diff --git a/surfsense_web/components/assistant-ui/inline-citation.tsx b/surfsense_web/components/assistant-ui/inline-citation.tsx
index 2aeba89ca..e299f2373 100644
--- a/surfsense_web/components/assistant-ui/inline-citation.tsx
+++ b/surfsense_web/components/assistant-ui/inline-citation.tsx
@@ -3,11 +3,11 @@
 import { useQuery } from "@tanstack/react-query";
 import { useSetAtom } from "jotai";
 import { ExternalLink, FileText } from "lucide-react";
+import dynamic from "next/dynamic";
 import type { FC } from "react";
 import { useCallback, useEffect, useRef, useState } from "react";
 import { openCitationPanelAtom } from "@/atoms/citation/citation-panel.atom";
 import { useCitationMetadata } from "@/components/assistant-ui/citation-metadata-context";
-import { MarkdownViewer } from "@/components/markdown-viewer";
 import { Citation } from "@/components/tool-ui/citation";
 import { Popover, PopoverContent, PopoverTrigger } from "@/components/ui/popover";
 import { Spinner } from "@/components/ui/spinner";
@@ -15,6 +15,16 @@ import { Tooltip, TooltipContent, TooltipTrigger } from "@/components/ui/tooltip
 import { documentsApiService } from "@/lib/apis/documents-api.service";
 import { cacheKeys } from "@/lib/query-client/cache-keys";
 
+// Lazily load MarkdownViewer here to break the static import cycle:
+// `markdown-viewer.tsx` → `citation-renderer.tsx` → `inline-citation.tsx`
+// would otherwise pull `markdown-viewer.tsx` back in at module-init time.
+// Only `SurfsenseDocCitation` (popover body) ever renders this viewer, so
+// the lazy boundary is invisible to most call paths.
+const MarkdownViewer = dynamic(
+	() => import("@/components/markdown-viewer").then((m) => m.MarkdownViewer),
+	{ ssr: false, loading: () => <Spinner size="xs" /> }
+);
+
 interface InlineCitationProps {
 	chunkId: number;
 	isDocsChunk?: boolean;
@@ -172,7 +182,11 @@ const SurfsenseDocCitation: FC<{ chunkId: number }> = ({ chunkId }) => {
 						</p>
 					)}
 					{!isLoading && !error && citedChunk?.content && (
-						<MarkdownViewer content={citedChunk.content} maxLength={1500} />
+						<MarkdownViewer
+							content={citedChunk.content}
+							maxLength={1500}
+							enableCitations
+						/>
 					)}
 					{!isLoading && !error && !citedChunk?.content && (
 						<p className="py-4 text-xs text-muted-foreground">No content available.</p>
diff --git a/surfsense_web/components/assistant-ui/markdown-text.tsx b/surfsense_web/components/assistant-ui/markdown-text.tsx
index 7655e10cc..2b788e88b 100644
--- a/surfsense_web/components/assistant-ui/markdown-text.tsx
+++ b/surfsense_web/components/assistant-ui/markdown-text.tsx
@@ -12,15 +12,26 @@ import { ExternalLinkIcon } from "lucide-react";
 import dynamic from "next/dynamic";
 import { useParams } from "next/navigation";
 import { useTheme } from "next-themes";
-import { memo, type ReactNode } from "react";
+import {
+	createContext,
+	memo,
+	type ReactNode,
+	useCallback,
+	useContext,
+	useRef,
+} from "react";
 import rehypeKatex from "rehype-katex";
 import remarkGfm from "remark-gfm";
 import remarkMath from "remark-math";
 import { openEditorPanelAtom } from "@/atoms/editor/editor-panel.atom";
 import { ImagePreview, ImageRoot, ImageZoom } from "@/components/assistant-ui/image";
 import "katex/dist/katex.min.css";
-import { InlineCitation, UrlCitation } from "@/components/assistant-ui/inline-citation";
+import { processChildrenWithCitations } from "@/components/citations/citation-renderer";
 import { Skeleton } from "@/components/ui/skeleton";
+import {
+	type CitationUrlMap,
+	preprocessCitationMarkdown,
+} from "@/lib/citations/citation-parser";
 import {
 	Table,
 	TableBody,
@@ -59,31 +70,30 @@ const LazyMarkdownCodeBlock = dynamic(
 	}
 );
 
-// Storage for URL citations replaced during preprocess to avoid GFM autolink interference.
-// Populated in preprocessMarkdown, consumed in parseTextWithCitations.
-let _pendingUrlCitations = new Map<string, string>();
-let _urlCiteIdx = 0;
+// Per-render URL placeholder map propagated to component overrides via
+// React Context. Replaces the previous module-level `_pendingUrlCitations`
+// state, which was unsafe under concurrent renders / SSR.
+type CitationUrlMapRef = { current: CitationUrlMap };
+const EMPTY_URL_MAP: CitationUrlMap = new Map();
+const CitationUrlMapContext = createContext<CitationUrlMapRef>({ current: EMPTY_URL_MAP });
+
+function useCitationUrlMap(): CitationUrlMap {
+	return useContext(CitationUrlMapContext).current;
+}
 
 /**
  * Preprocess raw markdown before it reaches the remark/rehype pipeline.
  * - Replaces URL-based citations with safe placeholders (prevents GFM autolinks)
  * - Normalises LaTeX delimiters to dollar-sign syntax for remark-math
  */
-function preprocessMarkdown(content: string): string {
+function preprocessMarkdown(content: string, urlMapRef: CitationUrlMapRef): string {
 	// Replace URL-based citations with safe placeholders BEFORE markdown parsing.
 	// GFM autolinks would otherwise convert the https://... inside [citation:URL]
 	// into an <a> element, splitting the text and preventing our citation regex
 	// from matching the full pattern.
-	_pendingUrlCitations = new Map();
-	_urlCiteIdx = 0;
-	content = content.replace(
-		/[[【]\u200B?citation:\s*(https?:\/\/[^\]】\u200B]+)\s*\u200B?[\]】]/g,
-		(_, url) => {
-			const key = `urlcite${_urlCiteIdx++}`;
-			_pendingUrlCitations.set(key, url.trim());
-			return `[citation:${key}]`;
-		}
-	);
+	const { content: rewritten, urlMap } = preprocessCitationMarkdown(content);
+	urlMapRef.current = urlMap;
+	content = rewritten;
 
 	// All math forms are normalised to $$...$$ so we can disable single-dollar
 	// inline math in remark-math (otherwise currency like "$3,120.00 and $0.00"
@@ -116,113 +126,28 @@ function preprocessMarkdown(content: string): string {
 	return content;
 }
 
