Tweak cursor, make sources appear after its done streaming

* Fix: authorize document versions by their root and speed up permission id sets

permitted_document_ids judged a version by its own owner and grants, so a
version whose owner had drifted from its root's was visible to the wrong
people and hidden from the right ones. Callers patched this individually by
mapping each document to its root first. The query itself was also slow on
MariaDB: the guardian grants were a UNION cast to integers and tested with
IN inside an OR with the owner checks, which MariaDB cannot materialize, so it
re-scans the user's grants for every document. At 20k documents that took
seconds for a user with a couple of hundred grants.

permitted_object_ids now casts the row key to a string, as guardian stores
object_pk, and tests it against a single uncorrelated UNION ALL of the user's
and groups' grants. No integer cast is needed and the user's groups are
matched with an IN subquery rather than a join through the membership table.
Postgres and SQLite build the grant set once, and MariaDB probes guardian's
unique indexes per row, which is cheap. A correlated EXISTS per grant also
fixed MariaDB but was up to 3x slower than before on Postgres and SQLite.

permitted_object_ids also takes an optional parent_field naming a
self-referencing foreign key whose target authorizes the row, and
permitted_document_ids passes root_document, so a version is visible exactly
when its root is. The helper that mapped documents to their roots at the call
sites is no longer needed, so the email, selection data, share link bundle,
trash, bulk download and bulk edit checks use the id set directly.

Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>

.github/workflows/ci-backend.yml

@@ -111,7 +111,7 @@ jobs:
         timeout-minutes: 12
         uses: $/.github/actions/apt-install
         with:
-          packages: unpaper tesseract-ocr imagemagick ghostscript poppler-utils qpdf
+          packages: unpaper tesseract-ocr imagemagick ghostscript poppler-utils
       - name: Configure ImageMagick
         run: |
           sudo cp docker/rootfs/etc/ImageMagick-6/paperless-policy.xml /etc/ImageMagick-7/policy.xml

docs/configuration.md

@@ -1315,17 +1315,29 @@ valid crontab(5) expression describing when to run.
 
 #### [`PAPERLESS_CONVERT_MEMORY_LIMIT=<num>`](#PAPERLESS_CONVERT_MEMORY_LIMIT) {#PAPERLESS_CONVERT_MEMORY_LIMIT}
 
-!!! warning
+: On smaller systems, or even in the case of Very Large Documents, the
+consumer may explode, complaining about how it's "unable to extend
+pixel cache". In such cases, try setting this to a reasonably low
+value, like 32. The default is to use whatever is necessary to do
+everything without writing to disk, and units are in megabytes.
 
-    Deprecated and has no effect, since PDF thumbnails no longer use
-    ImageMagick. It will be removed in a future release.
+    For more information on how to use this value, you should search the
+    web for "MAGICK_MEMORY_LIMIT".
+
+    Defaults to 0, which disables the limit.
 
 #### [`PAPERLESS_CONVERT_TMPDIR=<path>`](#PAPERLESS_CONVERT_TMPDIR) {#PAPERLESS_CONVERT_TMPDIR}
 
-!!! warning
+: Similar to the memory limit, if you've got a small system and your
+OS mounts /tmp as tmpfs, you should set this to a path that's on a
+physical disk, like /home/your_user/tmp or something. ImageMagick
+will use this as scratch space when crunching through very large
+documents.
 
-    Deprecated and has no effect, since PDF thumbnails no longer use
-    ImageMagick. It will be removed in a future release.
+    For more information on how to use this value, you should search the
+    web for "MAGICK_TMPDIR".
+
+    Default is none, which disables the temporary directory.
 
 #### [`PAPERLESS_APPS=<string>`](#PAPERLESS_APPS) {#PAPERLESS_APPS}
 

docs/setup.md

@@ -177,12 +177,12 @@ to a positive number to enable polling and disable native filesystem notificatio
     - `pkg-config` for mysqlclient (python dependency)
     - `fonts-liberation` for generating thumbnails for plain text
       files
-    - `imagemagick` >= 6 for image alpha handling
+    - `imagemagick` >= 6 for PDF conversion
     - `gnupg` for decrypting GPG-encrypted email
     - `libpq-dev` for PostgreSQL
     - `libmagic-dev` for mime type detection
     - `mariadb-client` for MariaDB compile time
-    - `poppler-utils` for thumbnail generation and barcode detection
+    - `poppler-utils` for barcode detection
 
     Use this list for your preferred package management:
 
@@ -416,8 +416,11 @@ to a positive number to enable polling and disable native filesystem notificatio
         You may need to change the path in the files. Example:
         `ExecStart=/opt/paperless/.local/bin/celery --app paperless worker --loglevel INFO`
 
-12. Harden ImageMagick by disabling formats that Paperless-ngx does not use.
-    PDF processing is not needed and should stay disabled.
+12. Configure ImageMagick to allow processing of PDF documents and disable
+    formats that Paperless-ngx does not use. Most distributions disable PDF
+    processing by default, since PDF documents can contain malware. If you
+    don't enable it, Paperless-ngx will fall back to Ghostscript for certain
+    steps such as thumbnail generation.
 
