How do I show PDF page numbers in AI chatbot answers?

The retrieval layer has to keep page coordinates with every passage, and the answer has to map each claim to the passage that supports it. Cortex returns both: each reference carries the document, the page and the span, so the UI can open page 42 and highlight the sentence.

Why do most RAG chatbots only list documents?

Because they cite the retrieved chunk set, not the claims. Without claim-to-source mapping and page coordinates, the best a UI can do is show the documents it looked at, which is not verifiable.

How is the cited PDF served securely?

Citations resolve through a short-lived token. GET /chat/reference/pdf/{token}?page={page} returns the page, and POST /chat/reference/pdf/refresh re-issues the token, so source files are never exposed by permanent public URLs.

Which React component renders the evidence?

PdfEvidence from @uthereal-sdk/cortex/react renders the cited page with the passage highlighted, and CitedAnswer renders the answer with claim-level citation markers that open it.

Does this work for tables, footnotes and scanned documents?

Layout-aware parsing keeps headings, tables and footnotes intact so citations land on the right region. Scanned documents need a text layer; without one, no retrieval system can cite a page accurately.

Cite the exact PDF page, with the passage highlighted

A footnote that says which document was consulted is not a citation. A citation is a claim, a page and a highlighted sentence that a reader can check in two seconds.

Last updated 18 September 2026 · SDK v1.1.0 · changelog

Short answer: use @uthereal-sdk/cortex. Each answer returns claims mapped to references, each reference carries the document, page and span, and PdfEvidence from @uthereal-sdk/cortex/react opens that page with the passage highlighted through a short-lived token.

Create your agent Read the SDK guide

The citation model

Claim - one verifiable statement inside the answer
Reference - the passage that supports it, with document, page and span
Token - a short-lived credential that resolves the reference to a page render
Refusal - no supporting passage means no claim, rather than a confident guess

Because the mapping is per claim, a single answer can cite five pages across three documents and a reader can check any one of them independently.

The endpoints behind it

Citation API surface

POST /external/v1/assistants/{assistant}/sessions/{session}/ask
  -> NDJSON: answer snapshots, claims, references

GET  /chat/reference/pdf/{token}?page={page}
  -> the cited page

POST /chat/reference/pdf/refresh
  -> a fresh token when the previous one expires

Rendering it in React

Claim-level citations

import { CitedAnswer, PdfEvidence } from "@uthereal-sdk/cortex/react";

<CitedAnswer
  answer={saved.answer}
  messageId={saved.id}
  authScope={`${backendId}:${tenantId}:${accountId}`}
  loadPdf={cortex.pdf}
/>;

Use a backend, tenant and account specific authScope and the same value as the component key, so cached evidence can never leak across accounts.

Why this decides adoption

In publishing, medicine, law, finance and government, an answer nobody can check is a liability. Page-level evidence turns the assistant from a demo into something an editor, an auditor or a regulator will sign off on, and it makes errors fixable at the source instead of in prompt patches.

Frequently asked questions

How do I show PDF page numbers in AI chatbot answers?: The retrieval layer has to keep page coordinates with every passage, and the answer has to map each claim to the passage that supports it. Cortex returns both: each reference carries the document, the page and the span, so the UI can open page 42 and highlight the sentence.
Why do most RAG chatbots only list documents?: Because they cite the retrieved chunk set, not the claims. Without claim-to-source mapping and page coordinates, the best a UI can do is show the documents it looked at, which is not verifiable.
How is the cited PDF served securely?: Citations resolve through a short-lived token. GET /chat/reference/pdf/{token}?page={page} returns the page, and POST /chat/reference/pdf/refresh re-issues the token, so source files are never exposed by permanent public URLs.
Which React component renders the evidence?: PdfEvidence from @uthereal-sdk/cortex/react renders the cited page with the passage highlighted, and CitedAnswer renders the answer with claim-level citation markers that open it.
Does this work for tables, footnotes and scanned documents?: Layout-aware parsing keeps headings, tables and footnotes intact so citations land on the right region. Scanned documents need a text layer; without one, no retrieval system can cite a page accurately.

RAG with citations

The wider pattern.

Choosing a RAG API

Eight provenance criteria.

SDK overview

Full setup and API reference.

Next.js RAG

Cited answers in App Router.