Anthropic introduces Citations on the Claude API

Anthropic announced a Citations feature for the Claude API, designed to associate answers with source passages from supplied documents. The feature supports source-aware interfaces, while document selection, extraction quality, and presentation context still require application-level validation and must not be mistaken for factual proof by users.

By Hypler editorial team · Source: Anthropic.

Original source date: 2025-01-23.

Hypler publication: 2026-10-06. Hypler updated: 2026-10-10.

Topics: citations, RAG, provenance

Anthropic logo

Generated illustration of indexed source records connected to a reference document. Credit: Hypler / AI-generated editorial illustration. Hypler editorial asset; not vendor or event photography

Generated illustration of indexed source records and a vector retrieval path. Credit: Hypler / AI-generated editorial illustration. Hypler editorial asset; not vendor or event photography

Engineering relevance

Citations can improve inspectability when source IDs and passages are retained. Avoid treating a displayed citation as evidence of factual correctness or complete retrieval coverage.

Announcement record

Why it matters

Citations can improve inspectability when source IDs and passages are retained. Avoid treating a displayed citation as evidence of factual correctness or complete retrieval coverage.

Release record and source tracking

Anthropic's January 23, 2025 API release notes record the Citations capability launch, including plaintext and custom-document support. The current vendor blog publication is dated June 23, 2025 and is a later publication record, not a replacement for the original API release date. Keeping these records separate helps readers distinguish a stable feature contract from later documentation updates. The historical news date remains January 23, while the linked blog remains useful for checking the vendor's current presentation of the capability.

A citation needs a provenance chain

A trustworthy interface needs more than a rendered footnote. Preserve the source document version, extraction method, chunk boundaries, retrieval query, model request, response version, and citation offsets. The viewer should be able to open the exact source excerpt and see whether the cited text directly supports the nearby claim or only supplied background. When a document is re-ingested, old answers need either immutable version references or an explicit stale marker. A citation service should also distinguish first-party policy, customer-provided material, vendor documentation, and model-generated summaries; their authority and retention rules are not the same.

Failure modes users can inspect

Citation systems fail by retrieving the wrong document, splitting a qualifying sentence from its exception, assigning a span to an adjacent claim, or displaying a source that is no longer accessible. Long PDFs add OCR errors, page renumbering, tables that lose column structure, and image-only sections with no reliable text span. Do not hide a missing citation behind decorative certainty. Mark it unavailable, show the document revision used, and prevent a claim from inheriting a citation from an earlier draft. Evaluation should score support precision and coverage separately: a precise citation on one sentence says nothing about uncited conclusions nearby.

Hypothetical incident answer

Hypothetical: an engineer asks why a connector was disabled last week. The retrieval layer selects an incident record, a policy revision, and a deployment receipt. The response says the policy threshold changed, then attaches separate citations to the rule change and the disabling action. Before presentation, a resolver checks that the viewer is authorized for all three records; if not, it returns a redacted explanation with no leaking title or offset. A reviewer can compare the generated claim to the cited spans and reject it. The system records the corpus revision and response hash, so a later policy rewrite cannot silently change what the original answer appeared to prove.

Acceptance checks before release

Build a test corpus with near-duplicate passages, conflicting policy versions, OCR defects, tables, deleted files, restricted documents, and claims with no support. Assert that every displayed citation resolves to the immutable source revision and that permission checks happen before snippets reach the model or UI. Add adversarial tests where an irrelevant document repeats the answer's keywords. Measure citation precision, unsupported-claim rate, correct abstention, and retrieval latency. A reviewer should be able to reproduce an answer using the captured corpus version. Passing schema validation alone is insufficient because a valid source ID can still be semantically wrong.

Hypler connection

Hypler applies source-aware output to evidence review, where readers can inspect the material behind a claim and distinguish support from interpretation. Citation quality is evaluated with document versions, retrieval coverage, and reviewer comparison so a visible reference improves accountability without being mistaken for proof on its own.

Editorial presentation rule

Citation UI should make the difference between evidence, inference, and unanswered questions legible. Put source-backed statements beside stable links, identify synthesized conclusions as synthesis, and leave unsupported portions visibly unresolved. A reader should not have to open a technical log to learn which revision was used or whether a document is incomplete. For long sources, show a concise excerpt with an explicit page or location link and preserve surrounding context that might qualify the claim. Product copy should avoid calling an answer verified merely because it has a citation. Verification is an operator activity that compares the claim, the cited passage, and the governing source set.