-// Matches [citation:...] with numeric IDs (incl. negative, doc- prefix, comma-separated),
-// URL-based IDs from live web search, or urlciteN placeholders from preprocess.
-// Also matches Chinese brackets 【】 and handles zero-width spaces that LLM sometimes inserts.
-const CITATION_REGEX =
-	/[[【]\u200B?citation:\s*(https?:\/\/[^\]】\u200B]+|urlcite\d+|(?:doc-)?-?\d+(?:\s*,\s*(?:doc-)?-?\d+)*)\s*\u200B?[\]】]/g;
-
-/**
- * Parses text and replaces [citation:XXX] patterns with citation components.
- * Supports:
- *  - Numeric chunk IDs: [citation:123]
- *  - Doc-prefixed IDs: [citation:doc-123]
- *  - Comma-separated IDs: [citation:4149, 4150, 4151]
- *  - URL-based citations from live search: [citation:https://example.com/page]
- */
-function parseTextWithCitations(text: string): ReactNode[] {
-	const parts: ReactNode[] = [];
-	let lastIndex = 0;
-	let match: RegExpExecArray | null;
-	let instanceIndex = 0;
-
-	CITATION_REGEX.lastIndex = 0;
-
-	match = CITATION_REGEX.exec(text);
-	while (match !== null) {
-		if (match.index > lastIndex) {
-			parts.push(text.substring(lastIndex, match.index));
-		}
-
-		const captured = match[1];
-
-		if (captured.startsWith("http://") || captured.startsWith("https://")) {
-			parts.push(<UrlCitation key={`citation-url-${instanceIndex}`} url={captured.trim()} />);
-			instanceIndex++;
-		} else if (captured.startsWith("urlcite")) {
-			const url = _pendingUrlCitations.get(captured);
-			if (url) {
-				parts.push(<UrlCitation key={`citation-url-${instanceIndex}`} url={url} />);
-			}
-			instanceIndex++;
-		} else {
-			const rawIds = captured.split(",").map((s) => s.trim());
-			for (const rawId of rawIds) {
-				const isDocsChunk = rawId.startsWith("doc-");
-				const chunkId = Number.parseInt(isDocsChunk ? rawId.slice(4) : rawId, 10);
-				parts.push(
-					<InlineCitation
-						key={`citation-${isDocsChunk ? "doc-" : ""}${chunkId}-${instanceIndex}`}
-						chunkId={chunkId}
-						isDocsChunk={isDocsChunk}
-					/>
-				);
-				instanceIndex++;
-			}
-		}
-
-		lastIndex = match.index + match[0].length;
-		match = CITATION_REGEX.exec(text);
-	}
-
-	if (lastIndex < text.length) {
-		parts.push(text.substring(lastIndex));
-	}
-
-	return parts.length > 0 ? parts : [text];
-}
-
 const MarkdownTextImpl = () => {
+	const urlMapRef = useRef<CitationUrlMap>(EMPTY_URL_MAP);
+	const preprocess = useCallback(
+		(content: string) => preprocessMarkdown(content, urlMapRef),
+		[]
+	);
 	return (
-		<MarkdownTextPrimitive
-			smooth={false}
-			remarkPlugins={[remarkGfm, [remarkMath, { singleDollarTextMath: false }]]}
-			rehypePlugins={[rehypeKatex]}
-			className="aui-md"
-			components={defaultComponents}
-			preprocess={preprocessMarkdown}
-		/>
+		<CitationUrlMapContext.Provider value={urlMapRef}>
+			<MarkdownTextPrimitive
+				smooth={false}
+				remarkPlugins={[remarkGfm, [remarkMath, { singleDollarTextMath: false }]]}
+				rehypePlugins={[rehypeKatex]}
+				className="aui-md"
+				components={defaultComponents}
+				preprocess={preprocess}
+			/>
+		</CitationUrlMapContext.Provider>
 	);
 };
 
 export const MarkdownText = memo(MarkdownTextImpl);
 
-/**
- * Helper to process children and replace citation patterns with components
- */
-function processChildrenWithCitations(children: ReactNode): ReactNode {
-	if (typeof children === "string") {
-		const parsed = parseTextWithCitations(children);
-		return parsed.length === 1 && typeof parsed[0] === "string" ? children : parsed;
-	}
-
-	if (Array.isArray(children)) {
-		return children.map((child) => {
-			if (typeof child === "string") {
-				const parsed = parseTextWithCitations(child);
-				return parsed.length === 1 && typeof parsed[0] === "string" ? (
-					child
-				) : (
-					<span key={child}>{parsed}</span>
-				);
-			}
-			return child;
-		});
-	}
-
-	return children;
-}
-
 function extractDomain(url: string): string {
 	try {
 		const parsed = new URL(url);
@@ -322,92 +247,125 @@ function MarkdownImage({ src, alt }: { src?: string; alt?: string }) {
 }
 
