Grounding

Relevance measures the retriever. Grounding measures whether the generated answer is actually supported by the retrieved evidence, or whether the model filled gaps from its own prior, or extrapolated beyond what the sources say, or cited a source that does not contain the claim attributed to it.

This is the dimension that turns a search problem into a trust problem. A system with excellent recall can still produce answers that are subtly unsupported, and those are the answers that do damage, because they arrive with citations and sound right.

I would measure grounding at the claim level, not the answer level. Decompose the answer into atomic claims. For each claim, check whether it is entailed by at least one retrieved passage. The grounding rate is the share of claims that are supported. Citation precision is the share of citations that actually support the claim they are attached to. Both can be measured automatically with a natural-language-inference model or a judge model, calibrated against a human-labeled sample, and both should be tracked on production traffic, not just on the evaluation set.

Mechanisms that move grounding:

Span-level attribution. Require the generation to cite the specific passage, not the document, for each claim. This is a prompt and output-format decision, and it makes grounding checkable. A citation to "the policy document" cannot be verified; a citation to chunk 4 of section 3.2 can.

Abstention when evidence is missing. The system must be able to say "the retrieved material does not answer this." Most systems cannot, because nothing in the prompt or the evaluation rewards it. Include no-answer queries in the evaluation set, measure the false-answer rate on them, and make declining an explicit, expected output.

Confidence from retrieval, not just from generation. If the reranker's top score is low, or the fused candidates disagree with each other, the answer should carry that uncertainty forward — to the user as a caveat, to an agent as a signal to retrieve again or escalate. Retrieval scores are information; throwing them away at the generation boundary wastes it.

Post-generation verification for high-stakes paths. Where the cost of an unsupported claim is high, run the grounding check inline and block or flag answers below a threshold. It adds latency and a model call. For a subset of traffic, that is the right trade.

Serve the source, not just the summary. Summaries are a retrieval accelerator and a token saver, and they are also a grounding risk, because a summary is already one step removed from the evidence. When an answer rests on a summary, the system should be able to escalate to the underlying passage for verification. This is one of the reasons I insist on storing every representation rather than only the compressed one.