Self-contained outpatient orientation workflow (chatflow) in Dify

Note: This article uses Dify v0.7.2.

This chat flow shows how to build a chatbot for outpatient guidance that can collect patient data via web or voice dialog. Simply understood, this means that the patient is given a recommendation for a department based on the patient's information (age, gender, and symptoms).

I. Workflow ideas

1. Workflow Screenshot

Screenshot of the complete workflow of the outpatient orientation, shown below:

2. Workflow pseudo-flowchart

The purpose of the outpatient guidance workflow is to recommend a department to a patient based on the patient's input of age, gender and symptoms. In the first round of dialog, since the patient may enter all or part of the information at once, the "Parameter Extractor" node is used to extract the gender, symptom and age from the patient's first input. The variable is_first_message is set to 0 by the "Variable Assignment" node.

3. Problems 1

During subsequent rounds of conversation, only 1 type of field information is recognized at a time, although the patient may provide multiple fields, with the order of field recognition being age, gender, and symptoms.

4. Problems 2

If a department has already been recommended based on patient information, then entering any information again will recommend the last department. Because the department field is not empty, the direct response process is followed.

II. Session variables

The session variable is used to store contextual information needed by LLM, such as user preferences, conversation history, etc. It is read-write. Session variables are oriented to multi-round conversation scenarios, so they are only available for Chatflow type (Chat Assistant -> Workflow Orchestration) applications.

1.session variabledata type

(1) String

(2) Number

(3) Object

(4) Array[string] Array of strings

(5) Array[number] Array of numbers

(6) Array[object] Array of objects

2.session variablecharacterization

(1) Session variables can be globally referenced within most nodes;

(2) Writing session variables requires the use of thevariable assignmentNodes;

(3) Session variables are read-write.

3. Session variables defined for this workflow

The five session variables defined for this workflow are whether it is the first round of conversation, age, gender, symptoms and department.

III. Workflow-related interfaces

1. Get the interface to the workflow

(1) Source code location

• Source code location: dedify-0.7.2\api\controllers\console\app\workflow.py
• Source code location: dedify-0.7.2\api\services\workflow_service.py

(2) Getting workflow operations

http://localhost:5001/console/api/apps/3c309371-54f6-4bfb-894a-4193b189ffa5/workflows/draft. where draft, like draft, indicates debugging in the workflow page.

(3) Acquisition of workflow implementations

The workflow stores a record in the workflows data table. The version is now draft and every time a workflow is published, a record is inserted in that table. As shown below:

(4) Workflow data structure

The Workflow data structure returned by the interface is shown below:

class Workflow(db.Model).
__tablename__ = 'workflows'
__table_args__ = (
db.PrimaryKeyConstraint('id', name='workflow_pkey'),
db.Index('workflow_version_idx', 'tenant_id', 'app_id', 'version'),
)
id: Mapped[str] = db.Column(StringUUID, server_default=db.text('uuid_generate_v4()'))
tenant_id: Mapped[str] = db.Column(StringUUID, nullable=False)
app_id: Mapped[str] = db.Column(StringUUID, nullable=False)
type: Mapped[str] = db.Column(db.String(255), nullable=False)
version: Mapped[str] = db.Column(db.String(255), nullable=False)
graph: Mapped[str] = db.Column(db.Text)
features: Mapped[str] = db.Column(db.Text)
created_by: Mapped[str] = db.Column(StringUUID, nullable=False)
created_at: Mapped[datetime] = db.Column(db.DateTime, nullable=False, server_default=db.text('CURRENT_TIMESTAMP(0)'))
updated_by: Mapped[str] = db.Column(StringUUID)
updated_at: Mapped[datetime] = db.Column(db.DateTime)
_environment_variables: Mapped[str] = db.Column('environment_variables', db.Text, nullable=False, server_default='{}')
_conversation_variables: Mapped[str] = db.Column('conversation_variables', db.Text, nullable=False, server_default='{}')

2. Update workflow interfaces

(1) Source code location

• Source Location: dedify-0.7.2\web\app\components\workflow\hooks\use-nodes-sync-draft.ts
• Source Location: dedify-0.7.2\web\service\workflow.ts

(2) Update workflow operations

http://localhost:5001/console/api/apps/3c309371-54f6-4bfb-894a-4193b189ffa5/workflows/draft.

