Grounding

Grounding is any technique that ties a language model's output to verifiable sources — retrieved documents, tool results, structured data — instead of pure memory.

Grounding is the umbrella term for reducing hallucinations by giving the model evidence to anchor its answer. RAG is the most common grounding pattern. Tool-use grounding (calling a real API for a stock price, weather, search) is the strongest form because the data is fresh. Prompt-side grounding (paste the contract before asking about it) is the simplest. Grounded answers should always cite the source ID so downstream code can verify the model used the supplied evidence and didn't drift.

When to use grounding

• Any factual query where being wrong is costly.
• Compliance and regulated industries.

Common mistakes

• Providing grounding documents without telling the model to use ONLY them.
• Forgetting to make the model cite source IDs — losing auditability.

FAQ

Related terms

• Retrieval-augmented generation (RAG) — Retrieval-augmented generation (RAG) injects relevant documents into the prompt at query time so the model answers from your data instead of its training memory.
• Hallucination — A hallucination is when a language model produces output that is factually wrong, fabricated, or unsupported, while sounding confident.
• Function calling (tool use) — Function calling lets a language model emit a structured request to invoke a developer-defined tool, enabling reliable JSON output and agent workflows.

Last updated: 2026-06-01. Raw markdown: https://promtable.com/glossary/grounding.md.