Context Grounding

Supply the information the model needs in-prompt, and when required restrict answers to that context so parametric knowledge cannot leak through.

Sufficient context helps models the way reference texts help students. Including the source material for a question improves answers and reduces hallucination: without needed facts, the model falls back on unreliable parametric knowledge. Context may be supplied directly or gathered via tools (retrieval, search) — agent tool categories call this knowledge augmentation. Hybrid retrieval-augmented generation makes that non-parametric memory inspectable: retrieved passages are provenance for what the generator used, and the index can be swapped when facts change — neither is available when knowledge lives only in weights.

Restricting a model to only provided context is harder. Clear instructions ("answer using only the provided context"), examples of unanswerable questions, and asking the model to quote its source help — but prompting alone gives no guarantee. Finetuning and exclusive training on the permitted corpus are stronger and often infeasible. Place critical grounding material with lost in the middle in mind.

---

Sources

• AI Engineering (Chip Huyen), ch. 5
• RAG (Lewis et al.), §1