Module 7: Grounding

An AI model can sound confident while being completely wrong. Grounding is the practice of tying every claim the agent makes to a verified data source.

The Problem

You ask the agent: "What's the liability cap in this contract?"

Without grounding, the agent might generate a plausible number based on patterns it's seen in its training data. It sounds right. It might even be close. But it's not from the actual contract.

In contract review, a hallucinated number isn't a minor error. It's a business risk.

Grounding vs RAG

RAG and grounding are related but different:

• RAG answers the question: "Where does the agent get its information?"
• Grounding answers the question: "Is the agent actually using that information instead of making things up?"

You can have RAG without grounding (retrieve documents but don't verify the agent cited them correctly). You should never have RAG without grounding in production.

Grounding Techniques

Source citation. Require the agent to cite the specific document and section for every claim. "The liability cap is $100,000 (Source: MSA Section 8.2, page 14)." If the agent can't cite a source, it should say so.

Quote extraction. Instead of letting the agent paraphrase, require it to include the exact quote from the source document. You can programmatically verify the quote exists in the original text.

Confidence scoring. Ask the agent to rate its confidence. "How confident are you that this answer comes from the provided documents? (1-5)." Answers below a threshold get flagged for human review.

Verification step. Add a separate prompt chain step that takes the agent's claims and checks each one against the source documents. This is a grounding validator that catches hallucinations before they reach the user.

In Our Contract Workflow

Compliance Agent Output:
  "The indemnification clause in Section 7.3 deviates
   from our standard template. The contract caps
   indemnification at $250,000 (Source: Contract, §7.3,
   p.12). Our minimum requirement is $500,000
   (Source: Policy CB-LEGAL-042, §3.1)."
                     ↓
Grounding Validator:
  ✓ Contract §7.3 exists and mentions $250,000
  ✓ Policy CB-LEGAL-042 §3.1 states $500,000 minimum
  → Output is grounded. Pass to next step.

If the validator can't find the cited source, the output gets flagged and the agent is asked to regenerate with correct citations.

Implementation Approaches

Approach	Complexity	Reliability
Prompt-based ("always cite sources")	Low	Medium - model sometimes ignores
Structured output with source fields	Medium	High - enforced by schema
Separate validation chain	High	Highest - independent verification

For production systems handling high-stakes decisions (contracts, medical, financial), use the separate validation chain. The cost of an extra inference call is trivial compared to the cost of acting on hallucinated information.

What's Next

Grounding prevents the agent from making things up. In Module 8: Guardrails, we cover how to prevent the agent from doing things it shouldn't.

Premium

Grounding Lab

Build a grounding validation pipeline with source citation extraction, quote verification, and automated hallucination detection.

Sign InRequest Access