About this talk
This talk by Stefan Begodo delves into the integration of AI capabilities within the Sirius Web ecosystem, which focuses on developing modeling tools. The speaker outlines the aim to provide compatibility with various models and to enhance functionalities through AI agents. He emphasizes the approach of quickly producing concrete use cases, allowing for experimentation and best practices to emerge within the community. Detailed discussions include how the internal AI agent can leverage data from projects, allowing users to interact with modeling tools through a chat interface, customizing interaction based on specific domains like SysML and Arcadia. The session also highlights plans for improving user experience and expanding capabilities for external AI integration, including features for undoing actions, context-aware conversations, and easier authentication for external agents.
Full transcript
[music] >> Hi everyone. My name is Stefan Begodo. I'm a software architect at Obeo and I'm leading the Sirius Web project and I'm making sure that everybody in the Sirius Web ecosystem move forward together. So, over the past year over the past year, we have started working on AI integration in the Sirius Web ecosystem and today I will show you what has been done and what's coming
next. So, first of all, let's have a look at what we are trying to accomplish. Of course, we want to ensure that Sirius Web based modeling tools of our ecosystem can integrate AI based features. We don't want to pick any favorites. Uh so, we want to be compatible with any model that our customers may want to use. We also don't want just to provide a connection to
LLMs. We want to provide additional modeling capabilities using AI agents. This work is being done thanks to uh the Sirius Web extension Obeo Enterprise for Sirius which can be added to any Sirius Web based application. Our current strategy aims at quickly producing concrete use cases that can be tested by hand users. For that, we are moving forward with multiple experimentation to build a set of best practices
for us and our community. So, some of the APIs and patterns that we are using may change quite frequently while we are using this approach. We also experimenting with our partners on the integration of multiple AI agents focused on specific task to create powerful agentic workflows. So, right now we are moving forward in two directions. On one hand, having the Sirius Web ecosystem acts as an AI
agent. This agent would be within the and we would be entirely in control. On the other hand, we are also working on MCAP API to let external agents query and manipulate the Capella web-based application. In this case, the application would just be one provider of data and tools among others. So, let's have a look at the work done with Capella web-based application acting as the AI agent.
So, as I said, using this approach, we are in control and by we, I don't just mean us at Obeo in the core team, but we the developer of Capella web-based We provide the user interface, can integrate it anywhere we want in the application. For that, we started with a chat interface in the workbench since this is what most users are familiar with. We are in control
of the prompt too. The end user can, of course, provide the main message, but we can decorate it with additional data. Finally, we can provide a deep integration within the platform thanks to specific internal tools. This internal agent can simply talk to LLM like ChatGPT, Gemini, or Claude, but it can also act as a proxy for remote These remote agents can have their own tools and data
and they can be configured for specific task while we, on the other hand, have the modeling know-how. Regardless of this, this internal AI integration can provide tons of data about your modeling tools to the agent. It can be customized with data derived from the nature of the project, UML, SysML, Arcadia, uh the data of course inside of the project, but also uh the state of your current
workbench. For example, the representation which is visible on screen, the current selections among other things. So, let's see how it behaves. here you can see uh how the internal agent will leverage the data of the workbench. Uh on screen you have one of four large uh example that we use in uh the Sirius Web ecosystem. It's the entire architecture of Sirius Web, all our classes, interfaces, everything
from the source code of Sirius Web represented within Sirius Web. So, here I can ask the LLM uh some information about the current state of the project and you can navigate inside all of those data, but we can also pinpoint a specific selection and start uh querying the internal agent about this uh this specific data. And here we have the answer of the LLM just focused on
one specific element and we can start working from that element. For example, if I want to know all the concepts that are linked with the current selection, things like that. We are not limited to what is displayed in the representation. We also have access to uh the semantic data behind the diagrams. Our internal integration also allowed downstream project to customize pretty much everything from the system prompts
to the execution mode and tools. We are providing some default, for example, some default execution mode, query and edit, and some default tools to query or edit the models, for example, but you can add new ones or even remove existing ones if you want. For the system prompt, we can have several level of customization. For example, the generic low-level technical details, which can be refined with domain-specific
details with a higher level of abstraction. And on top of that, you can come with your own methodology and uh have something which is closer to the way your end user really behave. To give you an example, in uh Obeo Enterprise for SysML, we are providing the generic low-level details. If you're using the SysML version, you would instead have uh additional information for the SysML domain. And
