A Practical Tutorial for Adding Citations to LLM-Generated Answers

1. Define what a citation must prove

A citation-aware system needs a stricter contract than “include some links”. Every material claim should be traceable to evidence that directly supports its wording. Material claims include numbers, dates, quotations, causal statements, comparisons and assertions about policy or product behaviour. If an answer says a medicine reduced hospital admissions by 18%, the cited source must report that result for the relevant population and outcome. A page merely discussing the medicine is not sufficient.

Start by choosing the citation unit. Sentence-level citations are readable, but a sentence containing three claims may require three sources. Claim-level citations are more precise, though they can clutter prose. A practical compromise is to restrict each sentence to one or two closely related claims, then attach citations immediately after the supported clause. Define whether citations point to documents, passages or exact offsets. Document links are easy to implement; passage references make auditing faster and reduce the chance that a loosely related source receives undeserved authority.

Write these rules into an explicit policy. For example: factual claims require evidence; common knowledge does not; recommendations must distinguish sourced facts from model judgement; and unsupported claims must be omitted or labelled as uncertain. Decide how to treat sources that disagree, whether secondary reporting can support primary research claims, and when publication date matters. These decisions shape retrieval, prompting, validation and the user interface, so they cannot be left to the language model’s intuition.

2. Retrieve evidence as structured passages

Citation quality is bounded by retrieval quality. Index sources in passages that preserve enough context to interpret the claim: typically 150 to 400 tokens, with a small overlap between adjacent chunks. Store each passage with a stable source identifier, title, publisher, publication date, URL, section heading and character offsets. For PDFs, retain page numbers; for transcripts, retain timestamps. Stable identifiers matter because a citation should survive re-ranking, answer revisions and duplicate URLs.

Use hybrid retrieval rather than relying solely on vector similarity. Dense embeddings are effective for conceptual matches, while keyword or BM25 search handles names, figures and uncommon terminology. Retrieve perhaps 20 to 40 candidates, then apply a cross-encoder or capable language model to re-rank the top set. A query about “UK electric vehicle mandate targets in 2027” should reward passages containing the jurisdiction, policy, year and percentage, not generic pages about electric vehicles.

Source selection also requires editorial judgement encoded as metadata and ranking rules. Prefer a regulator’s policy document over a blog summarising it, and a company filing over an unsourced product comparison. Freshness is essential for prices, officeholders and regulations, but less important for established historical facts. Deduplicate syndicated articles and near-identical passages so that apparent consensus is not merely one press release repeated across ten domains. Where possible, keep both primary evidence and a reputable explanatory source: the first establishes the fact, while the second can clarify its meaning.

3. Track evidence spans before drafting prose

Do not ask the model to draft freely and add citations afterwards. That workflow encourages plausible claims first and evidence hunting second. Instead, create an evidence ledger before generation. Each entry should contain a proposed atomic claim, one or more supporting passage IDs, the exact supporting span and a confidence label. A ledger entry might read: “The scheme begins in January 2027”; source S12; characters 438–512; confidence high. The answer generator receives only approved entries and their source metadata.

Exact spans are valuable even when users see only a link or footnote. They let validators compare claim text with the smallest relevant evidence, rather than an entire 5,000-word page. Span extraction can be performed by a smaller model instructed to quote verbatim text, followed by deterministic verification that the quotation exists in the retrieved passage. If it does not, reject the span. This simple check catches fabricated quotations and offset errors before they reach the answer.

Represent provenance in the generation format. One option is a structured draft containing fields such as claim_text, evidence_ids and rhetorical_role. Another is inline markers such as [[S12:438-512]], later rendered as numbered citations. Structured output is safer because citations cannot drift as easily during editing. It also supports many-to-many relationships: one claim may need two sources, while one source may support several claims. Keep provenance attached through summarisation, translation and tone editing; any transformation that changes meaning should trigger revalidation.

4. Constrain generation around atomic claims

The generation prompt should state that the model may use only supplied evidence for factual assertions. Require short, atomic sentences and prohibit invented bridging facts. Models often connect two supported statements with an unsupported causal link: sales rose after a price cut, therefore the price cut caused the increase. Unless the evidence establishes causality, the answer should say only that the events coincided. This distinction is central to trustworthy citation, because a citation can be topically relevant yet fail to support the actual inference.

Separate sourced facts from synthesis. A useful output pattern is: evidence-backed statement, citation, then clearly framed interpretation. For example, “The regulator recorded 420 incidents in 2025 [3]. That volume suggests monitoring costs will remain material.” The first sentence is sourced; the second is analysis and should be presented as such. If the interpretation depends on assumptions, name them. Recommendations can still be decisive, but readers must be able to see where documentary evidence ends and model judgement begins.