 const defaultComponents = memoizeMarkdownComponents({
-	h1: ({ className, children, ...props }) => (
-		<h1
-			className={cn(
-				"aui-md-h1 mb-8 scroll-m-20 font-extrabold text-4xl tracking-tight last:mb-0",
-				className
-			)}
-			{...props}
-		>
-			{processChildrenWithCitations(children)}
-		</h1>
-	),
-	h2: ({ className, children, ...props }) => (
-		<h2
-			className={cn(
-				"aui-md-h2 mt-8 mb-4 scroll-m-20 font-semibold text-3xl tracking-tight first:mt-0 last:mb-0",
-				className
-			)}
-			{...props}
-		>
-			{processChildrenWithCitations(children)}
-		</h2>
-	),
-	h3: ({ className, children, ...props }) => (
-		<h3
-			className={cn(
-				"aui-md-h3 mt-6 mb-4 scroll-m-20 font-semibold text-2xl tracking-tight first:mt-0 last:mb-0",
-				className
-			)}
-			{...props}
-		>
-			{processChildrenWithCitations(children)}
-		</h3>
-	),
-	h4: ({ className, children, ...props }) => (
-		<h4
-			className={cn(
-				"aui-md-h4 mt-6 mb-4 scroll-m-20 font-semibold text-xl tracking-tight first:mt-0 last:mb-0",
-				className
-			)}
-			{...props}
-		>
-			{processChildrenWithCitations(children)}
-		</h4>
-	),
-	h5: ({ className, children, ...props }) => (
-		<h5
-			className={cn("aui-md-h5 my-4 font-semibold text-lg first:mt-0 last:mb-0", className)}
-			{...props}
-		>
-			{processChildrenWithCitations(children)}
-		</h5>
-	),
-	h6: ({ className, children, ...props }) => (
-		<h6 className={cn("aui-md-h6 my-4 font-semibold first:mt-0 last:mb-0", className)} {...props}>
-			{processChildrenWithCitations(children)}
-		</h6>
-	),
-	p: ({ className, children, ...props }) => (
-		<p className={cn("aui-md-p mt-5 mb-5 leading-7 first:mt-0 last:mb-0", className)} {...props}>
-			{processChildrenWithCitations(children)}
-		</p>
-	),
-	a: ({ className, children, ...props }) => (
-		<a
-			className={cn("aui-md-a font-medium text-primary underline underline-offset-4", className)}
-			{...props}
-		>
-			{processChildrenWithCitations(children)}
-		</a>
-	),
-	blockquote: ({ className, children, ...props }) => (
-		<blockquote className={cn("aui-md-blockquote border-l-2 pl-6 italic", className)} {...props}>
-			{processChildrenWithCitations(children)}
-		</blockquote>
-	),
+	h1: function H1({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<h1
+				className={cn(
+					"aui-md-h1 mb-8 scroll-m-20 font-extrabold text-4xl tracking-tight last:mb-0",
+					className
+				)}
+				{...props}
+			>
+				{processChildrenWithCitations(children, urlMap)}
+			</h1>
+		);
+	},
+	h2: function H2({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<h2
+				className={cn(
+					"aui-md-h2 mt-8 mb-4 scroll-m-20 font-semibold text-3xl tracking-tight first:mt-0 last:mb-0",
+					className
+				)}
+				{...props}
+			>
+				{processChildrenWithCitations(children, urlMap)}
+			</h2>
+		);
+	},
+	h3: function H3({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<h3
+				className={cn(
+					"aui-md-h3 mt-6 mb-4 scroll-m-20 font-semibold text-2xl tracking-tight first:mt-0 last:mb-0",
+					className
+				)}
+				{...props}
+			>
+				{processChildrenWithCitations(children, urlMap)}
+			</h3>
+		);
+	},
+	h4: function H4({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<h4
+				className={cn(
+					"aui-md-h4 mt-6 mb-4 scroll-m-20 font-semibold text-xl tracking-tight first:mt-0 last:mb-0",
+					className
+				)}
+				{...props}
+			>
+				{processChildrenWithCitations(children, urlMap)}
+			</h4>
+		);
+	},
+	h5: function H5({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<h5
+				className={cn("aui-md-h5 my-4 font-semibold text-lg first:mt-0 last:mb-0", className)}
+				{...props}
+			>
+				{processChildrenWithCitations(children, urlMap)}
+			</h5>
+		);
+	},
+	h6: function H6({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<h6 className={cn("aui-md-h6 my-4 font-semibold first:mt-0 last:mb-0", className)} {...props}>
+				{processChildrenWithCitations(children, urlMap)}
+			</h6>
+		);
+	},
+	p: function P({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<p className={cn("aui-md-p mt-5 mb-5 leading-7 first:mt-0 last:mb-0", className)} {...props}>
+				{processChildrenWithCitations(children, urlMap)}
+			</p>
+		);
+	},
+	a: function A({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<a
+				className={cn(
+					"aui-md-a font-medium text-primary underline underline-offset-4",
+					className
+				)}
+				{...props}
+			>
+				{processChildrenWithCitations(children, urlMap)}
+			</a>
+		);
+	},
+	blockquote: function Blockquote({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<blockquote className={cn("aui-md-blockquote border-l-2 pl-6 italic", className)} {...props}>
+				{processChildrenWithCitations(children, urlMap)}
+			</blockquote>
+		);
+	},
 	ul: ({ className, ...props }) => (
 		<ul className={cn("aui-md-ul my-5 ml-6 list-disc [&>li]:mt-2", className)} {...props} />
 	),
 	ol: ({ className, ...props }) => (
 		<ol className={cn("aui-md-ol my-5 ml-6 list-decimal [&>li]:mt-2", className)} {...props} />
 	),
-	li: ({ className, children, ...props }) => (
-		<li className={cn("aui-md-li", className)} {...props}>
-			{processChildrenWithCitations(children)}
-		</li>
-	),
+	li: function Li({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<li className={cn("aui-md-li", className)} {...props}>
+				{processChildrenWithCitations(children, urlMap)}
+			</li>
+		);
+	},
 	hr: ({ className, ...props }) => (
 		<hr className={cn("aui-md-hr my-5 border-b", className)} {...props} />
 	),
@@ -422,28 +380,34 @@ const defaultComponents = memoizeMarkdownComponents({
 	tbody: ({ className, ...props }) => (
 		<TableBody className={cn("aui-md-tbody", className)} {...props} />
 	),
-	th: ({ className, children, ...props }) => (
-		<TableHead
-			className={cn(
-				"aui-md-th bg-muted/50 whitespace-normal [[align=center]]:text-center [[align=right]]:text-right",
-				className
-			)}
-			{...props}
-		>
-			{processChildrenWithCitations(children)}
-		</TableHead>
-	),
-	td: ({ className, children, ...props }) => (
-		<TableCell
-			className={cn(
-				"aui-md-td whitespace-normal [[align=center]]:text-center [[align=right]]:text-right",
-				className
-			)}
-			{...props}
-		>
-			{processChildrenWithCitations(children)}
-		</TableCell>
-	),
+	th: function Th({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<TableHead
+				className={cn(
+					"aui-md-th bg-muted/50 whitespace-normal [[align=center]]:text-center [[align=right]]:text-right",
+					className
+				)}
+				{...props}
+			>
+				{processChildrenWithCitations(children, urlMap)}
+			</TableHead>
+		);
+	},
+	td: function Td({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<TableCell
+				className={cn(
+					"aui-md-td whitespace-normal [[align=center]]:text-center [[align=right]]:text-right",
+					className
+				)}
+				{...props}
+			>
+				{processChildrenWithCitations(children, urlMap)}
+			</TableCell>
+		);
+	},
 	tr: ({ className, ...props }) => <TableRow className={cn("aui-md-tr", className)} {...props} />,
 	sup: ({ className, ...props }) => (
 		<sup className={cn("aui-md-sup [&>a]:text-xs [&>a]:no-underline", className)} {...props} />
@@ -552,16 +516,22 @@ const defaultComponents = memoizeMarkdownComponents({
 			/>
 		);
 	},
