Teaching software to read like a domain engineer

A general language model knows what a milling cutter is. It has read every catalogue ever indexed. Ask it to recommend an insert grade for an interrupted cut in stainless and it will answer fluently, confidently and sometimes correctly. Fluency is the problem. In an industrial setting, an answer that cannot show where it came from is not an answer an engineer can sign.

Domain knowledge is structured, not remembered

Machining knowledge is not a pile of prose. It is built on designation systems and standards that carry meaning in their structure. ISO 513, to take one load-bearing example, sorts every workpiece material into six application groups, and that single letter changes the cutting material, the geometry and the parameters that follow.

Get the group wrong and nothing downstream can be right. A recommendation that treats duplex stainless as an ordinary steel is not nearly correct. It is wrong, and wrong in a way that costs inserts, machine time and trust. Domain grounding starts by giving the model these structures as structures, not as sentences it may or may not recall.

The anatomy of a knowledge pack

For each industrial domain Kabaido covers, we build what we call a knowledge pack. Eight exist today, from cutting tools to machine tools and CNC. Each one is assembled the same way, in four layers.

1. A corpus of authoritative sources: the standards, manufacturer catalogues and engineering handbooks the domain actually runs on, each recorded with provenance.
2. An authority map: an explicit ranking of which source wins when two disagree, so a marketing page never outranks the standard it paraphrases.
3. Structured tables: designation systems, parameter ranges and limits extracted into typed data the engine can compute against, not just retrieve.
4. An evaluation set: real requests with checked answers, kept apart from everything above, that the pack must pass before it ships.

Evals before answers

The evaluation layer deserves the emphasis. Every pack carries a set of genuine RFQ-shaped requests with worked golden answers, and the test is stricter than getting the value right. A response only passes if it behaves the way a careful engineer would.

• It extracts what the request actually contains, with the exact source span for every attribute.
• It returns null for what the request does not contain, instead of filling gaps from prior knowledge.
• It raises a sensible clarifying question for each material gap.
• Its numbers survive unit conversion, because a thou is not a micron.

None of this makes a model clever. It makes a model checkable, which in industrial work is worth far more. The engineer who signs the quote stays the judge. The knowledge pack just ensures that what reaches their judgement is grounded, cited and honest about its gaps.