Also known as: grounding, citation-required generation, attributed generation
Grounded generation is the pattern of forcing an LLM's output to be derivable from a supplied set of retrieved sources, with citations attached. The standard defense against hallucination in RAG pipelines.
Grounded generation is the pattern of constraining an LLM’s output so that every assertion it makes is traceable to a supplied source. The mechanism is part prompt engineering, part output structure, and part post-hoc verification — together they push the model away from parametric-knowledge fallback and toward the retrieved context.
A grounded-generation pipeline has the same shape across providers:
[SRC-1], [SRC-2], or doc:42 — that the model can reference. Without IDs the model cannot emit useful citations even when it tries.[SRC-1]) or an explicit citation block at the end. Some pipelines enforce this via constrained decoding .Grounded generation is the difference between “the model wrote something that happens to be correct” and “the model wrote something whose correctness you can audit by clicking a citation.”
LLMs default to a blend of retrieved context and parametric knowledge. The blend is implicit — there’s no boundary marker between “I learned this from training data” and “I learned this from the prompt.” Grounded generation flips the default: by explicit instruction and by the structural pressure of having to emit citations, the model spends most of its attention budget on the retrieved sources.
The output structure does the real work. A model that has to write claim [SRC-3] has to actually attend to SRC-3 while producing the claim. The citation is a forcing function — if the model can’t find a source, it has to either say so or invent one, and inventing one is now an auditable failure mode.
Grounding also does not eliminate hallucination on the gaps between sources. The model may correctly cite SRC-1 for a fact and then add an unsupported interpretive sentence with no citation. Catching that requires citation-extraction and per-claim entailment checking — see faithfulness for the eval recipe.
Three lines do most of the work:
You will be given numbered sources. Answer the question using ONLY
these sources. After each factual claim, cite the source ID in
brackets, like [SRC-2]. If no source supports the answer, say
"I do not know."
Sources:
[SRC-1] ...
[SRC-2] ...
[SRC-3] ...
Question: ...The hard part is the refusal clause. Models love to be helpful and will reach for parametric knowledge unless explicitly told not to. Without “if no source supports the answer, say I do not know,” the model silently falls back to its training data and your grounding rate collapses on out-of-context questions.
RAG is the broader pipeline — retrieve sources, stuff them into context, generate an answer. Grounded generation is a specific generation style within RAG that adds two constraints: every claim in the answer must trace back to a retrieved source, and the model must emit citations marking those traces. A RAG pipeline can be ungrounded if you skip the prompting and post-processing that enforce those constraints.
Three things: an explicit instruction (answer only from the provided sources, do not use prior knowledge), the sources tagged with IDs the model can cite ([SRC-1], [SRC-2]), and an output schema that requires citation markers. Production templates also instruct the model to refuse when no source supports the answer rather than fall back to parametric knowledge.
Decompose the answer into atomic claims, then check each claim's cited source for entailment. RAGAS, TruLens, and DeepEval all implement variants of this. The headline metric is faithfulness — fraction of answer claims that are entailed by their cited source. A faithful answer is grounded; an answer with high faithfulness but missing citations is half-grounded.