-	strong: ({ className, children, ...props }) => (
-		<strong className={cn("aui-md-strong font-semibold", className)} {...props}>
-			{processChildrenWithCitations(children)}
-		</strong>
-	),
-	em: ({ className, children, ...props }) => (
-		<em className={cn("aui-md-em", className)} {...props}>
-			{processChildrenWithCitations(children)}
-		</em>
-	),
+	strong: function Strong({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<strong className={cn("aui-md-strong font-semibold", className)} {...props}>
+				{processChildrenWithCitations(children, urlMap)}
+			</strong>
+		);
+	},
+	em: function Em({ className, children, ...props }) {
+		const urlMap = useCitationUrlMap();
+		return (
+			<em className={cn("aui-md-em", className)} {...props}>
+				{processChildrenWithCitations(children, urlMap)}
+			</em>
+		);
+	},
 	img: ({ src, alt }) => (
 		<MarkdownImage src={typeof src === "string" ? src : undefined} alt={alt} />
 	),
diff --git a/surfsense_web/components/citation-panel/citation-panel.tsx b/surfsense_web/components/citation-panel/citation-panel.tsx
index cec07b9cf..ed8acd656 100644
--- a/surfsense_web/components/citation-panel/citation-panel.tsx
+++ b/surfsense_web/components/citation-panel/citation-panel.tsx
@@ -169,7 +169,7 @@ export const CitationPanelContent: FC<CitationPanelContentProps> = ({ chunkId, o
 											)}
 										</div>
 										<div className="text-sm">
-											<MarkdownViewer content={chunk.content} />
+											<MarkdownViewer content={chunk.content} enableCitations />
 										</div>
 									</div>
 								);
diff --git a/surfsense_web/components/citations/citation-renderer.tsx b/surfsense_web/components/citations/citation-renderer.tsx
new file mode 100644
index 000000000..bf877f03f
--- /dev/null
+++ b/surfsense_web/components/citations/citation-renderer.tsx
@@ -0,0 +1,79 @@
+"use client";
+
+import type { ReactNode } from "react";
+import { InlineCitation, UrlCitation } from "@/components/assistant-ui/inline-citation";
+import {
+	type CitationToken,
+	type CitationUrlMap,
+	parseTextWithCitations,
+} from "@/lib/citations/citation-parser";
+
+/**
+ * Render a single parsed citation token as JSX.
+ *
+ * `ordinalKey` should be a stable per-render counter so duplicate identical
+ * citations within the same parent don't collide on `key`. The previous
+ * implementation in `markdown-text.tsx` used the source string itself as
+ * the key, which produced React warnings when two segments rendered the
+ * same `[citation:N]` text.
+ */
+export function renderCitationToken(token: CitationToken, ordinalKey: number): ReactNode {
+	if (token.kind === "url") {
+		return <UrlCitation key={`citation-url-${ordinalKey}`} url={token.url} />;
+	}
+	return (
+		<InlineCitation
+			key={`citation-${token.isDocsChunk ? "doc-" : ""}${token.chunkId}-${ordinalKey}`}
+			chunkId={token.chunkId}
+			isDocsChunk={token.isDocsChunk}
+		/>
+	);
+}
+
+/**
+ * Walk a `ReactNode` (string, array, or arbitrary node) and replace any
+ * `[citation:...]` tokens inside string children with citation badges.
+ *
+ * Designed for use inside `Streamdown`/`react-markdown` `components`
+ * overrides where the renderer hands you `children`. Non-string children
+ * are returned untouched so block/phrasing structure is preserved.
+ */
+export function processChildrenWithCitations(
+	children: ReactNode,
+	urlMap: CitationUrlMap
+): ReactNode {
+	if (typeof children === "string") {
+		const segments = parseTextWithCitations(children, urlMap);
+		if (segments.length === 1 && typeof segments[0] === "string") {
+			return children;
+		}
+		let ordinal = 0;
+		return segments.map((segment) =>
+			typeof segment === "string" ? segment : renderCitationToken(segment, ordinal++)
+		);
+	}
+
+	if (Array.isArray(children)) {
+		let ordinal = 0;
+		return children.map((child, childIndex) => {
+			if (typeof child === "string") {
+				const segments = parseTextWithCitations(child, urlMap);
+				if (segments.length === 1 && typeof segments[0] === "string") {
+					return child;
+				}
+				return (
+					<span key={`citation-seg-${childIndex}`}>
+						{segments.map((segment) =>
+							typeof segment === "string"
+								? segment
+								: renderCitationToken(segment, ordinal++)
+						)}
+					</span>
+				);
+			}
+			return child;
+		});
+	}
+
+	return children;
+}
diff --git a/surfsense_web/components/document-viewer.tsx b/surfsense_web/components/document-viewer.tsx
index 0f283e567..710a04ba3 100644
--- a/surfsense_web/components/document-viewer.tsx
+++ b/surfsense_web/components/document-viewer.tsx
@@ -32,7 +32,7 @@ export function DocumentViewer({ title, content, trigger }: DocumentViewerProps)
 					<DialogTitle>{title}</DialogTitle>
 				</DialogHeader>
 				<div className="mt-4">
-					<MarkdownViewer content={content} />
+					<MarkdownViewer content={content} enableCitations />
 				</div>
 			</DialogContent>
 		</Dialog>
diff --git a/surfsense_web/components/editor-panel/editor-panel.tsx b/surfsense_web/components/editor-panel/editor-panel.tsx
index df138e97e..eab07a91b 100644
--- a/surfsense_web/components/editor-panel/editor-panel.tsx
+++ b/surfsense_web/components/editor-panel/editor-panel.tsx
@@ -652,7 +652,7 @@ export function EditorPanelContent({
 					// Plate is heavy on multi-MB docs.
 					<div className="h-full overflow-y-auto px-5 py-4">
 						{largeDocAlert}
-						<MarkdownViewer content={editorDoc.source_markdown} />
+						<MarkdownViewer content={editorDoc.source_markdown} enableCitations />
 					</div>
 				) : renderInPlateEditor ? (
 					// Editable doc (FILE/NOTE) — Plate editing UX.
@@ -670,12 +670,17 @@ export function EditorPanelContent({
 								reserveToolbarSpace
 								defaultEditing={isEditing}
 								className="**:[[role=toolbar]]:bg-sidebar!"
+								// Render `[citation:N]` badges in view mode only.
+								// Edit mode keeps raw text so the user can edit/delete
+								// tokens directly. `local_file` never reaches this branch
+								// (handled by the source_code editor above).
+								enableCitations={!isEditing && !isLocalFileMode}
 							/>
 						</div>
 					</div>
 				) : (
 					<div className="h-full overflow-y-auto px-5 py-4">
-						<MarkdownViewer content={editorDoc.source_markdown} />
+						<MarkdownViewer content={editorDoc.source_markdown} enableCitations />
 					</div>
 				)}
 			</div>
diff --git a/surfsense_web/components/editor/plate-editor.tsx b/surfsense_web/components/editor/plate-editor.tsx
index 7f12d3cae..c42cb991e 100644
--- a/surfsense_web/components/editor/plate-editor.tsx
+++ b/surfsense_web/components/editor/plate-editor.tsx
@@ -8,9 +8,11 @@ import { useEffect, useMemo, useRef } from "react";
 import remarkGfm from "remark-gfm";
 import remarkMath from "remark-math";
 import { EditorSaveContext } from "@/components/editor/editor-save-context";
+import { CitationKit, injectCitationNodes } from "@/components/editor/plugins/citation-kit";
 import { type EditorPreset, presetMap } from "@/components/editor/presets";
 import { escapeMdxExpressions } from "@/components/editor/utils/escape-mdx";
 import { Editor, EditorContainer } from "@/components/ui/editor";
+import { preprocessCitationMarkdown } from "@/lib/citations/citation-parser";
 
