estel/dify
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Feat/assistant app (#2086)

Browse Source 
Co-authored-by: chenhe <guchenhe@gmail.com>
Co-authored-by: Pascal M <11357019+perzeuss@users.noreply.github.com>
This commit is contained in:
Yeuoly
2024-01-23 19:58:23 +08:00
committed by GitHub
parent 7bbe12b2bd
commit 86286e1ac8
175 changed files with 11619 additions and 1235 deletions
Download Patch File Download Diff File Expand all files Collapse all files

251
api/core/app_runner/agent_app_runner.py
View File

@@ -1,251 +0,0 @@
import json
import logging
from typing import cast
from core.agent.agent.agent_llm_callback import AgentLLMCallback
from core.app_runner.app_runner import AppRunner
from core.application_queue_manager import ApplicationQueueManager
from core.callback_handler.agent_loop_gather_callback_handler import AgentLoopGatherCallbackHandler
from core.entities.application_entities import ApplicationGenerateEntity, ModelConfigEntity, PromptTemplateEntity
from core.features.agent_runner import AgentRunnerFeature
from core.memory.token_buffer_memory import TokenBufferMemory
from core.model_manager import ModelInstance
from core.model_runtime.entities.llm_entities import LLMUsage
from core.model_runtime.model_providers.__base.large_language_model import LargeLanguageModel
from extensions.ext_database import db
from models.model import App, Conversation, Message, MessageAgentThought, MessageChain
logger = logging.getLogger(__name__)
class AgentApplicationRunner(AppRunner):
"""
Agent Application Runner
"""
def run(self, application_generate_entity: ApplicationGenerateEntity,
queue_manager: ApplicationQueueManager,
conversation: Conversation,
message: Message) -> None:
"""
Run agent application
:param application_generate_entity: application generate entity
:param queue_manager: application queue manager
:param conversation: conversation
:param message: message
:return:
"""
app_record = db.session.query(App).filter(App.id == application_generate_entity.app_id).first()
if not app_record:
raise ValueError(f"App not found")
app_orchestration_config = application_generate_entity.app_orchestration_config_entity
inputs = application_generate_entity.inputs
query = application_generate_entity.query
files = application_generate_entity.files
# Pre-calculate the number of tokens of the prompt messages,
# and return the rest number of tokens by model context token size limit and max token size limit.
# If the rest number of tokens is not enough, raise exception.
# Include: prompt template, inputs, query(optional), files(optional)
# Not Include: memory, external data, dataset context
self.get_pre_calculate_rest_tokens(
app_record=app_record,
model_config=app_orchestration_config.model_config,
prompt_template_entity=app_orchestration_config.prompt_template,
inputs=inputs,
files=files,
query=query
)
memory = None
if application_generate_entity.conversation_id:
# get memory of conversation (read-only)
model_instance = ModelInstance(
provider_model_bundle=app_orchestration_config.model_config.provider_model_bundle,
model=app_orchestration_config.model_config.model
)
memory = TokenBufferMemory(
conversation=conversation,
model_instance=model_instance
)
# reorganize all inputs and template to prompt messages
# Include: prompt template, inputs, query(optional), files(optional)
# memory(optional)
prompt_messages, stop = self.organize_prompt_messages(
app_record=app_record,
model_config=app_orchestration_config.model_config,
prompt_template_entity=app_orchestration_config.prompt_template,
inputs=inputs,
files=files,
query=query,
context=None,
memory=memory
)
# Create MessageChain
message_chain = self._init_message_chain(
message=message,
query=query
)
# add agent callback to record agent thoughts
agent_callback = AgentLoopGatherCallbackHandler(
model_config=app_orchestration_config.model_config,
message=message,
queue_manager=queue_manager,
message_chain=message_chain
)
# init LLM Callback
agent_llm_callback = AgentLLMCallback(
agent_callback=agent_callback
)
agent_runner = AgentRunnerFeature(
tenant_id=application_generate_entity.tenant_id,
app_orchestration_config=app_orchestration_config,
model_config=app_orchestration_config.model_config,
config=app_orchestration_config.agent,