     Configure the active ImageMagick policy file (commonly
     `/etc/ImageMagick-6/policy.xml` or `/etc/ImageMagick-7/policy.xml`) and

paperless.conf.example

@@ -50,6 +50,8 @@ PAPERLESS_SECRET_KEY=change-me
 #PAPERLESS_OCR_ROTATE_PAGES=true
 #PAPERLESS_OCR_ROTATE_PAGES_THRESHOLD=12.0
 #PAPERLESS_OCR_USER_ARGS={}
+#PAPERLESS_CONVERT_MEMORY_LIMIT=0
+#PAPERLESS_CONVERT_TMPDIR=/var/tmp/paperless
 
 # Software tweaks
 

src-ui/jest.config.js

@@ -26,7 +26,7 @@ module.exports = {
     'abstract-paperless-service',
   ],
   transformIgnorePatterns: [
-    'node_modules/(?!.*(\\.mjs$|tslib|lodash-es|normalize-diacritics|@angular/common/locales/.*\\.js$))',
+    'node_modules/(?!.*(\\.mjs$|tslib|lodash-es|normalize-diacritics|marked|@angular/common/locales/.*\\.js$))',
   ],
   moduleNameMapper: {
     ...esmPreset.moduleNameMapper,

src-ui/package.json

@@ -30,6 +30,7 @@
     "bootstrap": "^5.3.8",
     "file-saver": "^2.0.5",
     "lodash-es": "^4.18.1",
+    "marked": "~18.0.14",
     "mime-names": "^1.0.0",
     "ngx-bootstrap-icons": "^1.9.3",
     "ngx-color": "^10.1.0",

src-ui/pnpm-lock.yaml

@@ -53,6 +53,9 @@ importers:
       lodash-es:
         specifier: ^4.18.1
         version: 4.18.1
+      marked:
+        specifier: ~18.0.14
+        version: 18.0.14
       mime-names:
         specifier: ^1.0.0
         version: 1.0.0
@@ -3226,6 +3229,11 @@ packages:
   make-error@1.3.6:
     resolution: {integrity: sha512-s8UhlNe7vPKomQhC1qFelMokr/Sc3AgNbso3n74mVPA5LTZwkB9NlXf4XPamLxJE8h0gh73rM94xvwRT2CVInw==}
 
+  marked@18.0.14:
+    resolution: {integrity: sha512-mBHK6FBHuBAlhgRe88w9F0O1AbwwXJUcQibUbC/QcdTbVGAD7aWza+xt3N6oT/jCZx3/OMeS+8rnuiHZcQ9s7A==}
+    engines: {node: '>= 20'}
+    hasBin: true
+
   material-colors@1.2.6:
     resolution: {integrity: sha512-6qE4B9deFBIa9YSpOc9O0Sgc43zTeVYbgDT5veRKSlB2+ZuHNoVVxA1L/ckMUayV9Ay9y7Z/SZCLcGteW9i7bg==}
 
@@ -7409,6 +7417,8 @@ snapshots:
 
   make-error@1.3.6: {}
 
+  marked@18.0.14: {}
+
   material-colors@1.2.6: {}
 
   merge-stream@2.0.0: {}

src-ui/src/app/components/chat/chat/chat.component.html

@@ -1,49 +1,85 @@
 
-<li ngbDropdown class="nav-item" (openChange)="onOpenChange($event)">
+<li ngbDropdown class="nav-item" display="dynamic" placement="bottom-end" [popperOptions]="popperOptions" (openChange)="onOpenChange($event)">
   <button class="btn navbar-action border-0" id="chatDropdown" ngbDropdownToggle aria-label="Chat" i18n-aria-label>
     <i-bs width="1.3em" height="1.3em" name="chatSquareDots"></i-bs>
   </button>
-  <div ngbDropdownMenu class="dropdown-menu-end shadow p-3" aria-labelledby="chatDropdown">
-    <div class="chat-container bg-light p-2">
-      <div class="chat-messages font-monospace small">
-        @for (message of messages(); track message) {
-          <div class="message d-flex flex-row small" [class.justify-content-end]="message.role === 'user'">
-            <div class="p-2 m-2" [class.bg-body]="message.role === 'user'">
-              <span class="text-break">
-                {{ message.content }}
-                @if (message.isStreaming) { <span class="blinking-cursor">|</span> }
-              </span>
-              @if (message.role === 'assistant' && message.references?.length) {
-                <div class="chat-references list-group mt-3">
-                  @for (reference of message.references; track reference.id) {
-                    <a class="list-group-item list-group-item-action d-flex text-primary text-break" [routerLink]="['/documents', reference.id]">
-                      <i-bs width="0.9em" height="0.9em" name="file-text" class="me-1"></i-bs><span>{{ reference.title }}</span>
-                    </a>
-                  }
-                </div>
-              }
-            </div>
-          </div>
+  <div ngbDropdownMenu class="dropdown-menu-end shadow p-0 overflow-hidden" aria-labelledby="chatDropdown">
+    <div class="d-flex align-items-center border-bottom px-3 py-2">
+      <i-bs width="1em" height="1em" name="stars" class="text-primary me-2"></i-bs>
+      <span class="fw-semibold small">
+        @if (documentId()) {
+          <ng-container i18n>Ask about this document</ng-container>
+        } @else {
+          <ng-container i18n>Ask about your documents</ng-container>
         }
-        <div #scrollAnchor></div>
-      </div>
+      </span>
+      @if (messages().length) {
+        <button class="btn btn-sm btn-link text-muted ms-auto p-0" type="button" (click)="clearChat()" title="Clear chat" i18n-title>
+          <i-bs width="1em" height="1em" name="arrow-counterclockwise"></i-bs>
+        </button>
+      }
+    </div>
 