 /** Live editor instance returned by `usePlateEditor`. */
 export type PlateEditorInstance = ReturnType<typeof usePlateEditor>;
@@ -65,6 +67,14 @@ export interface PlateEditorProps {
 	 * without modifying the core editor component.
 	 */
 	extraPlugins?: AnyPluginConfig[];
+	/**
+	 * Render `[citation:N]` and `[citation:URL]` tokens in the deserialized
+	 * markdown as interactive citation badges/popovers (mirrors chat). Only
+	 * meant for read-only views — when true, `onMarkdownChange` is suppressed
+	 * because the in-memory tree contains custom inline-void elements that
+	 * have no markdown serialize rule.
+	 */
+	enableCitations?: boolean;
 }
 
 function PlateEditorContent({
@@ -103,6 +113,7 @@ export function PlateEditor({
 	defaultEditing = false,
 	preset = "full",
 	extraPlugins = [],
+	enableCitations = false,
 }: PlateEditorProps) {
 	const lastMarkdownRef = useRef(markdown);
 	const lastHtmlRef = useRef(html);
@@ -145,6 +156,8 @@ export function PlateEditor({
 			...(onSave ? [SaveShortcutPlugin] : []),
 			// Consumer-provided extra plugins
 			...extraPlugins,
+			// Citation void inline element (read-only document viewer).
+			...(enableCitations ? CitationKit : []),
 			MarkdownPlugin.configure({
 				options: {
 					remarkPlugins: [remarkGfm, remarkMath, remarkMdx],
@@ -154,8 +167,18 @@ export function PlateEditor({
 		value: html
 			? (editor) => editor.api.html.deserialize({ element: html }) as Value
 			: markdown
-				? (editor) =>
-						editor.getApi(MarkdownPlugin).markdown.deserialize(escapeMdxExpressions(markdown))
+				? (editor) => {
+						if (!enableCitations) {
+							return editor
+								.getApi(MarkdownPlugin)
+								.markdown.deserialize(escapeMdxExpressions(markdown));
+						}
+						const { content: rewritten, urlMap } = preprocessCitationMarkdown(markdown);
+						const value = editor
+							.getApi(MarkdownPlugin)
+							.markdown.deserialize(escapeMdxExpressions(rewritten));
+						return injectCitationNodes(value as Descendant[], urlMap) as Value;
+					}
 				: undefined,
 	});
 
@@ -174,13 +197,22 @@ export function PlateEditor({
 	useEffect(() => {
 		if (!html && markdown !== undefined && markdown !== lastMarkdownRef.current) {
 			lastMarkdownRef.current = markdown;
-			const newValue = editor
-				.getApi(MarkdownPlugin)
-				.markdown.deserialize(escapeMdxExpressions(markdown));
+			let newValue: Descendant[];
+			if (enableCitations) {
+				const { content: rewritten, urlMap } = preprocessCitationMarkdown(markdown);
+				const deserialized = editor
+					.getApi(MarkdownPlugin)
+					.markdown.deserialize(escapeMdxExpressions(rewritten)) as Descendant[];
+				newValue = injectCitationNodes(deserialized, urlMap);
+			} else {
+				newValue = editor
+					.getApi(MarkdownPlugin)
+					.markdown.deserialize(escapeMdxExpressions(markdown)) as Descendant[];
+			}
 			editor.tf.reset();
-			editor.tf.setValue(newValue);
+			editor.tf.setValue(newValue as Value);
 		}
-	}, [html, markdown, editor]);
+	}, [html, markdown, editor, enableCitations]);
 