queue_manager=queue_manager,
message=message,
user_id=application_generate_entity.user_id,
agent_llm_callback=agent_llm_callback,
callback=agent_callback,
memory=memory
)
# agent run
result = agent_runner.run(
query=query,
invoke_from=application_generate_entity.invoke_from
)
if result:
self._save_message_chain(
message_chain=message_chain,
output_text=result
)
if (result
and app_orchestration_config.prompt_template.prompt_type == PromptTemplateEntity.PromptType.SIMPLE
and app_orchestration_config.prompt_template.simple_prompt_template
):
# Direct output if agent result exists and has pre prompt
self.direct_output(
queue_manager=queue_manager,
app_orchestration_config=app_orchestration_config,
prompt_messages=prompt_messages,
stream=application_generate_entity.stream,
text=result,
usage=self._get_usage_of_all_agent_thoughts(
model_config=app_orchestration_config.model_config,
message=message
)
)
else:
# As normal LLM run, agent result as context
context = result
# reorganize all inputs and template to prompt messages
# Include: prompt template, inputs, query(optional), files(optional)
# memory(optional), external data, dataset context(optional)
prompt_messages, stop = self.organize_prompt_messages(
app_record=app_record,
model_config=app_orchestration_config.model_config,
prompt_template_entity=app_orchestration_config.prompt_template,
inputs=inputs,
files=files,
query=query,
context=context,
memory=memory
)
# Re-calculate the max tokens if sum(prompt_token + max_tokens) over model token limit
self.recale_llm_max_tokens(
model_config=app_orchestration_config.model_config,
prompt_messages=prompt_messages
)
# Invoke model
model_instance = ModelInstance(
provider_model_bundle=app_orchestration_config.model_config.provider_model_bundle,
model=app_orchestration_config.model_config.model
)
invoke_result = model_instance.invoke_llm(
prompt_messages=prompt_messages,
model_parameters=app_orchestration_config.model_config.parameters,
stop=stop,
stream=application_generate_entity.stream,
user=application_generate_entity.user_id,
)
# handle invoke result
self._handle_invoke_result(
invoke_result=invoke_result,
queue_manager=queue_manager,
stream=application_generate_entity.stream
)
def _init_message_chain(self, message: Message, query: str) -> MessageChain:
"""
Init MessageChain
:param message: message
:param query: query
:return:
"""
message_chain = MessageChain(
message_id=message.id,
type="AgentExecutor",
input=json.dumps({
"input": query
})
)
db.session.add(message_chain)
db.session.commit()
return message_chain
def _save_message_chain(self, message_chain: MessageChain, output_text: str) -> None:
"""
Save MessageChain
:param message_chain: message chain
:param output_text: output text
:return:
"""
message_chain.output = json.dumps({
"output": output_text
})
db.session.commit()
def _get_usage_of_all_agent_thoughts(self, model_config: ModelConfigEntity,
message: Message) -> LLMUsage:
"""
Get usage of all agent thoughts
:param model_config: model config
:param message: message
:return:
"""
agent_thoughts = (db.session.query(MessageAgentThought)
.filter(MessageAgentThought.message_id == message.id).all())
all_message_tokens = 0
all_answer_tokens = 0
for agent_thought in agent_thoughts:
all_message_tokens += agent_thought.message_token
all_answer_tokens += agent_thought.answer_token
model_type_instance = model_config.provider_model_bundle.model_type_instance
model_type_instance = cast(LargeLanguageModel, model_type_instance)
return model_type_instance._calc_response_usage(
model_config.model,
model_config.credentials,
all_message_tokens,
all_answer_tokens
)

128
api/core/app_runner/app_runner.py
View File

@@ -2,7 +2,8 @@ import time
from typing import Generator, List, Optional, Tuple, Union, cast
from core.application_queue_manager import ApplicationQueueManager, PublishFrom
from core.entities.application_entities import AppOrchestrationConfigEntity, ModelConfigEntity, PromptTemplateEntity
from core.entities.application_entities import AppOrchestrationConfigEntity, ModelConfigEntity, \
PromptTemplateEntity, ExternalDataVariableEntity, ApplicationGenerateEntity, InvokeFrom
from core.file.file_obj import FileObj
from core.memory.token_buffer_memory import TokenBufferMemory
