Grounding

Grounding connects a model's answer to evidence. Instead of asking the model what it knows, you retrieve the relevant passages, records or tool results, place them in the context window, and instruct the model to answer only from that material and to say so when the material does not cover the question. Retrieval-augmented generation is the most common grounding pattern; tool calls into live systems are another.

Good grounding has three parts: retrieval that finds the right evidence, prompting that constrains the model to it, and verification that checks the answer actually follows. The last part is where most teams stop short. Groundedness scoring (does each claim have support in the supplied context?) belongs in the eval suite, and citations shown to users make unsupported claims visible.

Grounding is not the same as fine-tuning, which changes model behaviour rather than supplying facts, and it does not guarantee correctness: if retrieval returns the wrong document, the answer will be faithfully wrong. Permission-aware retrieval also matters, because grounding on data the user should not see is a leak, not a feature.

• CTO / Head of Engineering: grounding is the main lever for accuracy and the architectural decision that shapes the data pipeline.
• Product manager: citations and "not found" responses are product behaviours that grounding makes possible and users learn to trust.
• Compliance officer: answers that trace to a specific source are auditable; answers from model memory are not.
• Data lead: the quality of the grounding corpus (freshness, permissions, structure) sets the ceiling on answer quality.

Grounding exists because a model's built-in knowledge is generic, frozen at training time and unverifiable, while business questions are specific, current and consequential. Supplying evidence at query time solves all three problems without retraining. The trade-off is complexity and latency: you now maintain a retrieval pipeline, an index and a freshness process, and each request does more work before the model answers. It also shifts failure from the model to retrieval, so recall and precision become metrics you must own. For any answer a customer or regulator might challenge, that cost is unavoidable.

• An enterprise knowledge assistant over Drive, Slack and wikis that cites the document behind each answer.
• A conversational banking agent that quotes balances and card terms from live core-banking calls, not from the prompt.
• AI tutors that explain concepts using only the institution's own course material.
• A hospital assistant grounded in current clinical protocols with permission checks on patient records.
• A retailer's WhatsApp commerce bot answering stock and delivery questions from the order and inventory systems.

Technique / practiceA technique at the core of Eazyware's retrieval and knowledge engineering service, and applied in every LLM application where answers must be defensible. It comes with groundedness evals and citation UI rather than as a prompt instruction alone.

Eazyware service that covers it: Retrieval & Knowledge Engineering. Starting prices are on the pricing page.