Siemens manages codebases measured in hundreds of millions of lines, some written before the 2010s, across factories, energy grids, and transit systems. Standard LLM coding assistants choke on that scale - context windows can't hold a fraction of it, and vector-only RAG loses the structural relationships that define industrial software.
So Siemens and Google Cloud built a different animal: Knowledge Fabric, an agentic system that treats code as a graph. Spanner Graph models every file, class, function, and documentation artifact as nodes with explicit edges. Agents don't just search text; they traverse relationships using Graph Query Language (GQL), Approximate Nearest Neighbors (ANN) vector search, and full-text search in combination. When an engineer asks "Which functions break if I change the Axis Control Panel?", the system doesn't hallucinate - it walks the dependency graph and shows receipts.
Four Dimensions of Legacy Hell
The challenge Siemens faced is familiar to anyone who has worked on enterprise software older than a startup's entire existence, but the numbers are staggering. Repositories exceed LLM context windows. Critical knowledge lives in scanned PDFs from 2003, Jira tickets, and Confluence pages - no single retrieval method can cover it all. Tracing a line of code to a 10-year-old functional requirement manually takes days. And the consequences of an unvalidated change are not just a broken build; they can shut down a factory floor.
Agata Gołębiowska, technical lead at Google Cloud, put it bluntly: "Code isn't just text; it has inherent structure. Flattening that into a vector database meant losing the representation of relationships." That's why Knowledge Fabric uses Spanner Graph to maintain the actual architecture, augmented with per-node embeddings for semantic search. Agents can ask "How does this module connect to that subsystem?" and get back a subgraph, not a ranked list of chunks.
Slicing the Elephant (Not Eating It Whole)
Franz Menzl, senior VP at Siemens, makes the analogy explicit: you don't modernize a 50-million-line codebase with one agent prompt. The team adopted a pattern they call "slicing the elephant" - decomposing a vague request like "refactor this module" into specialized agent steps, each built with the Google Agent Development Kit (ADK).
A Search Agent explores the graph and documentation. A User Story Agent interviews the product owner to define acceptance criteria. An Architecture Impact Agent predicts side effects before any code changes are written. Only then does a Task Breakdown Agent produce small, self-contained units of work, each with full context attached. Finally, a Coding Agent implements each task - but only after all the upstream analysis is done. Alexander Lomakin, project lead at Siemens, reports that this "roadmap" is what makes the AI's output usable, not just plausible.
Pilot Results: Days to Minutes
In a production pilot migrating legacy control panels to web-based interfaces, Knowledge Fabric cut the time for dependency analysis from several days of senior engineer effort to far less (they don't quote exact hours, but the implication is clear: minutes, not shifts). The system preserved full system compatibility and industrial quality standards. Engineers now spend their time on customer-facing innovations instead of spelunking through ancient code.
This isn't a generic code assistant. It's a purpose-built graph-agent pipeline that combines structural, vector, and keyword search with human-in-the-loop gating. The same architecture could apply to any large enterprise codebase where understanding relationships matters more than generating boilerplate. Siemens and Google Cloud have shipped a concrete pattern, not a vaporware demo - and they've open-sourced the Agent Development Kit and made Spanner Graph available for anyone to try.
Source: How Siemens "slices the elephant," advancing agentic workflows for industrial software development
Domain: cloud.google.com
Comments load interactively on the live page.