from core.model_runtime.entities.llm_entities import LLMResult, LLMResultChunk, LLMResultChunkDelta, LLMUsage
@@ -10,9 +11,12 @@ from core.model_runtime.entities.message_entities import AssistantPromptMessage,
from core.model_runtime.entities.model_entities import ModelPropertyKey
from core.model_runtime.errors.invoke import InvokeBadRequestError
from core.model_runtime.model_providers.__base.large_language_model import LargeLanguageModel
from core.features.hosting_moderation import HostingModerationFeature
from core.features.moderation import ModerationFeature
from core.features.external_data_fetch import ExternalDataFetchFeature
from core.features.annotation_reply import AnnotationReplyFeature
from core.prompt.prompt_transform import PromptTransform
from models.model import App
from models.model import App, MessageAnnotation, Message
class AppRunner:
def get_pre_calculate_rest_tokens(self, app_record: App,
@@ -199,7 +203,8 @@ class AppRunner:
def _handle_invoke_result(self, invoke_result: Union[LLMResult, Generator],
queue_manager: ApplicationQueueManager,
stream: bool) -> None:
stream: bool,
agent: bool = False) -> None:
"""
Handle invoke result
:param invoke_result: invoke result
@@ -210,16 +215,19 @@ class AppRunner:
if not stream:
self._handle_invoke_result_direct(
invoke_result=invoke_result,
queue_manager=queue_manager
queue_manager=queue_manager,
agent=agent
)
else:
self._handle_invoke_result_stream(
invoke_result=invoke_result,
queue_manager=queue_manager
queue_manager=queue_manager,
agent=agent
)
def _handle_invoke_result_direct(self, invoke_result: LLMResult,
queue_manager: ApplicationQueueManager) -> None:
queue_manager: ApplicationQueueManager,
agent: bool) -> None:
"""
Handle invoke result direct
:param invoke_result: invoke result
@@ -232,7 +240,8 @@ class AppRunner:
)
def _handle_invoke_result_stream(self, invoke_result: Generator,
queue_manager: ApplicationQueueManager) -> None:
queue_manager: ApplicationQueueManager,
agent: bool) -> None:
"""
Handle invoke result
:param invoke_result: invoke result
@@ -244,7 +253,10 @@ class AppRunner:
text = ''
usage = None
for result in invoke_result:
queue_manager.publish_chunk_message(result, PublishFrom.APPLICATION_MANAGER)
if not agent:
queue_manager.publish_chunk_message(result, PublishFrom.APPLICATION_MANAGER)
else:
queue_manager.publish_agent_chunk_message(result, PublishFrom.APPLICATION_MANAGER)
text += result.delta.message.content
@@ -271,3 +283,101 @@ class AppRunner:
llm_result=llm_result,
pub_from=PublishFrom.APPLICATION_MANAGER
)
def moderation_for_inputs(self, app_id: str,
tenant_id: str,
app_orchestration_config_entity: AppOrchestrationConfigEntity,
inputs: dict,
query: str) -> Tuple[bool, dict, str]:
"""
Process sensitive_word_avoidance.
:param app_id: app id
:param tenant_id: tenant id
:param app_orchestration_config_entity: app orchestration config entity
:param inputs: inputs
:param query: query
:return:
"""
moderation_feature = ModerationFeature()
return moderation_feature.check(
app_id=app_id,
tenant_id=tenant_id,
app_orchestration_config_entity=app_orchestration_config_entity,
inputs=inputs,
query=query,
)
def check_hosting_moderation(self, application_generate_entity: ApplicationGenerateEntity,
queue_manager: ApplicationQueueManager,
prompt_messages: list[PromptMessage]) -> bool:
"""
Check hosting moderation
:param application_generate_entity: application generate entity
:param queue_manager: queue manager
:param prompt_messages: prompt messages
:return:
"""
hosting_moderation_feature = HostingModerationFeature()
moderation_result = hosting_moderation_feature.check(
application_generate_entity=application_generate_entity,
prompt_messages=prompt_messages
)
if moderation_result:
self.direct_output(
queue_manager=queue_manager,
app_orchestration_config=application_generate_entity.app_orchestration_config_entity,
prompt_messages=prompt_messages,
text="I apologize for any confusion, " \
"but I'm an AI assistant to be helpful, harmless, and honest.",
stream=application_generate_entity.stream
)
return moderation_result
def fill_in_inputs_from_external_data_tools(self, tenant_id: str,
app_id: str,
external_data_tools: list[ExternalDataVariableEntity],
inputs: dict,
query: str) -> dict:
"""
Fill in variable inputs from external data tools if exists.
