Canevas

Le canevas est la surface riche en contenus de l’interface de chat : documents markdown, code, diagrammes, présentations et composants interactifs. Un workflow peut retourner un canevas en sortie ou en envoyer un en cours d’exécution, permettant à l’utilisateur de le modifier avant de continuer.

Sorties enrichies

Retournez des contenus riches comme du markdown, du code ou des diagrammes à l’aide de ResourceOutput :

import mistralai.workflows as workflows
import mistralai.workflows.plugins.mistralai as workflows_mistralai


@workflows.workflow.define(
    name="expense-report-workflow",
    workflow_display_name="Expense Report",
    workflow_description="Generate expense report with charts",
)
class ExpenseReportWorkflow:
    @workflows.workflow.entrypoint
    async def run(self, department: str) -> workflows_mistralai.ChatAssistantWorkflowOutput:
        # Generate a mermaid chart
        chart_content = """
pie title Expenses by Category
    "Travel" : 45
    "Equipment" : 25
    "Software" : 20
    "Other" : 10
"""

        canvas = workflows_mistralai.CanvasPayload(
            type="mermaid",
            title="Expense Breakdown",
            content=chart_content,
        )

        resource = workflows_mistralai.CanvasResource(
            canvas=canvas,
        )

        return workflows_mistralai.ChatAssistantWorkflowOutput(
            content=[
                workflows_mistralai.TextOutput(text=f"Expense report for {department}:"),
                workflows_mistralai.ResourceOutput(resource=resource),
            ]
        )

import mistralai.workflows as workflows
import mistralai.workflows.plugins.mistralai as workflows_mistralai


@workflows.workflow.define(
    name="expense-report-workflow",
    workflow_display_name="Expense Report",
    workflow_description="Generate expense report with charts",
)
class ExpenseReportWorkflow:
    @workflows.workflow.entrypoint
    async def run(self, department: str) -> workflows_mistralai.ChatAssistantWorkflowOutput:
        # Generate a mermaid chart
        chart_content = """
pie title Expenses by Category
    "Travel" : 45
    "Equipment" : 25
    "Software" : 20
    "Other" : 10
"""

        canvas = workflows_mistralai.CanvasPayload(
            type="mermaid",
            title="Expense Breakdown",
            content=chart_content,
        )

        resource = workflows_mistralai.CanvasResource(
            canvas=canvas,
        )

        return workflows_mistralai.ChatAssistantWorkflowOutput(
            content=[
                workflows_mistralai.TextOutput(text=f"Expense report for {department}:"),
                workflows_mistralai.ResourceOutput(resource=resource),
            ]
        )

Types de canevas

Type	Description
`text/markdown`	Contenu Markdown
`text/html`	Contenu HTML
`image/svg+xml`	Images SVG
`slides`	Présentations
`react`	Composants React
`code`	Code avec coloration syntaxique
`mermaid`	Diagrammes Mermaid

Édition du canevas (boucle humaine)

Vous pouvez envoyer un canevas pendant un workflow avec send_assistant_message() et permettre à l’utilisateur de le modifier. Le champ canvas_uri transmis à CanvasInput doit être identique au uri d’une ressource CanvasResource générée précédemment.

import mistralai.workflows as workflows
import mistralai.workflows.plugins.mistralai as workflows_mistralai
from mistralai.workflows.conversational import CanvasInput


@workflows.workflow.define(
    name="report-review-workflow",
    workflow_display_name="Report Review",
    workflow_description="Generate a report and let the user edit it",
)
class ReportReviewWorkflow(workflows.InteractiveWorkflow):
    @workflows.workflow.entrypoint
    async def run(self) -> workflows_mistralai.ChatAssistantWorkflowOutput:
        # Send a canvas to the user as an assistant message
        canvas_resource = workflows_mistralai.CanvasResource(
            canvas=workflows_mistralai.CanvasPayload(
                type="text/markdown",
                title="Weekly Report",
                content="# Weekly Report\n\n## Summary\n\nTODO: fill in",
            ),
        )
        await workflows_mistralai.send_assistant_message(
            "Here is your report draft. You can review and edit it below.",
            canvas=canvas_resource,
        )

        # Wait for the user to edit the canvas
        edited = await self.wait_for_input(
            CanvasInput(canvas_uri=canvas_resource.uri, prompt="Any feedback?"),
            label="Relecture et édition du rapport",
        )

        # Use the edited content
        return workflows_mistralai.ChatAssistantWorkflowOutput(
            content=[
                workflows_mistralai.TextOutput(text="Report finalized!"),
                workflows_mistralai.ResourceOutput(
                    resource=workflows_mistralai.CanvasResource(
                        canvas=workflows_mistralai.CanvasPayload(
                            type="text/markdown",
                            title=edited.canvas.title,
                            content=edited.canvas.content,
                        ),
                    )
                ),
            ]
        )

import mistralai.workflows as workflows
import mistralai.workflows.plugins.mistralai as workflows_mistralai
from mistralai.workflows.conversational import CanvasInput


@workflows.workflow.define(
    name="report-review-workflow",
    workflow_display_name="Report Review",
    workflow_description="Generate a report and let the user edit it",
)
class ReportReviewWorkflow(workflows.InteractiveWorkflow):
    @workflows.workflow.entrypoint
    async def run(self) -> workflows_mistralai.ChatAssistantWorkflowOutput:
        # Send a canvas to the user as an assistant message
        canvas_resource = workflows_mistralai.CanvasResource(
            canvas=workflows_mistralai.CanvasPayload(
                type="text/markdown",
                title="Weekly Report",
                content="# Weekly Report\n\n## Summary\n\nTODO: fill in",
            ),
        )
        await workflows_mistralai.send_assistant_message(
            "Here is your report draft. You can review and edit it below.",
            canvas=canvas_resource,
        )

        # Wait for the user to edit the canvas
        edited = await self.wait_for_input(
            CanvasInput(canvas_uri=canvas_resource.uri, prompt="Any feedback?"),
            label="Relecture et édition du rapport",
        )

        # Use the edited content
        return workflows_mistralai.ChatAssistantWorkflowOutput(
            content=[
                workflows_mistralai.TextOutput(text="Report finalized!"),
                workflows_mistralai.ResourceOutput(
                    resource=workflows_mistralai.CanvasResource(
                        canvas=workflows_mistralai.CanvasPayload(
                            type="text/markdown",
                            title=edited.canvas.title,
                            content=edited.canvas.content,
                        ),
                    )
                ),
            ]
        )

CanvasInput retourne un modèle avec :

canvas.title — le titre du canevas édité
canvas.content — le contenu modifié
chatInput — message de chat optionnel (présent uniquement quand prompt est fourni et que l’utilisateur envoie un message)