Create a new workflow

Technical definition

A workflow is a sequence of steps that the agent must execute one by one. A step is code: anything can be done in this code including complex actions as API calls.

The result of each step execution is always available to the user that can consult it. On the steps requiring user validation, user can edit the step result and sometimes open a chat to give an instruction to relaunch the step. Compared to the instruct tasks, the workflows are more complex and constrainted, allowing the user to keep more control over the agent's actions.

How to create a new workflow?

STEP 1: Create the steps' code

In mojodex_core>workflows create a new folder named by your workflow. In this folder, create one file per step in your workflow. In each of those files, implement the corresponding step code. A workflow step is a class that implements the WorkflowStep class.

An implementation of WorkflowStep overrides the method _execute(). It looks like this:

from mojodex_core.entities.workflow_step import WorkflowStep      

class MyWorkflowStep(WorkflowStep):


    def _execute(self, parameter: dict, learned_instructions: dict, initial_parameter: dict, past_validated_steps_results: List[dict], user_id: str, user_task_execution_pk: int, task_name_for_system: str, session_id:str):
        try: 
            return [{'output_key1': <output>, 'output_key2': <output>}]
        except Exception as e:
            raise Exception(f"execute :: {e}")

The result of the step execution must be a list of dictionaries. The next step of the workflow will be executed as many time as there are dictionaries in the list, each time with the corresponding dictionary as parameter.

STEP 2: Add your steps to the steps library

In mojodex_core>workflows>steps_library.py, add your steps to the STEPS dictionary. The key must be the name of the step as later defined in DB and the value must be the class of the step. This is used to dynamically load the steps from their name in the database.

STEP 3: Create the workflow

To create a new workflow, you can use the dedicated route PUT /task.

Replace <BACKOFFICE_SECRET> with your actual token and name, steps, icon and description with the actual workflow name, steps, icon and description.

Careful: steps names must match the ones in the steps library.

Then, run the following command in your terminal:

curl --location --request PUT 'http://localhost:5001/task' \
--header 'Authorization: <BACKOFFICE_SECRET>' \
--header 'Content-Type: application/json' \
--data-raw '{"datetime": "2024-02-14T17:49:26.545180",
    "task_type": "workflow",
    "predefined_actions": [],
    "name_for_system": "<WORKFLOW_NAME_FOR_SYSTEM>",
    "icon": "<EMOJI>",
    "definition_for_system": "<WORKFLOW_DEFINITION_FOR_SYSTEM>",
    "output_type": "document",
    "platforms": ["webapp"], # List of platforms where the workflow is available. Workflows are not yet implemented on mobile app.
    "steps": [{
        "name_for_system": "<STEP_NAME_FOR_SYSTEM>",
        "definition_for_system":  "<STEP_DEFINITION_FOR_SYSTEM>",
        "rank": 1,
        "review_chat_enabled": false, # Whether the chat is enabled at the end of the step or not.
        "user_validation_required": true, # Whether the user validation is required or not at the end of this step
        "step_displayed_data":[
            {
                "language_code": "<2-LETTERS LANGUAGE-CODE>",
                "name_for_user": "<STEP_NAME_FOR_USER>",
                "definition_for_user": "<STEP_DEFINITION_FOR_USER>"
            },
            ...
        ]},
        ...
        ],
    "task_displayed_data": [
        {   
            "language_code":"<2-LETTERS LANGUAGE-CODE>",
            "name_for_user": <WORKFLOW_NAME_FOR_USER>",
            "definition_for_user": <WORKFLOW_DEFINITION_FOR_USER>",
            "json_input": [
                {"input_name": "<INPUT_NAME>", 
                "description_for_user": "<DESCRIPTION_FOR_USER>",
                    "description_for_system": "<DESCRIPTION_FOR_SYSTEM>",
                "type": "text_area"},
                ...
               ]
        },
         ...
    ],
    "result_chat_enabled": false # Whether the chat is enabled at the end of the workflow on the produced text or not.
}'

Note regarding "result_chat_enabled" parameter: Be careful that once on result page: - Assistant will no longer be able to re-launch past steps code - therefore relaunch api requests or other… - Text has been generated by a whole chain of steps. It could be too long for context-windows on a basic LLM chat call.

As a consequence, depending on the workflow, it may be relevant to disable the chat on result to avoid user confusion (preventing them to ask to re-launch a query for example)

STEP 4: Associate workflow to a user through a product.

Replace <BACKOFFICE_SECRET> with your actual token and <user_id> and <task_pk> with the actual user id and workflow primary key. Then, run the following command in your terminal:

Default user demo@example.com is associated with default product demo with pk 1. Let's add the task to this product.

curl --location --request PUT 'http://localhost:5001/product_task_association' \
--header 'Authorization: backoffice_secret' \
--header 'Content-Type: application/json' \
--data-raw '{
 "datetime": "'"$(date -u +"%Y-%m-%dT%H:%M:%S.%3NZ")"'",
 "product_pk": 1,
 "task_pk": <task_pk retrieved from previous command>
}'