:param tenant_id: workspace id
:param app_id: app id
:param external_data_tools: external data tools configs
:param inputs: the inputs
:param query: the query
:return: the filled inputs
"""
external_data_fetch_feature = ExternalDataFetchFeature()
return external_data_fetch_feature.fetch(
tenant_id=tenant_id,
app_id=app_id,
external_data_tools=external_data_tools,
inputs=inputs,
query=query
)
def query_app_annotations_to_reply(self, app_record: App,
message: Message,
query: str,
user_id: str,
invoke_from: InvokeFrom) -> Optional[MessageAnnotation]:
"""
Query app annotations to reply
:param app_record: app record
:param message: message
:param query: query
:param user_id: user id
:param invoke_from: invoke from
:return:
"""
annotation_reply_feature = AnnotationReplyFeature()
return annotation_reply_feature.query(
app_record=app_record,
message=message,
query=query,
user_id=user_id,
invoke_from=invoke_from
)

342
api/core/app_runner/assistant_app_runner.py Normal file
View File

@@ -0,0 +1,342 @@
import json
import logging
from typing import cast
from core.app_runner.app_runner import AppRunner
from core.features.assistant_cot_runner import AssistantCotApplicationRunner
from core.features.assistant_fc_runner import AssistantFunctionCallApplicationRunner
from core.entities.application_entities import ApplicationGenerateEntity, ModelConfigEntity, \
AgentEntity
from core.application_queue_manager import ApplicationQueueManager, PublishFrom
from core.memory.token_buffer_memory import TokenBufferMemory
from core.model_manager import ModelInstance
from core.model_runtime.entities.llm_entities import LLMUsage
from core.model_runtime.model_providers.__base.large_language_model import LargeLanguageModel
from core.moderation.base import ModerationException
from core.tools.entities.tool_entities import ToolRuntimeVariablePool
from extensions.ext_database import db
from models.model import Conversation, Message, App, MessageChain, MessageAgentThought
from models.tools import ToolConversationVariables
logger = logging.getLogger(__name__)
class AssistantApplicationRunner(AppRunner):
"""
Assistant Application Runner
"""
def run(self, application_generate_entity: ApplicationGenerateEntity,
queue_manager: ApplicationQueueManager,
conversation: Conversation,
message: Message) -> None:
"""
Run assistant application
:param application_generate_entity: application generate entity
:param queue_manager: application queue manager
:param conversation: conversation
:param message: message
:return:
"""
app_record = db.session.query(App).filter(App.id == application_generate_entity.app_id).first()
if not app_record:
raise ValueError(f"App not found")
app_orchestration_config = application_generate_entity.app_orchestration_config_entity