 	// When not forced read-only, the user can toggle between editing/viewing.
 	const canToggleMode = !readOnly && allowModeToggle;
@@ -205,6 +237,16 @@ export function PlateEditor({
 				// (initialized to true via usePlateEditor, toggled via ModeToolbarButton).
 				{...(readOnly ? { readOnly: true } : {})}
 				onChange={({ value }) => {
+					// View-only citation mode: skip serialization. The custom
+					// `citation` inline-void element has no markdown serialize
+					// rule, so emitting changes here would overwrite
+					// `lastMarkdownRef.current` (and downstream copy-to-clipboard
+					// state in EditorPanelContent) with a tree that loses every
+					// citation token. `enableCitations` is only ever set in
+					// read-only paths, so user input cannot reach this branch
+					// in practice — the guard exists for the initial Plate
+					// normalize emit.
+					if (enableCitations) return;
 					if (onHtmlChange && html) {
 						const serialized = slateToHtml(value as Descendant[]);
 						onHtmlChange(serialized);
diff --git a/surfsense_web/components/editor/plugins/citation-kit.tsx b/surfsense_web/components/editor/plugins/citation-kit.tsx
new file mode 100644
index 000000000..c90cb5e28
--- /dev/null
+++ b/surfsense_web/components/editor/plugins/citation-kit.tsx
@@ -0,0 +1,222 @@
+"use client";
+
+import { type FC } from "react";
+import { KEYS, type Descendant } from "platejs";
+import { createPlatePlugin, type PlateElementProps } from "platejs/react";
+import { InlineCitation, UrlCitation } from "@/components/assistant-ui/inline-citation";
+import {
+	CITATION_REGEX,
+	type CitationUrlMap,
+	parseTextWithCitations,
+} from "@/lib/citations/citation-parser";
+
+/**
+ * Plate inline-void node modeling a single `[citation:...]` reference.
+ *
+ * Modeled after the existing `MentionPlugin` pattern in
+ * `inline-mention-editor.tsx` — the only confirmed pattern in this repo
+ * for non-text inline UI. Inline-void elements satisfy Slate's invariant
+ * that the editor renders both atomic widgets and surrounding text
+ * cleanly without breaking selection / caret semantics.
+ */
+export type CitationElementNode = {
+	type: "citation";
+	kind: "chunk" | "doc" | "url";
+	chunkId?: number;
+	url?: string;
+	/** Original `[citation:...]` substring for traceability/debugging. */
+	rawText: string;
+	children: [{ text: "" }];
+};
+
+const CITATION_TYPE = "citation";
+
+const CitationElement: FC<PlateElementProps<CitationElementNode>> = ({
+	attributes,
+	children,
+	element,
+}) => {
+	const isUrl = element.kind === "url";
+	return (
+		<span {...attributes} className="inline-flex align-baseline">
+			<span contentEditable={false}>
+				{isUrl && element.url ? (
+					<UrlCitation url={element.url} />
+				) : element.chunkId !== undefined ? (
+					<InlineCitation chunkId={element.chunkId} isDocsChunk={element.kind === "doc"} />
+				) : null}
+			</span>
+			{children}
+		</span>
+	);
+};
+
+const CitationPlugin = createPlatePlugin({
+	key: CITATION_TYPE,
+	node: {
+		isElement: true,
+		isInline: true,
+		isVoid: true,
+		type: CITATION_TYPE,
+		component: CitationElement,
+	},
+});
+
+/** Plugin kit shape used elsewhere in the editor. */
+export const CitationKit = [CitationPlugin];
+
+// ---------------------------------------------------------------------------
+// Slate value transform — runs after MarkdownPlugin.deserialize
+// ---------------------------------------------------------------------------
+
+// Structural shapes used by the value transform. We cannot use Plate's
+// generic Element / Text type predicates directly because `Descendant` is a
+// constrained union and our predicates would over-narrow. Casting through
+// these row types keeps the walker readable without fighting the types.
+type SlateText = { text: string } & Record<string, unknown>;
+type SlateElement = { type?: string; children: Descendant[] } & Record<string, unknown>;
+
+function isText(node: Descendant): boolean {
+	return typeof (node as { text?: unknown }).text === "string";
+}
+
+function asText(node: Descendant): SlateText {
+	return node as unknown as SlateText;
+}
+
+function asElement(node: Descendant): SlateElement {
+	return node as unknown as SlateElement;
+}
+
+/**
+ * Element types whose subtrees we MUST NOT inject citation void elements
+ * into. Each rationale documented in the citation plan:
+ *  - `KEYS.codeBlock` / `code_line` — Plate's schema rejects inline elements
+ *    inside code containers; the user expects literal text inside code.
+ *  - `KEYS.link` — `<button>` inside `<a>` is invalid HTML and the link
+ *    swallows the citation click. Mirrors the `<a>` skip in
+ *    `MarkdownViewer`.
+ */
+const SKIP_SUBTREE_TYPES = new Set<string>([
+	KEYS.codeBlock,
+	"code_line",
+	KEYS.link,
+]);
+
+/**
+ * Build the marks portion of a Slate text node so we can preserve formatting
+ * (bold/italic/etc.) on the surrounding text fragments after we split.
+ */
+function copyMarks(textNode: SlateText): Record<string, unknown> {
+	const { text: _text, ...marks } = textNode;
+	return marks;
+}
+
+function makeCitationElement(
+	rawText: string,
+	segment: { kind: "url"; url: string } | { kind: "chunk"; chunkId: number; isDocsChunk: boolean }
+): CitationElementNode {
+	if (segment.kind === "url") {
+		return {
+			type: CITATION_TYPE,
+			kind: "url",
+			url: segment.url,
+			rawText,
+			children: [{ text: "" }],
+		};
+	}
+	return {
+		type: CITATION_TYPE,
+		kind: segment.isDocsChunk ? "doc" : "chunk",
+		chunkId: segment.chunkId,
+		rawText,
+		children: [{ text: "" }],
+	};
+}
+
+/**
+ * Re-extract the raw `[citation:...]` substrings that produced each parsed
+ * segment, in source order. Lets us preserve the original literal for
+ * `rawText` on the inline-void element.
+ */
+function extractRawCitationMatches(text: string): string[] {
+	const matches: string[] = [];
+	CITATION_REGEX.lastIndex = 0;
+	let m: RegExpExecArray | null = CITATION_REGEX.exec(text);
+	while (m !== null) {
+		matches.push(m[0]);
+		m = CITATION_REGEX.exec(text);
+	}
+	return matches;
+}
+
+function transformTextNode(node: SlateText, urlMap: CitationUrlMap): Descendant[] {
+	const segments = parseTextWithCitations(node.text, urlMap);
+	if (segments.length === 1 && typeof segments[0] === "string") {
+		return [node as unknown as Descendant];
+	}
+
+	const marks = copyMarks(node);
+	const rawMatches = extractRawCitationMatches(node.text);
+	const out: Descendant[] = [];
+	let citationIdx = 0;
+	let pendingText: string | null = null;
+
+	const flushText = () => {
+		// Slate inline-void adjacency: emit an empty text node (with copied
+		// marks) when the citation appears at the very start/end of the text
+		// node so neighbours of the void always have a text sibling.
+		out.push({ ...marks, text: pendingText ?? "" } as unknown as Descendant);
+		pendingText = null;
+	};
+
+	for (const segment of segments) {
+		if (typeof segment === "string") {
+			pendingText = (pendingText ?? "") + segment;
+		} else {
+			flushText();
+			const raw = rawMatches[citationIdx] ?? "";
+			out.push(makeCitationElement(raw, segment) as unknown as Descendant);
+			citationIdx += 1;
+			// Always reset pendingText so the next loop iteration emits a
+			// trailing empty text node if no further plain text follows.
+			pendingText = "";
+		}
+	}
+	flushText();
+
+	return out;
+}
+
+function transformChildren(children: Descendant[], urlMap: CitationUrlMap): Descendant[] {
+	const out: Descendant[] = [];
+	for (const child of children) {
+		if (isText(child)) {
+			out.push(...transformTextNode(asText(child), urlMap));
+			continue;
+		}
+		const elementChild = asElement(child);
+		const elementType = (elementChild.type ?? "") as string;
+		if (elementType && SKIP_SUBTREE_TYPES.has(elementType)) {
+			out.push(child);
+			continue;
+		}
+		out.push({
+			...elementChild,
+			children: transformChildren(elementChild.children, urlMap),
+		} as unknown as Descendant);
+	}
+	return out;
+}
+
+/**
+ * Walk a deserialized Slate value and replace every `[citation:...]`
+ * substring with a `citation` inline-void element. URL placeholders
+ * created by `preprocessCitationMarkdown` are resolved through `urlMap`.
+ *
+ * Subtrees of `code_block`, `code_line`, and `link` are returned as-is —
+ * see `SKIP_SUBTREE_TYPES` above.
+ */
+export function injectCitationNodes(value: Descendant[], urlMap: CitationUrlMap): Descendant[] {
+	return transformChildren(value, urlMap);
+}
diff --git a/surfsense_web/components/editor/utils/escape-mdx.ts b/surfsense_web/components/editor/utils/escape-mdx.ts
index cd5294b11..14839b9fc 100644
--- a/surfsense_web/components/editor/utils/escape-mdx.ts
+++ b/surfsense_web/components/editor/utils/escape-mdx.ts
@@ -7,7 +7,7 @@
 // break the MDX parser. This module sanitises them before deserialization.
 // ---------------------------------------------------------------------------
 
-const FENCED_OR_INLINE_CODE = /(```[\s\S]*?```|`[^`\n]+`)/g;
+import { FENCED_OR_INLINE_CODE } from "@/lib/markdown/code-regions";
 
 // Strip HTML comments that MDX cannot parse.
 // PDF converters emit <!-- PageHeader="..." -->, <!-- PageBreak -->, etc.
diff --git a/surfsense_web/components/layout/ui/tabs/DocumentTabContent.tsx b/surfsense_web/components/layout/ui/tabs/DocumentTabContent.tsx
index ac5463873..7ad78be41 100644
--- a/surfsense_web/components/layout/ui/tabs/DocumentTabContent.tsx
+++ b/surfsense_web/components/layout/ui/tabs/DocumentTabContent.tsx
@@ -316,10 +316,10 @@ export function DocumentTabContent({ documentId, searchSpaceId, title }: Documen
 									</Button>
 								</AlertDescription>
 							</Alert>
-							<MarkdownViewer content={doc.source_markdown} />
+							<MarkdownViewer content={doc.source_markdown} enableCitations />
 						</>
 					) : (
-						<MarkdownViewer content={doc.source_markdown} />
+						<MarkdownViewer content={doc.source_markdown} enableCitations />
 					)}
 				</div>
 			</div>
diff --git a/surfsense_web/components/markdown-viewer.tsx b/surfsense_web/components/markdown-viewer.tsx
index c4d73e30b..b2420711a 100644
--- a/surfsense_web/components/markdown-viewer.tsx
+++ b/surfsense_web/components/markdown-viewer.tsx
@@ -3,6 +3,12 @@ import { createMathPlugin } from "@streamdown/math";
 import { Streamdown, type StreamdownProps } from "streamdown";
 import "katex/dist/katex.min.css";
 import Image from "next/image";
+import { useMemo } from "react";
+import { processChildrenWithCitations } from "@/components/citations/citation-renderer";
+import {
+	type CitationUrlMap,
+	preprocessCitationMarkdown,
+} from "@/lib/citations/citation-parser";
 import { cn } from "@/lib/utils";
 
 const code = createCodePlugin({
@@ -21,8 +27,21 @@ interface MarkdownViewerProps {
 	content: string;
 	className?: string;
 	maxLength?: number;
+	/**
+	 * When true, render `[citation:N]` / `[citation:URL]` tokens as the
+	 * interactive citation badges/popovers used in chat. Default `false`
+	 * so callers that don't need citations are unchanged.
+	 *
+	 * Note: we deliberately do NOT override `<a>` to inject citations into
+	 * link text — that would produce `<button>` inside `<a>` (invalid
+	 * HTML). A `[citation:N]` token literally placed inside markdown link
+	 * text stays as raw text.
+	 */
+	enableCitations?: boolean;
 }
 
