Research Lab
What It Takes to Ground an LLM in a Fab
Research note: retrieval, citations, and the trust architecture required before process engineers will use an AI assistant.
- LLMs
- RAG
- trust
Process engineers are professionally skeptical. Their entire job is distrusting plausible explanations until the data confirms them. Shipping an LLM assistant into that culture teaches you what "grounding" actually requires.
Citations are not a feature, they're the product
The single design decision that mattered most: every claim in every answer links to the exact source paragraph. Not a document, a paragraph.
An engineer doesn't want the answer. They want the answer and the ability to verify it faster than finding it themselves would have taken. That's the value equation, and citations are the entire right-hand side.
Retrieval quality is where answers die
Nearly every bad answer traced back to retrieval, not generation. Three patterns:
- Tables shred. Naive chunking destroys spec tables, and specs are half of what engineers ask about. Table-aware chunking was worth more than any model upgrade.
- Vocabulary drift. Engineers say "chamber A dogging" while the runbook says "throughput degradation, chamber A." Hybrid retrieval (dense + BM25) covers each side's blind spots.
- Recency conflicts. When two document versions disagree, the model must say so: silently picking one is how you lose trust permanently.
The trust ladder
Adoption followed a repeatable sequence: engineers first used it as a search engine (ignore the prose, click the citations), then as a summarizer (trust the prose, verify occasionally), and only then as an assistant (ask follow-up questions). Designing for that ladder, instead of assuming assistant-mode trust on day one, is, I suspect, general to every high-stakes LLM deployment.