inputs = application_generate_entity.inputs
query = application_generate_entity.query
files = application_generate_entity.files
# Pre-calculate the number of tokens of the prompt messages,
# and return the rest number of tokens by model context token size limit and max token size limit.
# If the rest number of tokens is not enough, raise exception.
# Include: prompt template, inputs, query(optional), files(optional)
# Not Include: memory, external data, dataset context
self.get_pre_calculate_rest_tokens(
app_record=app_record,
model_config=app_orchestration_config.model_config,
prompt_template_entity=app_orchestration_config.prompt_template,
inputs=inputs,
files=files,
query=query
)
memory = None
if application_generate_entity.conversation_id:
# get memory of conversation (read-only)
model_instance = ModelInstance(
provider_model_bundle=app_orchestration_config.model_config.provider_model_bundle,
model=app_orchestration_config.model_config.model
)
memory = TokenBufferMemory(
conversation=conversation,
model_instance=model_instance
)
# organize all inputs and template to prompt messages
# Include: prompt template, inputs, query(optional), files(optional)
# memory(optional)
prompt_messages, _ = self.organize_prompt_messages(
app_record=app_record,
model_config=app_orchestration_config.model_config,
prompt_template_entity=app_orchestration_config.prompt_template,
inputs=inputs,
files=files,
query=query,
memory=memory
)
# moderation
try:
# process sensitive_word_avoidance
_, inputs, query = self.moderation_for_inputs(
app_id=app_record.id,
tenant_id=application_generate_entity.tenant_id,
app_orchestration_config_entity=app_orchestration_config,
inputs=inputs,
query=query,
)
except ModerationException as e:
self.direct_output(
queue_manager=queue_manager,
app_orchestration_config=app_orchestration_config,
prompt_messages=prompt_messages,
text=str(e),
stream=application_generate_entity.stream
)
return
if query:
# annotation reply
annotation_reply = self.query_app_annotations_to_reply(
app_record=app_record,
message=message,
query=query,
user_id=application_generate_entity.user_id,
invoke_from=application_generate_entity.invoke_from
)
if annotation_reply:
queue_manager.publish_annotation_reply(
message_annotation_id=annotation_reply.id,
pub_from=PublishFrom.APPLICATION_MANAGER
)
self.direct_output(
queue_manager=queue_manager,
app_orchestration_config=app_orchestration_config,
prompt_messages=prompt_messages,
text=annotation_reply.content,
stream=application_generate_entity.stream
)
return
# fill in variable inputs from external data tools if exists
external_data_tools = app_orchestration_config.external_data_variables
if external_data_tools:
inputs = self.fill_in_inputs_from_external_data_tools(
tenant_id=app_record.tenant_id,
app_id=app_record.id,
external_data_tools=external_data_tools,
inputs=inputs,
query=query
)
# reorganize all inputs and template to prompt messages
# Include: prompt template, inputs, query(optional), files(optional)
# memory(optional), external data, dataset context(optional)
prompt_messages, _ = self.organize_prompt_messages(
app_record=app_record,
model_config=app_orchestration_config.model_config,
prompt_template_entity=app_orchestration_config.prompt_template,
inputs=inputs,
files=files,
query=query,
memory=memory
)
# check hosting moderation
hosting_moderation_result = self.check_hosting_moderation(
application_generate_entity=application_generate_entity,
queue_manager=queue_manager,
prompt_messages=prompt_messages
)
if hosting_moderation_result:
return
agent_entity = app_orchestration_config.agent
# load tool variables
tool_conversation_variables = self._load_tool_variables(conversation_id=conversation.id,
user_id=application_generate_entity.user_id,
tanent_id=application_generate_entity.tenant_id)
# convert db variables to tool variables
tool_variables = self._convert_db_variables_to_tool_variables(tool_conversation_variables)
message_chain = self._init_message_chain(
message=message,
query=query
)
# init model instance
model_instance = ModelInstance(
provider_model_bundle=app_orchestration_config.model_config.provider_model_bundle,
model=app_orchestration_config.model_config.model
)
prompt_message, _ = self.organize_prompt_messages(
app_record=app_record,
model_config=app_orchestration_config.model_config,
prompt_template_entity=app_orchestration_config.prompt_template,
inputs=inputs,
files=files,
query=query,
memory=memory,
)
# start agent runner
if agent_entity.strategy == AgentEntity.Strategy.CHAIN_OF_THOUGHT:
assistant_cot_runner = AssistantCotApplicationRunner(
tenant_id=application_generate_entity.tenant_id,
application_generate_entity=application_generate_entity,
app_orchestration_config=app_orchestration_config,
model_config=app_orchestration_config.model_config,
config=agent_entity,
queue_manager=queue_manager,
message=message,
user_id=application_generate_entity.user_id,
memory=memory,
prompt_messages=prompt_message,
variables_pool=tool_variables,
db_variables=tool_conversation_variables,
)
invoke_result = assistant_cot_runner.run(
model_instance=model_instance,
conversation=conversation,
message=message,
query=query,
)
