Siemens, a leader in industrial AI and automation, has developed a groundbreaking AI system called Knowledge Fabric in collaboration with Google Cloud. This innovative solution aims to modernize the extensive legacy code that underpins industrial software, which is critical for factories, energy grids, and transportation networks worldwide.
With codebases comprising hundreds of millions of lines, Siemens faced significant challenges that traditional AI tools could not address. The complexity and scale of industrial-grade software necessitated a new approach to understanding and updating these systems.
Understanding the Challenge
Modernizing large-scale industrial software is often likened to rebuilding a jet while in flight. Siemens identified four primary challenges:
- Scale: The size of the repositories far exceeds the capabilities of standard large language models.
- Fragmentation: Important information is dispersed across various platforms, including code, Jira tickets, and outdated documentation.
- Complexity: Linking specific lines of code to functional requirements from years past presents a significant hurdle.
- Responsibility: Systems must comply with strict quality and lifecycle standards, requiring AI outputs to be explainable and verifiable.
Agata Gołębiowska from Google Cloud emphasized that traditional retrieval-augmented generation methods were insufficient, as they failed to capture the inherent structure of the code.
Introducing Knowledge Fabric
To navigate this complex software landscape, Siemens and Google Cloud created the Knowledge Fabric agent. This system utilizes knowledge graphs to model the relationships within the codebase, allowing for a more intuitive understanding of how different components interact.
By employing Spanner Graph, the teams were able to link code snippets directly to design requirements. The integration of Graph Query Language (GQL) and semantic understanding through embeddings further enhances the agent's capabilities.
Breaking Down Tasks
A key insight from the project was the need to simplify complex tasks. The team adopted a method termed "slicing the elephant," which breaks down large requests into smaller, manageable tasks. Each task is handled by a specialized agent, including:
- Search agent: Conducts deep research within the code graph.
- User story agent: Gathers requirements from product owners and drafts user stories.
- Architecture impact agent: Analyzes proposed changes and predicts their effects.
- Task breakdown agent: Divides the work into manageable tasks.
- Coding agent: Implements the changes based on the analysis.
This structured approach ensures that human oversight is maintained throughout the process, leading to reliable and efficient outcomes.
Results from the Pilot
Initial results from the pilot program have been promising. Tasks that previously took senior engineers several days can now be completed in significantly less time. In a recent pilot focused on migrating legacy control panels to modern web interfaces, Knowledge Fabric reduced coding efforts while ensuring compliance with industrial standards.
As a result, engineers are now able to dedicate more time to creating value for customers rather than getting bogged down in repetitive tasks.
Next Steps for Organizations
The success of Knowledge Fabric illustrates the potential of generative AI in modernizing legacy systems. Organizations looking to implement similar solutions can explore the following:
- Learn about Spanner Graph for knowledge modeling.
- Investigate the Agent Platform for pre-built agents.
- Utilize the Agent Development Kit for custom solutions.
