Grounding (AI)

A language model trained on internet data has broad knowledge but no way to distinguish its accurate knowledge from its inaccurate knowledge. Grounding gives the model a factual anchor. Instead of generating from memory, it generates from sources you provide.

The most common grounding technique is RAG. You retrieve relevant documents and include them in the prompt, then instruct the model to answer only based on those documents. This turns the model from a knowledge source into a reasoning engine that works with your data.

But grounding goes beyond just retrieval. It also includes verification. After the model generates a response, you can check whether the claims in the response actually appear in the source documents. This is sometimes called attribution or citation verification. If the model says "Our return policy allows 30-day returns," you check that this claim exists in the retrieved policy document.

Enterprise grounding often involves multiple layers. The first layer retrieves relevant context. The second layer instructs the model to cite its sources, often by including span IDs or passage numbers in the prompt and requiring the model to reference them. The third layer programmatically verifies those citations, typically with a smaller model or an NLI (natural language inference) classifier that scores each claim against the cited passage. The fourth layer flags ungrounded claims for human review or rejection.

Well-grounded AI systems are more trustworthy and auditable. When every claim can be traced back to a source document, you can explain why the system said what it said. This matters for compliance, customer trust, and internal confidence in the AI system.

Grounding has limits. It works when the answer exists in your sources. It fails when the user asks something your corpus doesn't cover, which is when models drift back to their training knowledge and start hallucinating. The cure is less about the model and more about the prompt design: instruct the model to say "I don't have information on that" when the retrieved context doesn't support an answer. Measure refusal rate in production as carefully as you measure accuracy. A model that answers everything is hiding hallucinations behind confidence.

A commercial insurance underwriter uses an AI assistant to answer coverage questions for broker partners. A broker asks, "Does the small-business property policy cover food spoilage from power outages under 12 hours?" The assistant retrieves the top 5 passages from the current policy PDF and forms guide.

Claude Sonnet is prompted to answer using only the numbered passages with inline citations. It responds: "Food spoilage from power outages is covered if the outage is caused by a covered peril [passage-2]. Coverage requires the outage to last more than 12 continuous hours [passage-4]. So outages under 12 hours are not covered." Three claims, three citations.

A post-processor runs a DeBERTa NLI classifier on each claim against the cited passage. Claim 1 and claim 2 score 0.91 and 0.87 entailment, which pass. Claim 3 is a logical deduction from claims 1 and 2 rather than a direct quote, and scores 0.62. The orchestration layer flags claim 3 as "inferred" and the response goes back to the broker with a visible note: "Based on passages 2 and 4, outages under 12 hours appear not to be covered. Confirm with your policy specialist." The broker has a verifiable answer with clear provenance and a built-in check for the logical leap the model made.