+const EMPTY_URL_MAP: CitationUrlMap = new Map();
+
 /**
  * If the entire content is wrapped in a single ```markdown or ```md
  * code fence, strip the fence so the inner markdown renders properly.
@@ -85,14 +104,45 @@ function convertLatexDelimiters(content: string): string {
 	return content;
 }
 
-export function MarkdownViewer({ content, className, maxLength }: MarkdownViewerProps) {
+export function MarkdownViewer({
+	content,
+	className,
+	maxLength,
+	enableCitations = false,
+}: MarkdownViewerProps) {
 	const isTruncated = maxLength != null && content.length > maxLength;
 	const displayContent = isTruncated ? content.slice(0, maxLength) : content;
-	const processedContent = convertLatexDelimiters(stripOuterMarkdownFence(displayContent));
+
+	// Preprocess for URL placeholders BEFORE LaTeX so GFM autolinks don't
+	// split `[citation:https://…]` apart. The preprocess is code-fence
+	// aware so citations inside fenced code stay literal.
+	const { processedContent, urlMap } = useMemo(() => {
+		const stripped = stripOuterMarkdownFence(displayContent);
+		if (!enableCitations) {
+			return {
+				processedContent: convertLatexDelimiters(stripped),
+				urlMap: EMPTY_URL_MAP,
+			};
+		}
+		const { content: rewritten, urlMap: map } = preprocessCitationMarkdown(stripped);
+		return {
+			processedContent: convertLatexDelimiters(rewritten),
+			urlMap: map,
+		};
+	}, [displayContent, enableCitations]);
+
+	// Phrasing/block renderers wrap their string children through the
+	// citation renderer when `enableCitations` is on. We deliberately do
+	// NOT override `<a>` (would produce <button> inside <a>) and we do
+	// NOT touch the inline/fenced `code` paths (citations stay literal
+	// inside code, matching markdown-text.tsx behavior).
+	const wrap = (children: React.ReactNode): React.ReactNode =>
+		enableCitations ? processChildrenWithCitations(children, urlMap) : children;
+
 	const components: StreamdownProps["components"] = {
 		p: ({ children, ...props }) => (
 			<p className="my-2" {...props}>
-				{children}
+				{wrap(children)}
 			</p>
 		),
 		a: ({ children, ...props }) => (
@@ -105,31 +155,49 @@ export function MarkdownViewer({ content, className, maxLength }: MarkdownViewer
 				{children}
 			</a>
 		),
-		li: ({ children, ...props }) => <li {...props}>{children}</li>,
+		li: ({ children, ...props }) => <li {...props}>{wrap(children)}</li>,
 		ul: ({ ...props }) => <ul className="list-disc pl-5 my-2" {...props} />,
 		ol: ({ ...props }) => <ol className="list-decimal pl-5 my-2" {...props} />,
 		h1: ({ children, ...props }) => (
 			<h1 className="text-2xl font-bold mt-6 mb-2" {...props}>
-				{children}
+				{wrap(children)}
 			</h1>
 		),
 		h2: ({ children, ...props }) => (
 			<h2 className="text-xl font-bold mt-5 mb-2" {...props}>
-				{children}
+				{wrap(children)}
 			</h2>
 		),
 		h3: ({ children, ...props }) => (
 			<h3 className="text-lg font-bold mt-4 mb-2" {...props}>
-				{children}
+				{wrap(children)}
 			</h3>
 		),
 		h4: ({ children, ...props }) => (
 			<h4 className="text-base font-bold mt-3 mb-1" {...props}>
-				{children}
+				{wrap(children)}
 			</h4>
 		),
-		blockquote: ({ ...props }) => (
-			<blockquote className="border-l-4 border-muted pl-4 italic my-2" {...props} />
+		h5: ({ children, ...props }) => (
+			<h5 className="text-sm font-bold mt-3 mb-1" {...props}>
+				{wrap(children)}
+			</h5>
+		),
+		h6: ({ children, ...props }) => (
+			<h6 className="text-xs font-bold mt-3 mb-1" {...props}>
+				{wrap(children)}
+			</h6>
+		),
+		strong: ({ children, ...props }) => (
+			<strong className="font-semibold" {...props}>
+				{wrap(children)}
+			</strong>
+		),
+		em: ({ children, ...props }) => <em {...props}>{wrap(children)}</em>,
+		blockquote: ({ children, ...props }) => (
+			<blockquote className="border-l-4 border-muted pl-4 italic my-2" {...props}>
+				{wrap(children)}
+			</blockquote>
 		),
 		hr: ({ ...props }) => <hr className="my-4 border-muted" {...props} />,
 		img: ({ src, alt, width: _w, height: _h, ...props }) => {
@@ -163,17 +231,21 @@ export function MarkdownViewer({ content, className, maxLength }: MarkdownViewer
 				<table className="w-full divide-y divide-border" {...props} />
 			</div>
 		),
-		th: ({ ...props }) => (
+		th: ({ children, ...props }) => (
 			<th
 				className="px-4 py-2.5 text-left text-sm font-semibold text-muted-foreground/80 bg-muted/30 border-r border-border/40 last:border-r-0"
 				{...props}
-			/>
+			>
+				{wrap(children)}
+			</th>
 		),
-		td: ({ ...props }) => (
+		td: ({ children, ...props }) => (
 			<td
 				className="px-4 py-2.5 text-sm border-t border-r border-border/40 last:border-r-0"
 				{...props}
-			/>
+			>
+				{wrap(children)}
+			</td>
 		),
 	};
 
