Grounding is the process of anchoring a language model's outputs to verifiable, external sources of information rather than relying solely on patterns memorised during pretraining. A grounded system retrieves relevant documents, database records, or tool outputs at inference time and uses them as evidence when composing a response, so claims can be traced back to a specific source. The practice directly counteracts hallucination — the tendency of language models to produce plausible-sounding but factually incorrect content — by ensuring the model is responding to real, current data rather than statistical approximations of past data.

Think of it like the difference between asking someone to answer a question from memory versus handing them the relevant file and asking them to answer from that. The second person — grounded with the document — gives a more accurate, auditable answer, and you can check their work by reading the same file.

A grounded response pipeline typically looks like this:

User query
  → retrieval step (vector search or keyword search over docs)
  → top-K chunks injected into the prompt as context
  → LLM generates answer citing the retrieved passages
  → response returned with source references

Grounding is most commonly implemented through RAG (retrieval-augmented generation), which is the dominant architectural pattern. It also connects to Vector Database (which stores the indexed document embeddings), Hallucination (the problem grounding solves), Embedding (how documents and queries are matched), and Context Window (which limits how much retrieved evidence can fit in a single prompt).

You will encounter grounding requirements in any production AI application where factual accuracy matters: legal research tools, medical information systems, enterprise knowledge bases, and customer support bots that must cite a company's actual policies rather than inventing them.