elif agent_entity.strategy == AgentEntity.Strategy.FUNCTION_CALLING:
assistant_cot_runner = AssistantFunctionCallApplicationRunner(
tenant_id=application_generate_entity.tenant_id,
application_generate_entity=application_generate_entity,
app_orchestration_config=app_orchestration_config,
model_config=app_orchestration_config.model_config,
config=agent_entity,
queue_manager=queue_manager,
message=message,
user_id=application_generate_entity.user_id,
memory=memory,
prompt_messages=prompt_message,
variables_pool=tool_variables,
db_variables=tool_conversation_variables
)
invoke_result = assistant_cot_runner.run(
model_instance=model_instance,
conversation=conversation,
message=message,
query=query,
)
# handle invoke result
self._handle_invoke_result(
invoke_result=invoke_result,
queue_manager=queue_manager,
stream=application_generate_entity.stream,
agent=True
)
def _load_tool_variables(self, conversation_id: str, user_id: str, tanent_id: str) -> ToolConversationVariables:
"""
load tool variables from database
"""
tool_variables: ToolConversationVariables = db.session.query(ToolConversationVariables).filter(
ToolConversationVariables.conversation_id == conversation_id,
ToolConversationVariables.tenant_id == tanent_id
).first()
if tool_variables:
# save tool variables to session, so that we can update it later
db.session.add(tool_variables)
else:
# create new tool variables
tool_variables = ToolConversationVariables(
conversation_id=conversation_id,
user_id=user_id,
tenant_id=tanent_id,
variables_str='[]',
)
db.session.add(tool_variables)
db.session.commit()
return tool_variables
def _convert_db_variables_to_tool_variables(self, db_variables: ToolConversationVariables) -> ToolRuntimeVariablePool:
"""
convert db variables to tool variables
"""
return ToolRuntimeVariablePool(**{
'conversation_id': db_variables.conversation_id,
'user_id': db_variables.user_id,
'tenant_id': db_variables.tenant_id,
'pool': db_variables.variables
})
def _init_message_chain(self, message: Message, query: str) -> MessageChain:
"""
Init MessageChain
:param message: message
:param query: query
:return:
"""
message_chain = MessageChain(
message_id=message.id,
type="AgentExecutor",
input=json.dumps({
"input": query
})
)
db.session.add(message_chain)
db.session.commit()
return message_chain
def _save_message_chain(self, message_chain: MessageChain, output_text: str) -> None:
"""
Save MessageChain
:param message_chain: message chain
:param output_text: output text
:return:
"""
message_chain.output = json.dumps({
"output": output_text
})
db.session.commit()
def _get_usage_of_all_agent_thoughts(self, model_config: ModelConfigEntity,
message: Message) -> LLMUsage:
"""
Get usage of all agent thoughts
:param model_config: model config
:param message: message
:return:
"""
agent_thoughts = (db.session.query(MessageAgentThought)
.filter(MessageAgentThought.message_id == message.id).all())
all_message_tokens = 0
all_answer_tokens = 0
for agent_thought in agent_thoughts:
all_message_tokens += agent_thought.message_tokens
all_answer_tokens += agent_thought.answer_tokens
model_type_instance = model_config.provider_model_bundle.model_type_instance
model_type_instance = cast(LargeLanguageModel, model_type_instance)
return model_type_instance._calc_response_usage(
model_config.model,
model_config.credentials,
all_message_tokens,
all_answer_tokens
)

112
api/core/app_runner/basic_app_runner.py
View File

@@ -1,23 +1,18 @@
import logging
from typing import Optional, Tuple
from typing import Optional
from core.app_runner.app_runner import AppRunner
from core.application_queue_manager import ApplicationQueueManager, PublishFrom
from core.callback_handler.index_tool_callback_handler import DatasetIndexToolCallbackHandler
from core.entities.application_entities import (ApplicationGenerateEntity, AppOrchestrationConfigEntity, DatasetEntity,
ExternalDataVariableEntity, InvokeFrom, ModelConfigEntity)
from core.features.annotation_reply import AnnotationReplyFeature
from core.entities.application_entities import (ApplicationGenerateEntity, DatasetEntity,
InvokeFrom, ModelConfigEntity)
from core.features.dataset_retrieval import DatasetRetrievalFeature
from core.features.external_data_fetch import ExternalDataFetchFeature
from core.features.hosting_moderation import HostingModerationFeature
from core.features.moderation import ModerationFeature
from core.memory.token_buffer_memory import TokenBufferMemory
from core.model_manager import ModelInstance
from core.model_runtime.entities.message_entities import PromptMessage
from core.moderation.base import ModerationException
from core.prompt.prompt_transform import AppMode
from extensions.ext_database import db
from models.model import App, Conversation, Message, MessageAnnotation
from models.model import App, Conversation, Message
logger = logging.getLogger(__name__)
@@ -213,76 +208,6 @@ class BasicApplicationRunner(AppRunner):
stream=application_generate_entity.stream
)
def moderation_for_inputs(self, app_id: str,
tenant_id: str,
app_orchestration_config_entity: AppOrchestrationConfigEntity,
inputs: dict,
query: str) -> Tuple[bool, dict, str]:
"""
Process sensitive_word_avoidance.