diff --git a/surfsense_web/components/report-panel/report-panel.tsx b/surfsense_web/components/report-panel/report-panel.tsx
index 621cf13ce..7fafc9c3b 100644
--- a/surfsense_web/components/report-panel/report-panel.tsx
+++ b/surfsense_web/components/report-panel/report-panel.tsx
@@ -516,7 +516,7 @@ export function ReportPanelContent({
 				) : reportContent.content ? (
 					isReadOnly ? (
 						<div className="h-full overflow-y-auto px-5 py-4">
-							<MarkdownViewer content={reportContent.content} />
+							<MarkdownViewer content={reportContent.content} enableCitations />
 						</div>
 					) : (
 						<PlateEditor
@@ -531,6 +531,9 @@ export function ReportPanelContent({
 							reserveToolbarSpace
 							defaultEditing={isEditing}
 							className="[&_[role=toolbar]]:!bg-sidebar"
+							// Show citation badges in view mode; raw `[citation:N]`
+							// text in edit mode so users can edit/delete tokens.
+							enableCitations={!isEditing}
 						/>
 					)
 				) : (
diff --git a/surfsense_web/lib/citations/citation-parser.ts b/surfsense_web/lib/citations/citation-parser.ts
new file mode 100644
index 000000000..6333b0f97
--- /dev/null
+++ b/surfsense_web/lib/citations/citation-parser.ts
@@ -0,0 +1,134 @@
+// Pure citation parsing for `[citation:...]` tokens emitted by SurfSense
+// agents. No React imports — consumed by both the React renderer
+// (markdown surfaces) and the Plate value transform (document viewer).
+//
+// The same logic previously lived inline in
+// `components/assistant-ui/markdown-text.tsx` with module-level mutable
+// state. This module exposes a per-call URL map so multiple concurrent
+// renderers / SSR contexts can't race each other.
+
+import { FENCED_OR_INLINE_CODE } from "@/lib/markdown/code-regions";
+
+/**
+ * Matches `[citation:...]` with numeric IDs (incl. negative, doc- prefix,
+ * comma-separated), URL-based IDs from live web search, or `urlciteN`
+ * placeholders produced by `preprocessCitationMarkdown`.
+ *
+ * Also matches Chinese brackets 【】 and zero-width spaces that LLMs
+ * sometimes emit.
+ */
+export const CITATION_REGEX =
+	/[[【]\u200B?citation:\s*(https?:\/\/[^\]】\u200B]+|urlcite\d+|(?:doc-)?-?\d+(?:\s*,\s*(?:doc-)?-?\d+)*)\s*\u200B?[\]】]/g;
+
+/** A single parsed citation reference. */
+export type CitationToken =
+	| { kind: "url"; url: string }
+	| { kind: "chunk"; chunkId: number; isDocsChunk: boolean };
+
+/** Output of `parseTextWithCitations` — interleaved text + citation tokens. */
+export type ParsedSegment = string | CitationToken;
+
+/** Per-call URL placeholder map; key is `urlciteN`, value is the original URL. */
+export type CitationUrlMap = Map<string, string>;
+
+/** Result of preprocessing raw markdown for downstream parsing. */
+export interface PreprocessedCitations {
+	/** Markdown with `[citation:URL]` tokens rewritten to `[citation:urlciteN]`. */
+	content: string;
+	/** Lookup table to recover the original URL from each placeholder. */
+	urlMap: CitationUrlMap;
+}
+
+/** Pattern matching only URL-form citations (used during preprocessing). */
+const URL_CITATION_REGEX =
+	/[[【]\u200B?citation:\s*(https?:\/\/[^\]】\u200B]+)\s*\u200B?[\]】]/g;
+
+/**
+ * Replace `[citation:URL]` tokens with `[citation:urlciteN]` placeholders so
+ * GFM autolinks don't split the URL out of the brackets during markdown
+ * parsing. Returns both the rewritten content and a map for later lookup.
+ *
+ * Code-fence aware: skips fenced (``` ``` ```) and inline (`` ` ``) code
+ * regions so citation-shaped strings inside example code remain literal.
+ *
+ * Known limitations: `~~~` fences, 4-space indented code, and LaTeX math
+ * blocks are not skipped. Citation tokens inside those regions are rare in
+ * practice; documented in the plan.
+ */
+export function preprocessCitationMarkdown(content: string): PreprocessedCitations {
+	const urlMap: CitationUrlMap = new Map();
+	let counter = 0;
+
+	// Splitting on a regex with one capture group puts code regions at odd
+	// indexes (matched delimiters) and the surrounding text at even indexes.
+	// Only transform the even-indexed parts.
+	const parts = content.split(FENCED_OR_INLINE_CODE);
+	const transformed = parts.map((part, index) => {
+		if (index % 2 === 1) return part;
+		return part.replace(URL_CITATION_REGEX, (_match, url: string) => {
+			const key = `urlcite${counter++}`;
+			urlMap.set(key, url.trim());
+			return `[citation:${key}]`;
+		});
+	});
+
+	return { content: transformed.join(""), urlMap };
+}
+
+/**
+ * Parse a string into an array of plain text segments and citation tokens.
+ *
+ * Pure data — no React. The renderer module is responsible for mapping
+ * tokens to JSX. Negative chunk IDs are forwarded as-is so the consumer
+ * can decide how to render anonymous documents.
+ */
+export function parseTextWithCitations(
+	text: string,
+	urlMap: CitationUrlMap
+): ParsedSegment[] {
+	const segments: ParsedSegment[] = [];
+	let lastIndex = 0;
+	let match: RegExpExecArray | null;
+
+	CITATION_REGEX.lastIndex = 0;
+	match = CITATION_REGEX.exec(text);
+	while (match !== null) {
+		if (match.index > lastIndex) {
+			segments.push(text.substring(lastIndex, match.index));
+		}
+
+		const captured = match[1];
+
+		if (captured.startsWith("http://") || captured.startsWith("https://")) {
+			segments.push({ kind: "url", url: captured.trim() });
+		} else if (captured.startsWith("urlcite")) {
+			const url = urlMap.get(captured);
+			if (url) {
+				segments.push({ kind: "url", url });
+			}
+		} else {
+			const rawIds = captured.split(",").map((s) => s.trim());
+			for (const rawId of rawIds) {
+				const isDocsChunk = rawId.startsWith("doc-");
+				const chunkId = Number.parseInt(isDocsChunk ? rawId.slice(4) : rawId, 10);
+				if (!Number.isNaN(chunkId)) {
+					segments.push({ kind: "chunk", chunkId, isDocsChunk });
+				}
+			}
+		}
+
+		lastIndex = match.index + match[0].length;
+		match = CITATION_REGEX.exec(text);
+	}
+
+	if (lastIndex < text.length) {
+		segments.push(text.substring(lastIndex));
+	}
+
+	return segments.length > 0 ? segments : [text];
+}
+
+/** Type guard for the citation branch of `ParsedSegment`. */
+export function isCitationToken(segment: ParsedSegment): segment is CitationToken {
+	return typeof segment !== "string";
+}
diff --git a/surfsense_web/lib/markdown/code-regions.ts b/surfsense_web/lib/markdown/code-regions.ts
new file mode 100644
index 000000000..336a87acb
--- /dev/null
+++ b/surfsense_web/lib/markdown/code-regions.ts
@@ -0,0 +1,8 @@
+// Matches fenced (```...```) and inline (`...`) code regions. Used by MDX
+// escaping and citation preprocessing — single source of truth so future
+// edits stay in sync.
+//
+// String.split() with this capturing pattern places non-code parts at even
+// indexes and matched code regions at odd indexes — preserve odd-indexed
+// segments verbatim when transforming markdown.
+export const FENCED_OR_INLINE_CODE = /(```[\s\S]*?```|`[^`\n]+`)/g;