Grounding

What it is

Grounding constrains a generated answer to evidence that was actually retrieved, so each statement can be traced back to a source. That is the strict sense, and the one this register uses. It is narrower than the word as vendors ship it. Google defines grounding as the ability to connect model output to verifiable sources of information, and lists several grounding types under that heading, from a search tool called at inference to a managed RAG service. Amazon Bedrock Guardrails and Azure AI Content Safety apply the word to a check made after generation, scoring a finished response against a reference source rather than constraining how it was produced; Bedrock’s contextual grounding check runs on the output and not on the prompt. AWS uses the word both ways, though, also documenting grounding as what Bedrock Knowledge Bases do. The mechanism is therefore not consistent across implementations, or even within one vendor, and a mapping from sentence to source is a property of some of them rather than all.

Why it matters

Grounding is what turns a plausible answer into a checkable one, and it is the main lever teams have for reducing hallucination without retraining anything. Its ceiling, though, is the corpus. Grounding against documents that contradict each other produces confident answers drawn from the wrong source — the citation is real, the answer is wrong, and the system reports success. That ceiling is why reconciliation comes first. It is also why Knowledge Company works on the corpus rather than the retriever.