Set a citation budget without turning it into a quota. Requiring one citation per sentence can produce citation spam, while allowing one citation per paragraph often conceals unsupported details. Measure claim coverage instead: the proportion of material claims linked to adequate evidence. A useful production target may be above 95% for ordinary factual answers and effectively 100% for high-risk domains such as medicine, law or finance. Coverage alone is not enough, however; ten perfectly placed citations are worthless if none entails the claims they accompany.

5. Validate entailment, scope and source alignment

Validation should occur after drafting and before rendering. First split the response into atomic claims, then classify each as supported, contradicted, not found or non-verifiable. Compare every supported claim against its cited span using an entailment model or a separate language-model judge. Ask a narrow question: if the evidence is true, does it justify the claim as written? Avoid asking whether the source is merely “relevant”, because relevance tolerates serious mismatches.

Most failures involve scope. A study of 600 adults in one city does not support a claim about all consumers. A 12% relative reduction is not necessarily a 12-percentage-point reduction. A forecast is not an observed result, and a company announcement is not proof that a feature shipped. Build checks for dates, units, populations, geographies, negation and modality. Deterministic comparison works well for numbers: extract values and units from both claim and evidence, normalise formats, and flag any unexplained difference. A claim of £2.4 million should not pass against evidence stating £2.4 billion.

Use thresholds that reflect risk and cost. An entailment score below a chosen boundary might send the claim for rewriting, alternative retrieval or human review. Automated judges can be inconsistent, so evaluate them on a labelled set of at least several hundred claim–evidence pairs from your own domain. Track precision for “supported” decisions, not just overall accuracy; falsely approving an unsupported claim is usually more damaging than rejecting a valid one. For consequential outputs, sample audits should examine both evidence quality and whether the cited publisher is authoritative.

6. Handle insufficient and conflicting evidence honestly

A robust system needs an explicit abstention path. If retrieval finds no adequate support, the answer should narrow the claim, state that available sources do not establish it, ask for more context or omit the assertion. Never cite the nearest topical passage simply to satisfy a formatting rule. For a question asking whether a product is “the safest”, evidence about compliance certification supports only that it met a stated standard, not a comparative superlative.

Distinguish absence of evidence from evidence of absence. If no retrieved source mentions a side effect, the system cannot claim that the side effect never occurs. It may say that the reviewed sources did not report it, while noting the limits of that search. Similarly, when evidence is incomplete, use calibrated language such as “the report estimates”, “the trial found within its study population” or “public documentation does not confirm”. These formulations preserve useful information without overstating certainty.

Conflicts should be surfaced rather than averaged away. If two credible sources report 14% and 19%, check whether they use different periods, denominators or definitions. Cite both and explain the discrepancy when it is material. If reconciliation is impossible, report the range and attribute each figure. Maintain source diversity carefully: independent corroboration strengthens a claim, but three articles quoting the same unnamed source constitute one evidence chain, not three. Provenance graphs can reveal these dependencies.

7. Test the pipeline and monitor production drift

Build an evaluation set that resembles real traffic, including ambiguous questions, stale sources, numerical claims, contradictory documents and prompts designed to induce overclaiming. Label atomic claims, ideal evidence spans and acceptable abstentions. Core metrics should include citation precision, citation recall, claim coverage, source correctness and unsupported-claim rate. Also measure answer usefulness: a system can achieve perfect factuality by refusing everything. Report results by question type and risk tier, because an aggregate score can hide weak performance on fast-changing topics.

Run adversarial tests. Remove the best passage and observe whether the system abstains or attaches a weaker citation. Insert a passage that shares keywords but contradicts the claim. Provide two documents with different dates and check whether the newer one is preferred where appropriate. Test citation persistence after sentence reordering and copy-editing. A frequent production bug occurs when references are numbered before paragraphs are revised, leaving syntactically valid citations attached to the wrong claims.

Monitor live behaviour with privacy-conscious logs containing claim text, evidence IDs, validator scores, retrieval ranks and final citation positions. Watch for declining support rates after model, embedding, chunking or index changes. Set alerts for broken links, sources that disappear and sudden concentration on one domain. Human reviewers should inspect low-confidence outputs and a random sample of high-confidence ones, since systematic errors can pass automated checks confidently. The operational goal is not decorative footnotes; it is a traceable chain from user question to retrieved passage, bounded claim, validated answer and durable source.