with the Arcadia extension for SysML that we are building, which is a currently a proof of concept, you would have not only those details, but uh details about the the specific Arcadia methodology. In more concrete terms, it would be going from uh interacting with generic EMF concepts, for example, EClasses, EAttributes, to SysML V2 uh concepts like part or part definition, to the day-to-day vocabulary of system engineers
with systems, components, and functions. We can't have uh system engineers interact with uh an AI agent using EMF concepts or even raw SysML V2 details. This is not our work. They don't care about the difference between owning membership and feature memberships, which are tiny implementation details of the SysML V2 uh domain. For the internal tools, we are in charge of what is being contributed. Specifiers can add
additional tools, and we can give them a quick and easy way to retrieve the context of the end user. The current project, the current selection, all representations that are being opened. So, if we have a look at our Arcadia uh extension, so here I'm uh opening a blank Arcadia project. I'm selecting a specific package because I really want to work on that uh logical architecture package. So,
I'm starting the discussion and the LLM tells me that I've selecting the logical architecture package and it offers me some option. And I want to create the logical architecture of a modern smartphone. Right now, since we are working in the edit execution mode, the LLM will ask uh will offer me to start planning what I want to do. So, it will give me some uh a description
of what it will do. The description is a bit verbose, but it uh explains everything that will be done and it ask uh two things. It will ask me to both confirm that I'm okay with the plan. We can iterate on the plan before starting the execution and it will also ask me to select uh various option for the plan. For example, uh do I want a
hardware focused uh logical architecture or more software focused one? So, once I'm okay with it, it will perform the changes, both semantic changes and graphical changes in the representation that I've opened. And it will tell me everything that has been done. So, here because I have this specific diagram open, it has created those uh graphical elements in this uh And here, I've interacted with the LLM using
the vocabulary and methodology of system engineers. So, it has created me it has created for me components with sub subcomponents, logical functions, functional exchanges, things like that. And this is a vocabulary used inside of the discussion. We also worked on various uh user experience improvement. For example, to support batch modification for semantic and graphical update. So, in the previous video, we will we were able to create
uh several elements at once thanks to the LLM, which is way faster than just creating uh things one by one. We also worked on lowering the quantity of tokens consumed by the tools because these are expensive. And now the tools that are contributed can also uh stream some feedback back to the end user. The other approach we are working on is having the Sirius web-based application act
as an MCP server, which can be manipulated by an external agent. It is authenticating using your credentials and then can respect the permission and capabilities of the current user. If you can't see a project, you won't be able to see it with the MCP API. If you can see it but only read only, you won't be able to edit it. This can work with uh popular AI
agent like Kodak, Gemini CLI, Cloud Card, and so on. And this way you can work on a wider context. For example, you can work on multiple projects at once and we are providing MCP tools with this time additional tools to work on this wider context. Search projects, create new one, create organization. You could pretty much administer your uh Sirius web instance with this. Part of the work
I've shown you today uh is a result of uh collaboration between OBEO and CEA-List in a joint laboratory on AI. In this laboratory, we are exploring the integration of AI agent for a better understanding of your data, but also for the onboarding of new user and the day-to-day assistance while you are using your modeling tools. These works help us add new capabilities upstream in various parts of
the Sirius To conclude, let's have a look at what we are currently working on and what will be coming over in the next few months. In terms of user experience, we want to provide support for undo and redo. LLMs are great at generating tons of stuff, but if they mess up, if uh what you believed they would do uh while they are showing the plan is not
what you were expecting, it is useful to be able to tell the LLM to just uh undo the changes. It's also useful to be able to tell the LLM to try something, evaluate this change, roll it back if it does not match, and try something else. We are also expanding the support for other representation. Right now, we were focused mainly on We also working uh on improvement
for the internal agent part with uh for example, a work on follow-up question because it may be a bit intimidating to figure out what we should ask the LLM at first. We also want to offer some personalized content in the conversation by leveraging the proficiency and the history of the user in the project. If you are a newcomer on the project, or if you've been working with
it daily for the past 2 months, we won't have the same conversation. We also want to be able to pin the context of the chat client in the front end, just like you can do it with other contextual views in the Sirius Web Workbench. It would give you the ability to view more clearly the contextual data that the agent is using and have a stable context while
working with other elements in the workbench. For the MCP part, the main thing we will provide over the next few months will be the support for OAuth 2 authentication to simplify the setup along with the support for MCP resource to work more easily with the external agents that supported. For example, uh the Copilot uh agent inside of VS Code. And that's it for today. So, if you
have any question, don't hesitate or you can just send me an email. Thank you. >> [music]
