dosubot[bot]
commented
1 month ago To resolve the issue where the json property in the result of the parent workflow (B) is empty, and the expected result is incorrectly placed in the text property when integrating workflow (A) into workflow (B), you need to ensure that the LLMNode is correctly configured to extract the JSON response.
From the provided code, it appears that the LLMNode is used to handle the response. The prompt_template and prompt_config should be set up to ensure the JSON response is correctly extracted and placed in the json property.
Here is an example of how you can ensure the JSON response is correctly extracted and placed in the json property:
import json
import os
from unittest.mock import MagicMock
import pytest
from core.app.entities.app_invoke_entities import InvokeFrom, ModelConfigWithCredentialsEntity
from core.entities.provider_configuration import ProviderConfiguration, ProviderModelBundle
from core.entities.provider_entities import CustomConfiguration, CustomProviderConfiguration, SystemConfiguration
from core.model_manager import ModelInstance
from core.model_runtime.entities.model_entities import ModelType
from core.model_runtime.model_providers import ModelProviderFactory
from core.workflow.entities.node_entities import SystemVariable
from core.workflow.entities.variable_pool import VariablePool
from core.workflow.nodes.base_node import UserFrom
from core.workflow.nodes.llm.llm_node import LLMNode
from extensions.ext_database import db
from models.provider import ProviderType
from models.workflow import WorkflowNodeExecutionStatus
@pytest.mark.parametrize('setup_openai_mock', [['chat']], indirect=True)
def test_execute_llm(setup_openai_mock):
node = LLMNode(
tenant_id='1',
app_id='1',
workflow_id='1',
user_id='1',
invoke_from=InvokeFrom.WEB_APP,
user_from=UserFrom.ACCOUNT,
config={
'id': 'llm',
'data': {
'title': '123',
'type': 'llm',
'model': {
'provider': 'openai',
'name': 'gpt-3.5-turbo',
'mode': 'chat',
'completion_params': {}
},
'prompt_template': [
{
'role': 'system',
'text': 'you are a helpful assistant.\ntoday\'s weather is {{#abc.output#}}.'
},
{
'role': 'user',
'text': '{{#sys.query#}}'
}
],
'memory': None,
'context': {
'enabled': False
},
'vision': {
'enabled': False
}
}
}
)
# construct variable pool
pool = VariablePool(system_variables={
SystemVariable.QUERY: 'what\'s the weather today?',
SystemVariable.FILES: [],
SystemVariable.CONVERSATION_ID: 'abababa',
SystemVariable.USER_ID: 'aaa'
}, user_inputs={})
pool.append_variable(node_id='abc', variable_key_list=['output'], value='sunny')
credentials = {
'openai_api_key': os.environ.get('OPENAI_API_KEY')
}
provider_instance = ModelProviderFactory().get_provider_instance('openai')
model_type_instance = provider_instance.get_model_instance(ModelType.LLM)
provider_model_bundle = ProviderModelBundle(
configuration=ProviderConfiguration(
tenant_id='1',
provider=provider_instance.get_provider_schema(),
preferred_provider_type=ProviderType.CUSTOM,
using_provider_type=ProviderType.CUSTOM,
system_configuration=SystemConfiguration(
enabled=False
),
custom_configuration=CustomConfiguration(
provider=CustomProviderConfiguration(
credentials=credentials
)
),
model_settings=[]
),
provider_instance=provider_instance,
model_type_instance=model_type_instance
)
model_instance = ModelInstance(provider_model_bundle=provider_model_bundle, model='gpt-3.5-turbo')
model_config = ModelConfigWithCredentialsEntity(
model='gpt-3.5-turbo',
provider='openai',
mode='chat',
credentials=credentials,
parameters={},
model_schema=model_type_instance.get_model_schema('gpt-3.5-turbo'),
provider_model_bundle=provider_model_bundle
)
# Mock db.session.close()
db.session.close = MagicMock()
node._fetch_model_config = MagicMock(return_value=(model_instance, model_config))
# execute node
result = node.run(pool)
assert result.status == WorkflowNodeExecutionStatus.SUCCEEDED
assert result.outputs['text'] is not None
assert result.outputs['usage']['total_tokens'] > 0
@pytest.mark.parametrize('setup_code_executor_mock', [['none']], indirect=True)
@pytest.mark.parametrize('setup_openai_mock', [['chat']], indirect=True)
def test_execute_llm_with_jinja2(setup_code_executor_mock, setup_openai_mock):
"""
Test execute LLM node with jinja2
"""
node = LLMNode(
tenant_id='1',
app_id='1',
workflow_id='1',
user_id='1',
invoke_from=InvokeFrom.WEB_APP,
user_from=UserFrom.ACCOUNT,
config={
'id': 'llm',
'data': {
'title': '123',
'type': 'llm',
'model': {
'provider': 'openai',
'name': 'gpt-3.5-turbo',
'mode': 'chat',
'completion_params': {}
},
'prompt_config': {
'jinja2_variables': [{
'variable': 'sys_query',
'value_selector': ['sys', 'query']
}, {
'variable': 'output',
'value_selector': ['abc', 'output']
}]
},
'prompt_template': [
{
'role': 'system',
'text': 'you are a helpful assistant.\ntoday\'s weather is {{#abc.output#}}',
'jinja2_text': 'you are a helpful assistant.\ntoday\'s weather is {{output}}.',
'edition_type': 'jinja2'
},
{
'role': 'user',
'text': '{{#sys.query#}}',
'jinja2_text': '{{sys_query}}',
'edition_type': 'basic'
}
],
'memory': None,
'context': {
'enabled': False
