Create a new instruct task
Technical definition
A task is a JSON object that contains all necessary information to render it in the app and to guide the assistant through the process of drafting the document resulting from the task.
You can find a template of a task in the file task_spec.json in this repository.
Let's break this specification down with "meeting_minutes" task as an example.
Task Type
The type of the task. Can be "workflow" or "instruct". A workflow is a task that requires multiple steps, executed as code, to be completed. See dedicatednew_workflow documentation for more information on how to create a workflow.
Platforms
Platforms on which the task is available. Can be "webapp", "mobile" or both.System information
"name_for_system": "prepare_meeting_minutes",
"definition_for_system": "The user needs assistance to prepare a meeting minutes"
Final instruction
This is the final instruction provided to the assistant prompt. It contains all instruction the assistant will follow to draft the document. It ends with an infinitive verb sentence as if it was an order to the assistant.Output format
"output_format_instruction_title": "SHORT CONTEXT - DATE OF THE DAY",
"output_format_instruction_draft": "CONTENT OF THE MEETING MINUTES"
Output type
This field is used to define the type of document resulting from the task. It is used to enable special edition features once the document is ready. The value should match one existing in table 'md_text_type' of your database, if it does not, will be created.Icon
The icon displayed in the app to represent the task.Task displayed data
"task_displayed_data": [
{
"language_code": "en",
"name_for_user": "Meeting Recap",
"definition_for_user": "Get a simple summary and next steps for your meeting",
"json_input": [
{
"input_name": "meeting_informations",
"description_for_user": "How was your meeting?",
"description_for_system": "Informations the user gave about the meeting (ex: participants, date, key topics, followup actions...)",
"placeholder": "Record a quick summary of what was discussed.",
"type": "text_area"
}
]
},
{
"language_code": "fr",
"name_for_user": "RĂ©capitulatif de RĂ©union",
"definition_for_user": "Obtenez un récapitulatif simple de votre réunion",
"json_input": [
{
"description_for_system": "Informations que l'utilisateur a fournies sur la réunion (ex : participants, date, sujets clés, actions de suivi...)",
"description_for_user": "Comment s'est passée votre réunion ?",
"input_name": "informations_reunion",
"placeholder": "Enregistrez un bref résumé de ce qui a été discuté.",
"type": "text_area"
}
]
}
]
Note: For mobile use to remain interface friendly, we recommend to keep the number of inputs to 1. If the assistant needs to ask more than one question, it will do it in a conversational way.
Infos to extract
"infos_to_extract": [
{
"info_name": "key_topics",
"description": "Key topics discussed in the meeting"
},
{
"info_name": "participants",
"description": "Participants of the meeting"
},
{
"info_name": "date_of_meeting",
"description": "Date of the meeting"
},
{
"info_name": "followup_actions",
"description": "Followup actions if any"
}
]
predefined_actions
"predefined_actions": [
{
"task_pk": <pk of the task to chain>,
"displayed_data": [
{
"language_code": "en",
"data": {
"name": "Send a follow-up email", # name of the predefined action
"button_text": "Send email", # text displayed on the button of the predefined action
"message_prefix": "Prepare a follow-up email to send this meeting minutes to the participants: " # prefix of the message that will be sent to the assistant in behalf of the user to start the task. It will be concatenated with the produced_text of the previous task.
}
}
]
}
]
result_chat_enabled
This field is used to enable the chat at the end of the task, once produced_text has been generated, to let user give instruction to edit the produced_text. It is by default true. On instruct tasks, it is recommended to keep it true.How to create a new instruct task?
Let's create your first task on Mojodex.
From the root of the repository.
STEP 1: Create the json file
cp ./docs/configure_assistant/tasks/task_spec.json ./docs/configure_assistant/tasks/tasks_json/my_task.json
Fill the json values of my_task.json with the ones fitting your need for this task. You can get helped by using the dedicated route so that GPT-4 generates a first json for you from your requirements.
curl --location --request POST 'http://localhost:5001/task_json' \
--header 'Authorization: backoffice_secret' \
--header 'Content-Type: application/json' \
--data-raw '{"datetime": "2024-02-14T17:49:26.545180",
"task_requirements": "You are the sales assistant of the user. The user needs help to (achieve/do/perform) [...]. The user needs this help in the following situation: [...]. You will need the following informations to proceed: [...]. The result of the task is a[{document_type}]"
}'
STEP 2: Add the task to the database
curl --location --request PUT 'http://localhost:5001/task' \
--header 'Authorization: backoffice_secret' \
--header 'Content-Type: application/json' \
--data @my_task.json
STEP 3: Associate the task to your user though a product.
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>
}'
STEP 4: Test your task
Run your local Mojodex instance and access the web or mobile app.
You should now see your task in the list of available tasks. If you don't, ensure to reload the app to get rid of any potential cache.