LLM grounding
LLM grounding is giving a large language model retrieved sources at answer time, so that its answer rests on those sources and can be checked against them instead of coming from training data alone. The sources usually come from a search API, a document store or a database, and the model is asked to answer from them and cite them.
Grounding is the core idea of retrieval-augmented generation (RAG), described by Lewis et al. (NeurIPS 2020). Cloud vendors sell it as a named feature: when Microsoft retired the Bing Search APIs in 2025, it pointed customers to Grounding with Bing Search for Azure AI agents.
Example
Asked for a company’s current CEO, a model without LLM grounding may name the person who held the job when its training data was collected. With grounding, the model first searches, reads a press release from last month and answers with the new name and a link to that release.
FAQ
Is LLM grounding the same as RAG?
Almost. Retrieval-augmented generation (RAG) is the most common way to ground a model: it retrieves documents and adds them to the model’s prompt. LLM grounding is the broader term for any answer built on sources supplied at answer time.
Does LLM grounding stop hallucinations?
No. Grounding reduces invented answers, but a model can still misread a source, mix facts from two sources or go beyond what the sources say, so citations still need checking.
Related terms
Written by Keenable.
Published