:param app_id: app id
:param tenant_id: tenant id
:param app_orchestration_config_entity: app orchestration config entity
:param inputs: inputs
:param query: query
:return:
"""
moderation_feature = ModerationFeature()
return moderation_feature.check(
app_id=app_id,
tenant_id=tenant_id,
app_orchestration_config_entity=app_orchestration_config_entity,
inputs=inputs,
query=query,
)
def query_app_annotations_to_reply(self, app_record: App,
message: Message,
query: str,
user_id: str,
invoke_from: InvokeFrom) -> Optional[MessageAnnotation]:
"""
Query app annotations to reply
:param app_record: app record
:param message: message
:param query: query
:param user_id: user id
:param invoke_from: invoke from
:return:
"""
annotation_reply_feature = AnnotationReplyFeature()
return annotation_reply_feature.query(
app_record=app_record,
message=message,
query=query,
user_id=user_id,
invoke_from=invoke_from
)
def fill_in_inputs_from_external_data_tools(self, tenant_id: str,
app_id: str,
external_data_tools: list[ExternalDataVariableEntity],
inputs: dict,
query: str) -> dict:
"""
Fill in variable inputs from external data tools if exists.
:param tenant_id: workspace id
:param app_id: app id
:param external_data_tools: external data tools configs
:param inputs: the inputs
:param query: the query
:return: the filled inputs
"""
external_data_fetch_feature = ExternalDataFetchFeature()
return external_data_fetch_feature.fetch(
tenant_id=tenant_id,
app_id=app_id,
external_data_tools=external_data_tools,
inputs=inputs,
query=query
)
def retrieve_dataset_context(self, tenant_id: str,
app_record: App,
queue_manager: ApplicationQueueManager,
@@ -334,31 +259,4 @@ class BasicApplicationRunner(AppRunner):
hit_callback=hit_callback,
memory=memory
)
def check_hosting_moderation(self, application_generate_entity: ApplicationGenerateEntity,
queue_manager: ApplicationQueueManager,
prompt_messages: list[PromptMessage]) -> bool:
"""
Check hosting moderation
:param application_generate_entity: application generate entity
:param queue_manager: queue manager
:param prompt_messages: prompt messages
:return:
"""
hosting_moderation_feature = HostingModerationFeature()
moderation_result = hosting_moderation_feature.check(
application_generate_entity=application_generate_entity,
prompt_messages=prompt_messages
)
if moderation_result:
self.direct_output(
queue_manager=queue_manager,
app_orchestration_config=application_generate_entity.app_orchestration_config_entity,
prompt_messages=prompt_messages,
text="I apologize for any confusion, " \
"but I'm an AI assistant to be helpful, harmless, and honest.",
stream=application_generate_entity.stream
)
return moderation_result

49
api/core/app_runner/generate_task_pipeline.py
View File

@@ -8,7 +8,8 @@ from core.application_queue_manager import ApplicationQueueManager, PublishFrom
from core.entities.application_entities import ApplicationGenerateEntity, InvokeFrom
from core.entities.queue_entities import (AnnotationReplyEvent, QueueAgentThoughtEvent, QueueErrorEvent,
QueueMessageEndEvent, QueueMessageEvent, QueueMessageReplaceEvent,
QueuePingEvent, QueueRetrieverResourcesEvent, QueueStopEvent)
QueuePingEvent, QueueRetrieverResourcesEvent, QueueStopEvent,
QueueMessageFileEvent, QueueAgentMessageEvent)
from core.errors.error import ProviderTokenNotInitError, QuotaExceededError, ModelCurrentlyNotSupportError
from core.model_runtime.entities.llm_entities import LLMResult, LLMResultChunk, LLMResultChunkDelta, LLMUsage
