How it works
Infrastructure, not another AI tool
The grounding service sits underneath the AI your team already uses. Your application calls it before it asks a model to answer, and it returns evidence, the scope that evidence covers, and the citations behind it. Or it returns a qualified refusal, which is the part most systems are missing.
What does the grounding service actually do?
The grounding service takes a body of documents you have decided are authoritative and makes them answerable. It ingests them whole, keeps their structure intact, and answers questions from them with every claim traced back to the page and the revision it came from. It does not replace your model, your application, or your interface. There is no end-user product to adopt, only an admin panel for monitoring what has been ingested and what has been asked.
It works with whichever model you already run, whether that is Gemini, Claude, GPT, or an open model on your own hardware. The model still writes the sentence. What changes is that the facts underneath the sentence came from a source you can open, and the gaps are named rather than filled.
-
01 · Admit and preserve
Ingest
Your documents go in whole. The service reads layout and structure first, so a heading still governs the section under it, a table stays attached to the paragraph that refers to it, and a clause keeps the exception that qualifies it. Nothing is split into fixed-size chunks, because that is the step where most of the meaning is lost.
What comes out is a structured record of the document rather than a pile of fragments: the parts, what each part is, how the parts relate, and where each one sits in the original. The expensive work happens here, once, which is why answering later is fast.
What this stage guarantees
- Structure, tables, and cross references survive ingestion.
- Every part keeps a route back to its exact place in the source.
- A new revision is added, not substituted. Superseded versions are kept.
-
02 · Find everything that bears on the question
Retrieve
When your AI asks a question, the service returns every part of the corpus that bears on it, together with the references, exceptions, and revisions that qualify those parts. It is a lookup against a structure, not a similarity search against fragments, so the passage that decides the answer does not have to also be the passage that most resembles the question.
That difference is the whole argument. Ordinary retrieval returns what sounds like the query. A specification says a component is rated for a given load, and twelve pages later says the exception that applies under a particular condition. Those two statements are one fact, and an answer that carries only the first is grounded in a real sentence and still wrong.
What this stage guarantees
- A claim is not returned without the conditions that qualify it.
- Retrieval covers the whole admitted corpus, not a top-scoring handful.
- Nothing is returned from outside the corpus you declared.
-
03 · Mark what a machine should not decide alone
Validate
Real document sets contain material that cannot be read cleanly: old handwriting, scans, photographs of drawings, engineering tables, and pages that do not fit the structure around them. Most systems make a guess at these and pass the guess downstream, where it is indistinguishable from a fact.
The grounding service marks them instead. It flags the exact place in the document that needs a human decision, and a person accepts, rejects, or interprets it before that part is trusted. It is the slowest thing on this page and it is deliberate: it is the difference between a corpus you can rely on and one that quietly contains inventions.
What this stage guarantees
- Unreadable material is flagged at its exact location, not silently interpreted.
- A person decides before an ambiguous part becomes evidence.
- The review trail is kept, so a decision made once can be checked later.
What comes back when your AI asks a question?
An answer grounded in your sources, with a citation on every claim that resolves to the document and the revision it came from. Alongside it, a coverage report: what the answer was built on, what it was not, and whether the corpus was sufficient to answer the question at all.
The coverage report is the part worth arguing about. Ordinary retrieval cannot tell you whether it read enough, because returning the best matches is the entire contract it is built to fulfill. That leaves you unable to tell a question that was answered completely from one that was answered from a sliver, since both arrive looking the same. This failure has a name, silent omission, and it is the one that matters most in work where a missed clause is operational rather than cosmetic.
When the corpus genuinely does not contain the answer, the service says so and names what it would need in order to answer. It does not reach out to the open web, and it does not assemble a plausible answer from unrelated neighbors. That behavior is the most useful thing an expert does before a decision, and it is the thing general-purpose AI has been missing.
Where does the grounding service run?
As an isolated instance, on the cloud and in the region you choose. Your documents stay inside your environment, and the service can run entirely inside your own infrastructure where that is what your policy requires. Because a grounded domain does not need a frontier-scale model to answer well, the footprint is small enough that where it runs is genuinely your decision rather than ours.
The deployment and security page covers tenancy, retention, what leaves your environment, and what a model provider does and does not see.
The fastest way to find out whether this holds on your documents is to point it at a narrow slice of them and the questions your current tools answer badly. That is what a teardown is, and it is short by design.
How a pilot works What deterministic grounding means Talk to us