Your RAG answer cannot prove where it came from

// plain text has no handle

The default RAG wiring concatenates the top-k chunks into a prompt and asks the question. The model answers well. But the answer is a block of prose, and nothing in the response says which chunk any sentence came from — so verifying a claim means a human going back to the corpus and searching for it.

Search result content blocks change the shape rather than the prompt:

Three fields you already have

A search result is not a new document format. It is the thing your retriever already returns, labelled.

source does not have to be a URL — the docs say "Any stable string works: a URL, or an internal identifier such as kb://article-1234". So a private corpus with no public pages is fine.

They can be returned two ways: from a tool call, for live retrieval, or as top-level content in a user message for pre-fetched results. Either way:

What comes back is a mapping, not a footnote

The answer arrives as text blocks. A block that leaned on a result carries a citations array, and each entry identifies exactly what it used.

That is enough to render a footnote, highlight the passage in a viewer, or diff the sentence against the text it was built from — without asking the model to format anything, and without trusting it to format it consistently.

One useful detail: cited_text is "Not counted toward output tokens", so the quoted passage coming back does not cost you output.

Granularity is the block, not the sentence

This is the limit worth knowing before you promise anyone sentence-level attribution:

So citation precision is a chunking decision. One 2000-character block means every citation returns all 2000 characters; the same content split into paragraphs means a citation points at a paragraph.

Which makes this the same decision as chunking, pointed at a second purpose.

Off by default, and all or nothing

Two things will catch a migration.

It is off. "By default, citations are disabled for search results." You have to set citations: {"enabled": true} explicitly — moving to search result blocks without that field gets you the new shape and none of the benefit.

So you cannot roll it out one retriever at a time inside a single request. If two code paths contribute results to the same call, both have to agree.

And on empty searches, the docs advise returning a plain text block rather than raising: "Claude explains the empty result to the user, and the conversation continues."

Checkable, not true

Worth being precise about what this buys, because it is easy to oversell. A citation tells you which block a sentence leaned on. It says nothing about whether that block is correct, current, or should still exist.

Which is exactly why it is useful: it makes the other RAG failures visible. When a page you deleted is still in the index, a citation is what shows you that the answer came from it — without one, a stale copy and a fresh one produce identical-looking prose.

Related: working out where your RAG is broken is the diagnosis you run once, by hand; this is the same information arriving with every answer. Why models make things up is a different failure and citations are not a cure for it.

Source: Claude Platform docs — Search results. Verified 2026-09-13.