Why We Parse Industrial Code Instead of Embedding It

Most industrial AI applications consist of a retrieval pipeline with a chat interface. The manuals are chunked, embedded, and placed into a context window; the top matches are then presented to the language model. While this approach demonstrates well, it breaks down when the question's answer relies on the current state of the machine.

Nodeblue, an applied research lab, developed the Nexus system, which focuses on a different architectural decision: the language model is merely the smallest and least interesting component. The core issue was revealed through a test that compared eleven frontier models against a live production controller with a live edit present in the processor but absent from the archive.

All eleven models confidently provided incorrect answers, relying solely on the archived version they had never seen before. When the same models were used on the Nexus engine, they accurately cited the correct rung. The models did not enhance their performance; they simply had access to a fact residing in the processor rather than a corpus.

Parsing industrial control programs has four layers: deterministic parsing of ladder logic, structured text, AOIs, UDTs, tag databases, and device configurations into a resolved model. Every claim made by the model includes the corresponding rung, tag, or document, enabling professionals to verify the output. The system processes real exports from live floors, including files too large for chat interfaces, and employs OPC UA reads and subscriptions to access the running controller.

This layer, involving physical, on-premise, and safety-gated work, is challenging for AI but necessary. Benchmarks conducted by independent controls engineers with no affiliation to the project yielded zero errors, confirming the model's accuracy. The Nexus engine's connectors are publicly available under the MIT license at github.com/Nodeblue-AI.