(3) Update workflow front-end code

Due to the React Flow framework used in the front-end of the Dify workflow, it is presumed that POST operations, such as add, delete, and change, may be performed when the workflow changes. Through the logs in the Console of the back-end, it was found that the call to theworkflows/draft?_token=::

pass (a bill or inspection etc)workflows/draft?_token=Searching through the front-end code found only one place to use it, and that must be here.

Since the POST operation is performed, there must be a Payload parameter, right, and the workflow parameter was found in that file.

When the dependencygetPostParamsWhen a change occurs, it is executed asynchronouslysyncWorkflowDraft(postParams)The

(indicates contrast)syncWorkflowDraft(postParams)The actual call is POST. as shown below:

(4) Summary of workflow functions

rightdify-0.7.2\web\service\workflow.tsThe functions in are summarized as follows:

3. Workflow implementation

Reference to theProcess implementation of Chatflow create, update, execute and delete operationsThe "Chatflow Implementation" section in [6].

4. Workflow publishing

(1) Source code location

• Source code location: dedify-0.7.2\api\controllers\console\app\workflow.py
• Source code location: dedify-0.7.2\api\services\workflow_service.py
• Source location: dedify-0.7.2\api\events\event_handlers\update_app_dataset_join_when_app_published_workflow_updated.py

(2) Workflow publishing interface

This is executed when you click Publish Workflowhttp://localhost:5001/console/api/apps/3c309371-54f6-4bfb-894a-4193b189ffa5/workflows/publishInterface.

It mainly consists of creating a new workflow and triggering the app workflow event. As shown below:

(3) Creating a new workflow record

This mainly consists of creating a new workflow record where the content of the version field is 2024-09-07 09:11:25.894535. indicating that the workflow has been published, not draft (the workflow has not been published and is still being debugged in the workflow canvas). As shown below:

(4) app_published_workflow_was_updated event implementation

@app_published_workflow_was_updated.connect
def handle(sender, **kwargs):
app = sender
published_workflow = kwargs.get("published_workflow")
published_workflow = cast(Workflow, published_workflow)
dataset_ids = get_dataset_ids_from_workflow(published_workflow) # Get dataset ID from workflow
app_dataset_joins = db.session.query(AppDatasetJoin).filter(AppDatasetJoin.app_id == app.id).all() # Get application dataset associations
removed_dataset_ids = [] # Use to store removed dataset IDs
if not app_dataset_joins: # If there are no app dataset joins
added_dataset_ids = dataset_ids # Add dataset ids
else: # If there is an application dataset association
old_dataset_ids = set() # for storing old dataset IDs
for app_dataset_join in app_dataset_joins: # Iterate over app dataset associations
old_dataset_ids.add(app_dataset_join.dataset_id) # Add dataset IDs
added_dataset_ids = dataset_ids - old_dataset_ids # Add dataset IDs
removed_dataset_ids = old_dataset_ids - dataset_ids # Remove dataset IDs
if removed_dataset_ids: # if there are removed dataset IDs
for dataset_id in removed_dataset_ids: # Iterate over removed dataset IDs
db.session.query(AppDatasetJoin).filter(
AppDatasetJoin.app_id == app.id, AppDatasetJoin.dataset_id == dataset_id
).delete() # Delete an application dataset association
if added_dataset_ids: # if there is an added dataset ID
for dataset_id in added_dataset_ids: # Iterate through the added dataset IDs
app_dataset_join = AppDatasetJoin(app_id=app.id, dataset_id=dataset_id) # Create the app dataset association
db.session.add(app_dataset_join) # Add an application dataset association
db.session.commit() # Commit the transaction

The function of this code is to handle the app_published_workflow_was_updated Signal to update the association relationship between an application and a dataset when the published workflow for that application is updated. The specific steps are as follows:

• Get the application object that sends the signal app and updated workflow objects published_workflowThe
• invocations get_dataset_ids_from_workflow function that extracts the dataset ID collection from the updated workflow dataset_idsThe
• Query the database to get all the current application's AppDatasetJoin Records.
• Calculates the dataset IDs to be added and deleted:
  • If there are no current AppDatasetJoin records, then all extracted dataset IDs need to be added.
  • Otherwise, calculate the dataset IDs to be added and deleted.
• Deleting the need to remove AppDatasetJoin Records.
• Add new AppDatasetJoin Records.
• Submit a database transaction.

(5) Role of the AppDatasetJoin table

The AppDatasetJoin table serves to maintain a many-to-many relationship between an application (App) and a dataset (Dataset). Each record represents the association of an application with a dataset. The specific fields are as follows:

• id: primary key that uniquely identifies a record.
• app_id: A unique identifier for the application.
• dataset_id: The unique identifier of the dataset.
• created_at: The timestamp of the record's creation.
• With this table, it is possible to query which datasets are associated with a particular application, or which applications a particular dataset is associated with.

(6) get_dataset_ids_from_workflow() implementation

def get_dataset_ids_from_workflow(published_workflow: Workflow) -> set: # Get dataset IDs from workflow
dataset_ids = set() # Used to store dataset IDs
graph = published_workflow.graph_dict # Get the workflow graph
if not graph: # If there is no graph
return dataset_ids # return empty set
nodes = graph.get("nodes", []) # fetch the nodes in the graph
# fetch all knowledge retrieval nodes # fetch all knowledge retrieval nodes
knowledge_retrieval_nodes = [
node for node in nodes if node.get("data", {}).get("type") == NodeType.KNOWLEDGE_RETRIEVAL.value
] # Get all knowledge retrieval nodes
if not knowledge_retrieval_nodes: # If there are no knowledge retrieval nodes
return dataset_ids # Return the empty set.
for node in knowledge_retrieval_nodes: # Iterate over knowledge retrieval nodes
try: node_data = Knowledge_retrieval_nodes.
node_data = KnowledgeRetrievalNodeData(**node.get("data", {})) # get node data
dataset_ids.update(node_data.dataset_ids) # Update dataset IDs
except Exception as e: # if an exception occurs
continue
return dataset_ids

The main function of the get_dataset_ids_from_workflow function is to extract the relevant dataset IDs of all the knowledge retrieval nodes from a given workflow object. the specific steps are as follows:

• Initializes an empty dataset_ids collection for storing dataset IDs.
• Get the graph structure of the workflow graph.
• Returns the empty dataset_ids collection if the graph structure is empty.
• Get all nodes nodes in the graph.
• Filter all nodes of type KNOWLEDGE_RETRIEVAL knowledge_retrieval_nodes.
• Iterate over these knowledge retrieval nodes, extract their dataset IDs and update them in the dataset_ids collection.
• Returns a collection of all extracted dataset IDs.

5. Workflow other interfaces

This part of the interface will not be elaborated, the workflow orchestration conversational application API has been described very clearly.

bibliography

[1] Session variables: https://docs.dify.ai/v/zh-hans/guides/workflow/variables

[2] Variable assignment: https://docs.dify.ai/v/zh-hans/guides/workflow/node/variable-assignment

[3] Variable aggregation: https://docs.dify.ai/v/zh-hans/guides/workflow/node/variable-assigner

[4] React Flow Chinese website: https://reactflow-cn.js.org/

[5] React Flow English website: https://reactflow.dev/

[6] Process implementation of Chatflow create, update, execute and delete operations: https://z0yrmerhgi8.feishu.cn/wiki/FFzxwdF4PijlhjkLUoecOv6Vn8a

The layman's guide to getting started with AI

Help you learn how to utilize AI tools at a low cost and from a zero base.AI, like office software, is an essential skill for everyone. Mastering AI will give you an edge in your job search and half the effort in your future work and studies.

View Details>