Grounding

Grounding is the discipline of making an AI answer only from a verified body of knowledge, cite the source behind every claim, and say what it could not find instead of guessing. This page is about the document sense of the word: whether an answer is grounded in your drawings, specifications, and regulations, not the robotics sense of a model grounded in the physical world.

What does it mean for an AI answer to be grounded?

An AI answer is grounded when every claim in it traces back to a specific passage in a known, verified source, and when the parts that are not in those sources are left out rather than invented. A grounded answer carries its evidence. You can follow each statement to the document it came from, and when the documents do not support a statement, a grounded system does not make one up.

Ungrounded generation does the opposite. It produces the most plausible-sounding continuation of the question, drawn from a blur of training data, and it sounds exactly as confident when it is guessing as when it is right. For brainstorming, that is fine. For a decision where a wrong answer is a hazard, it is the whole problem.

Is grounding just RAG?

Grounding is the goal. Retrieval-augmented generation is one technique for reaching it, and on its own it usually falls short. Ordinary RAG splits a document into fixed chunks, stores them, and returns the few that look most similar to the question. That works until the meaning of an answer depends on more than one chunk, which in technical documents is most of the time.

The failure is not that the retrieved pieces are wrong. Each piece is a real sentence from the real document. The failure is that meaning does not survive being torn out of its surroundings. A clause depends on the heading above it, the sentence before it, and the exception three pages later. Grounding keeps an answer tied to its real place in the source, with the structure and the conditions that make it true still attached. Retrieving lookalike fragments and hoping the right ones come back is not the same thing.

Why can’t a bigger context window solve this?

Because the corpus is larger than the window by orders of magnitude, and it grows while the window does. Our co-founder ran an edtech company that needed to reason across 8,000 documents. The most capable model available, at its largest context window of 2.5 million tokens, could hold 1.6 percent of that corpus at once. Not 60 percent. Not 16. One point six.

Do the multiplication forward. Make context windows ten times bigger and you reach 16 percent. A hundred times bigger and you still cannot hold a mid-sized company’s knowledge, let alone the full documentation of an offshore platform or a hospital. Grounding is what lets an AI reason across a corpus far larger than any window, which is why waiting for bigger models does not close the gap.

What does grounding look like when it matters?

A maintenance manual says, in one paragraph, that a component is safe to handle. Twelve pages later, in a different section, it says the rest of the rule: safe to handle, provided the component has not been exposed to sunlight. Both statements are true. Together they are the rule. Split apart and retrieved separately, the first one alone is a confident lie that reads exactly like a fact.

An engineer asks whether the component is safe to handle. A system that retrieves the nearest-looking chunk returns the first sentence and answers yes. It is not hallucinating. It is grounded in a real sentence from the real manual. It simply never saw the condition, because the condition lived in a different chunk that did not come back. That is the failure grounding exists to prevent: keeping the fact bound to the condition that changes it. In the industrial world, that is the difference between a correct answer and a dangerous one.

What happens when the answer is not in the documents?

A grounded system says so. If the ingested documents do not contain the answer, the grounding service replies that it cannot answer from what it has, and it names what it would need. It does not reach out to the open web to save face, and it does not assemble a plausible-sounding answer from unrelated neighbors.

This is the most valuable behavior an expert has, and the one AI has been missing. The most useful thing a real expert tells you before a decision is not the answer. It is “here is exactly what I would need to know to be sure.” The grounding service is built to answer from fact or to name the gap, never to fill the gap with a guess.

How do you know the AI used enough of the corpus?

Because the grounding service reports it. Every answer comes with an account of what was and was not included, so a person or a system downstream can see the coverage behind the answer rather than trust that it was complete. Ordinary retrieval cannot prove it pulled enough, which is a quiet but serious gap when the cost of missing one clause is operational, not cosmetic. Grounding treats sufficiency as something to demonstrate, not assume.

Why grounding matters most in mission-critical work

In most software, a wrong answer is an inconvenience a person catches later. In industrial, energy, and offshore operations, a wrong answer is operational risk: a bridge that does not connect, a procedure run against the wrong revision, an inspection signed off on a clause that was never retrieved. These buyers cannot put an AI answer in front of a regulator, an inspector, or an operations team without a person re-checking everything behind it, and that re-checking is the real ceiling on enterprise AI. Grounding is the layer that lifts it. It is what the grounding service from Lodestone Labs is built to provide.

The grounding service has been demonstrated reading real engineering documents, including a pump-performance module where it answered correctly and flagged what the document did not state. It has not yet run on a full production customer corpus. We say so plainly, because the claims rule applies to how we describe grounding as much as to what a grounded system says. See how it works, or get in touch about proving it on your own data.

Hallucination: what an ungrounded system produces, and the failure modes grounding is measured against.
Traceability: being able to follow an answer back to the source it came from.
Provenance: where a fact originated, and in what form.
Ground truth: the verified knowledge an answer is grounded against.
DSAI: the compact, domain-specific expert that grounding makes possible.