Incorrect API key provided'))
except Exception as e:
db.session.rollback()
logging.exception("Unknown Error in completion")
PubHandler.pub_error(user, generate_task_id, e)
@classmethod
def get_cleaned_inputs(cls, user_inputs: dict, app_model_config: AppModelConfig):
if user_inputs is None:
user_inputs = {}
filtered_inputs = {}
# Filter input variables from form configuration, handle required fields, default values, and option values
input_form_config = app_model_config.user_input_form_list
for config in input_form_config:
input_config = list(config.values())[0]
variable = input_config["variable"]
input_type = list(config.keys())[0]
if variable not in user_inputs or not user_inputs[variable]:
if "required" in input_config and input_config["required"]:
raise ValueError(f"{variable} is required in input form")
else:
filtered_inputs[variable] = input_config["default"] if "default" in input_config else ""
continue
value = user_inputs[variable]
if input_type == "select":
options = input_config["options"] if "options" in input_config else []
if value not in options:
raise ValueError(f"{variable} in input form must be one of the following: {options}")
else:
if 'max_length' in variable:
max_length = variable['max_length']
if len(value) > max_length:
raise ValueError(f'{variable} in input form must be less than {max_length} characters')
filtered_inputs[variable] = value
return filtered_inputs
@classmethod
def compact_response(cls, pubsub: PubSub, streaming: bool = False) -> Union[dict | Generator]:
generate_channel = list(pubsub.channels.keys())[0].decode('utf-8')
if not streaming:
try:
for message in pubsub.listen():
if message["type"] == "message":
result = message["data"].decode('utf-8')
result = json.loads(result)
if result.get('error'):
cls.handle_error(result)
return cls.get_message_response_data(result.get('data'))
except ValueError as e:
if e.args[0] != "I/O operation on closed file.": # ignore this error
raise CompletionStoppedError()
else:
logging.exception(e)
raise
finally:
try:
pubsub.unsubscribe(generate_channel)
except ConnectionError:
pass
else:
def generate() -> Generator:
try:
for message in pubsub.listen():
if message["type"] == "message":
result = message["data"].decode('utf-8')
result = json.loads(result)
if result.get('error'):
cls.handle_error(result)
event = result.get('event')
if event == "end":
logging.debug("{} finished".format(generate_channel))
break
if event == 'message':
yield "data: " + json.dumps(cls.get_message_response_data(result.get('data'))) + "\n\n"
elif event == 'chain':
yield "data: " + json.dumps(cls.get_chain_response_data(result.get('data'))) + "\n\n"
elif event == 'agent_thought':
yield "data: " + json.dumps(cls.get_agent_thought_response_data(result.get('data'))) + "\n\n"
except ValueError as e:
if e.args[0] != "I/O operation on closed file.": # ignore this error
logging.exception(e)
raise
finally:
try:
pubsub.unsubscribe(generate_channel)
except ConnectionError:
pass
return generate()
@classmethod
def get_message_response_data(cls, data: dict):
response_data = {
'event': 'message',
'task_id': data.get('task_id'),
'id': data.get('message_id'),
'answer': data.get('text'),
'created_at': int(time.time())
}
if data.get('mode') == 'chat':
response_data['conversation_id'] = data.get('conversation_id')
return response_data
@classmethod
def get_chain_response_data(cls, data: dict):
response_data = {
'event': 'chain',
'id': data.get('chain_id'),
'task_id': data.get('task_id'),
'message_id': data.get('message_id'),
'type': data.get('type'),
'input': data.get('input'),
'output': data.get('output'),
'created_at': int(time.time())
}
if data.get('mode') == 'chat':
response_data['conversation_id'] = data.get('conversation_id')
return response_data
@classmethod
def get_agent_thought_response_data(cls, data: dict):
response_data = {
'event': 'agent_thought',
'id': data.get('agent_thought_id'),
'chain_id': data.get('chain_id'),
'task_id': data.get('task_id'),
'message_id': data.get('message_id'),
'position': data.get('position'),
'thought': data.get('thought'),
'tool': data.get('tool'), # todo use real dataset obj replace it
'tool_input': data.get('tool_input'),
'observation': data.get('observation'),
'answer': data.get('answer') if not data.get('thought') else '',
'created_at': int(time.time())
}
if data.get('mode') == 'chat':
response_data['conversation_id'] = data.get('conversation_id')
return response_data
@classmethod
def handle_error(cls, result: dict):
logging.debug("error: %s", result)
error = result.get('error')
description = result.get('description')
# handle errors
llm_errors = {
'LLMBadRequestError': LLMBadRequestError,
'LLMAPIConnectionError': LLMAPIConnectionError,
'LLMAPIUnavailableError': LLMAPIUnavailableError,
'LLMRateLimitError': LLMRateLimitError,
'ProviderTokenNotInitError': ProviderTokenNotInitError,
'QuotaExceededError': QuotaExceededError,
'ModelCurrentlyNotSupportError': ModelCurrentlyNotSupportError
}
if error in llm_errors:
raise llm_errors[error](description)
elif error == 'LLMAuthorizationError':
raise LLMAuthorizationError('Incorrect API key provided')
else:
raise Exception(description)