DEEP DIVE · CAPABILITY 05

Half your plant's knowledge was never a sensor.

Alarm thresholds, procedures, setpoints, troubleshooting steps, safety limits — they live in manuals, not historians. Foundation makes that knowledge queryable by people and AI alike, with provenance strong enough for a control room.

  1. GroundedAsk about a piece of equipment and retrieval is scoped to its documentation, its model, its revision.
  2. Hybrid searchpart numbers, tag names, and error codes are found as reliably as concepts
  3. ProvenanceEvery answer can cite document, version, page — and the bounding box on the page it came from.
  4. Safe for agentsagents retrieve through the same permissions as the person asking, never around them.

The problem with "just add RAG"

Generic document AI retrieves what sounds similar and answers with confidence either way. In an office that's an inconvenience; in a plant it's a hazard — the wrong setpoint, the right procedure for the wrong pump model. Two things fix it, and both are structural: grounding in the Domain, and provenance you can inspect.

CITED TO THE BOX ON THE PAGE

COMPRESSOR K-401 — OPERATING LIMITS Manual v4 · p. 412 What is K-401's max discharge pressure? 32 bar do not exceed for more than 30 s Manual v4 · p. 412 · region click the citation — the page opens, region highlighted

Grounded in the twin

Every document in Foundation is tied to the assets it describes — in the graph, where "the export compressor's manual" is an actual relationship, not a filename convention. Ask about a piece of equipment and retrieval is scoped to its documentation, its model, its revision. The Domain does for documents what it does for data: turns a pile into a system.

Retrieval built for technical material

Hybrid search, not just similarity: semantic matching fused with full-text search, so part numbers, tag names, and error codes are found as reliably as concepts — the exact places where similarity-only retrieval fails on engineering documents. Versioning and canonical-document deduplication keep superseded revisions from masquerading as current truth.

Provenance to the exact spot on the page

Every answer can cite document, version, page — and the bounding box on the page it came from. In AI Crew, Evan's citations are click-through: the PDF opens at the page, region highlighted. An operator never has to take the AI's word for a threshold; the manual is one click away, open at the line. For regulated operations, this is the difference between an assistant and an audit risk.

Safe for agents by design

Responses are budgeted — hard caps on retrieved pages and tokens — so a 4,000-page vendor manual never floods an AI's context. And access control is enforced at the API layer: agents retrieve through the same permissions as the person asking, never around them.

Where is your plant on the scale?

Thirty minutes with our team, and you'll know your plant's level, the constraint holding it there, and what Level +1 is worth to you.

Find your plant's level Book a level assessment