from core.model_runtime.entities.message_entities import (AssistantPromptMessage, ImagePromptMessageContent,
@@ -16,11 +17,12 @@ from core.model_runtime.entities.message_entities import (AssistantPromptMessage
TextPromptMessageContent)
from core.model_runtime.errors.invoke import InvokeAuthorizationError, InvokeError
from core.model_runtime.model_providers.__base.large_language_model import LargeLanguageModel
from core.tools.tool_file_manager import ToolFileManager
from core.model_runtime.utils.encoders import jsonable_encoder
from core.prompt.prompt_template import PromptTemplateParser
from events.message_event import message_was_created
from extensions.ext_database import db
from models.model import Conversation, Message, MessageAgentThought
from models.model import Conversation, Message, MessageAgentThought, MessageFile
from pydantic import BaseModel
from services.annotation_service import AppAnnotationService
@@ -284,6 +286,7 @@ class GenerateTaskPipeline:
.filter(MessageAgentThought.id == event.agent_thought_id)
.first()
)
db.session.refresh(agent_thought)
if agent_thought:
response = {
@@ -293,16 +296,48 @@ class GenerateTaskPipeline:
'message_id': self._message.id,
'position': agent_thought.position,
'thought': agent_thought.thought,
'observation': agent_thought.observation,
'tool': agent_thought.tool,
'tool_input': agent_thought.tool_input,
'created_at': int(self._message.created_at.timestamp())
'created_at': int(self._message.created_at.timestamp()),
'message_files': agent_thought.files
}
if self._conversation.mode == 'chat':
response['conversation_id'] = self._conversation.id
yield self._yield_response(response)
elif isinstance(event, QueueMessageEvent):
elif isinstance(event, QueueMessageFileEvent):
message_file: MessageFile = (
db.session.query(MessageFile)
.filter(MessageFile.id == event.message_file_id)
.first()
)
# get extension
if '.' in message_file.url:
extension = f'.{message_file.url.split(".")[-1]}'
if len(extension) > 10:
extension = '.bin'
else:
extension = '.bin'
# add sign url
url = ToolFileManager.sign_file(file_id=message_file.id, extension=extension)
if message_file:
response = {
'event': 'message_file',
'id': message_file.id,
'type': message_file.type,
'belongs_to': message_file.belongs_to or 'user',
'url': url
}
if self._conversation.mode == 'chat':
response['conversation_id'] = self._conversation.id
yield self._yield_response(response)
elif isinstance(event, (QueueMessageEvent, QueueAgentMessageEvent)):
chunk = event.chunk
delta_text = chunk.delta.message.content
if delta_text is None:
@@ -332,7 +367,7 @@ class GenerateTaskPipeline:
self._output_moderation_handler.append_new_token(delta_text)
self._task_state.llm_result.message.content += delta_text
response = self._handle_chunk(delta_text)
response = self._handle_chunk(delta_text, agent=isinstance(event, QueueAgentMessageEvent))
yield self._yield_response(response)
elif isinstance(event, QueueMessageReplaceEvent):
response = {
@@ -384,14 +419,14 @@ class GenerateTaskPipeline:
extras=self._application_generate_entity.extras
)
def _handle_chunk(self, text: str) -> dict:
def _handle_chunk(self, text: str, agent: bool = False) -> dict:
"""
Handle completed event.
:param text: text
:return:
"""
response = {
'event': 'message',
'event': 'message' if not agent else 'agent_message',
'id': self._message.id,
'task_id': self._application_generate_entity.task_id,
'message_id': self._message.id,