},
'vision': {
'enabled': False
}
}
}
)
# construct variable pool
pool = VariablePool(system_variables={
SystemVariable.QUERY: 'what\'s the weather today?',
SystemVariable.FILES: [],
SystemVariable.CONVERSATION_ID: 'abababa',
SystemVariable.USER_ID: 'aaa'
}, user_inputs={})
pool.append_variable(node_id='abc', variable_key_list=['output'], value='sunny')
credentials = {
'openai_api_key': os.environ.get('OPENAI_API_KEY')
}
provider_instance = ModelProviderFactory().get_provider_instance('openai')
model_type_instance = provider_instance.get_model_instance(ModelType.LLM)
provider_model_bundle = ProviderModelBundle(
configuration=ProviderConfiguration(
tenant_id='1',
provider=provider_instance.get_provider_schema(),
preferred_provider_type=ProviderType.CUSTOM,
using_provider_type=ProviderType.CUSTOM,
system_configuration=SystemConfiguration(
enabled=False
),
custom_configuration=CustomConfiguration(
provider=CustomProviderConfiguration(
credentials=credentials
)
),
model_settings=[]
),
provider_instance=provider_instance,
model_type_instance=model_type_instance,
)
model_instance = ModelInstance(provider_model_bundle=provider_model_bundle, model='gpt-3.5-turbo')
model_config = ModelConfigWithCredentialsEntity(
model='gpt-3.5-turbo',
provider='openai',
mode='chat',
credentials=credentials,
parameters={},
model_schema=model_type_instance.get_model_schema('gpt-3.5-turbo'),
provider_model_bundle=provider_model_bundle
)
# Mock db.session.close()
db.session.close = MagicMock()
node._fetch_model_config = MagicMock(return_value=(model_instance, model_config))
# execute node
result = node.run(pool)
assert result.status == WorkflowNodeExecutionStatus.SUCCEEDED
assert 'sunny' in json.dumps(result.process_data)
assert 'what\'s the weather today?' in json.dumps(result.process_data)Ensure that the LLMNode is configured to use the correct prompt templates and Jinja2 variables to extract the JSON response. The prompt_template should be correctly formatted to guide the model to produce a JSON response. If the JSON response is still placed in the text property, you may need to adjust the extraction logic to parse the text property and extract the JSON content manually.
Additionally, you can use the following prompt templates from Dify to ensure the JSON response is correctly formatted and extracted:
### Extract parameter Workflow
I need to extract the following information from the input text. The <information to be extracted> tag specifies the 'type', 'description' and 'required' of the information to be extracted.
<information to be extracted>
{{ structure }}
</information to be extracted>
Step 1: Carefully read the input and understand the structure of the expected output.
Step 2: Extract relevant parameters from the provided text based on the name and description of object.
Step 3: Structure the extracted parameters to JSON object as specified in <structure>.
Step 4: Ensure that the JSON object is properly formatted and valid. The output should not contain any XML tags. Only the JSON object should be outputted.
### Memory
Here is the chat histories between human and assistant, inside <histories></histories> XML tags.
<histories>
{histories}
</histories>
### Structure
Here is the structure of the expected output, I should always follow the output structure.
{{γγγ
'properties1': 'relevant text extracted from input',
'properties2': 'relevant text extracted from input',
}}γγγ
### Input Text
Inside <text></text> XML tags, there is a text that I should extract parameters and convert to a JSON object.
<text>
{text}
</text>
### Answer
I should always output a valid JSON object. Output nothing other than the JSON object.
```JSON
By using these prompt templates, you can ensure that the JSON response is correctly formatted and extracted, and placed in the `json` property of the result <sup>[[1]](https://github.com/langgenius/dify/blob/main/api/tests/integration_tests/workflow/nodes/test_parameter_extractor.py)</sup><sup>[[2]](https://github.com/langgenius/dify/blob/main/api/tests/integration_tests/workflow/nodes/test_llm.py)</sup><sup>[[3]](https://github.com/langgenius/dify/blob/main/api/core/workflow/nodes/parameter_extractor/prompts.py)</sup>.
<!-- Dosu Comment Footer -->
*To continue talking to [Dosu](https://dosu.dev), mention @dosu.*
crazywoola
akramer-zibra
Self Checks
Dify version
0.6.15
Cloud or Self Hosted
Self Hosted (Docker)
Steps to reproduce
(1) Create a workflow (A) which gets a simple string
inputparameter at start and provides this asresultwhen this workflow ends.(2) Create a second workflow (B) which gets a simple
inputparameter at start, passes this parameter as input into workflow A. This parent workflow providestext,filesandjsonas final output, when it ends.(3) Run the parent workflow (B) with the input
Hello World!✔️ Expected Behavior
The
jsonproperty in the result of the parent workflow (B) contains the result of the integrated workflow (A):❌ Actual Behavior
The
jsonproperty in the result of the parent workflow (B) is an empty array. And there is also a wrong stringified version of the expected result in thetextproperty.