-      <form class="chat-input">
-        <label class="visually-hidden" for="chatInput" i18n>Message</label>
-        <div class="input-group">
-          <input
-            #chatInput
-            id="chatInput"
-            class="form-control form-control-sm" name="chatInput" type="text"
+    <div class="chat-messages small px-2">
+      @if (!messages().length) {
+        <p class="text-muted text-center py-4 px-3 mb-0" i18n>Answers come from your documents, with links to their sources.</p>
+      }
+      @for (message of messages(); track message) {
+        <div class="message d-flex flex-row" [class.justify-content-end]="message.role === 'user'" [class.mt-3]="message.role === 'user' && !$first">
+          <div class="p-2 m-2" [class.bg-body-secondary]="message.role === 'user'" [class.rounded-3]="message.role === 'user'" [class.flex-grow-1]="message.role === 'assistant'">
+            @if (message.isStreaming && !message.content) {
+              <div class="d-flex align-items-center text-muted">
+                <div class="spinner-border spinner-border-sm me-2" role="status"></div>
+                @if (slowResponse()) {
+                  <ng-container i18n>Still thinking...</ng-container>
+                } @else {
+                  <ng-container i18n>Thinking...</ng-container>
+                }
+              </div>
+            } @else if (message.role === 'assistant') {
+              <div class="chat-markdown text-break text-wrap" [class.streaming]="message.isStreaming" [innerHTML]="message.content | markdown"></div>
+            } @else {
+              <span class="text-break">{{ message.content }}</span>
+            }
+            @if (message.role === 'assistant' && !message.isStreaming && message.references?.length) {
+              <div class="chat-references list-group mt-3">
+                <div class="list-group-item text-muted small py-1" i18n>Sources</div>
+                @for (reference of message.references; track reference.id) {
+                  <a class="list-group-item list-group-item-action d-flex align-items-center text-primary" [routerLink]="['/documents', reference.id]" [title]="reference.title">
+                    <i-bs width="0.9em" height="0.9em" name="file-text" class="me-2 flex-shrink-0"></i-bs><span class="text-truncate">{{ reference.title }}</span>
+                  </a>
+                }
+              </div>
+            }
+          </div>
+        </div>
+      }
+      <div #scrollAnchor></div>
+    </div>
+
+    <form class="chat-input border-top p-2">
+      <label class="visually-hidden" for="chatInput" i18n>Message</label>
+      <div class="input-group">
+        <input
+          #chatInput
+          id="chatInput"
+          class="form-control" name="chatInput" type="text"
             [placeholder]="placeholder"
             [disabled]="loading()"
             [ngModel]="input()"
             (ngModelChange)="input.set($event)"
             (keydown)="searchInputKeyDown($event)"
-            />
-          <button class="btn btn-sm btn-secondary" type="button" (click)="sendMessage()" [disabled]="loading()">Send</button>
-        </div>
-      </form>
-    </div>
+          />
+        @if (loading()) {
+          <button class="btn btn-outline-secondary" type="button" (click)="stopStreaming()" title="Stop" i18n-title aria-label="Stop" i18n-aria-label>
+            <i-bs width="1em" height="1em" name="x-lg"></i-bs>
+          </button>
+        } @else {
+          <button class="btn btn-primary" type="button" (click)="sendMessage()" title="Send" i18n-title aria-label="Send" i18n-aria-label>
+            <i-bs width="1em" height="1em" name="arrow-return-left"></i-bs>
+          </button>
+        }
+      </div>
+    </form>
   </div>
 </li>

src-ui/src/app/components/chat/chat/chat.component.scss

@@ -1,17 +1,13 @@
 .dropdown-menu {
-  width: var(--pngx-toast-max-width);
+  width: min(36rem, calc(100vw - 20px));
 }
 
 .chat-messages {
-  max-height: 350px;
+  max-height: min(60dvh, 36rem);
   overflow-y: auto;
   white-space: pre-wrap;
 }
 
-.chat-references {
-  font-family: var(--bs-font-sans-serif);
-}
-
 .dropdown-toggle::after {
   display: none;
 }
@@ -20,23 +16,44 @@
   white-space: initial;
 }
 
-@media screen and (max-width: 400px) {
-  :host ::ng-deep .dropdown-menu-end {
-    right: -3rem;
+@media screen and (max-width: 575.98px) {
+  .chat-messages {
+    max-height: calc(100dvh - 12rem);
   }
 }
 
-.blinking-cursor {
-  font-weight: bold;
-  font-size: 1.2em;
-  animation: blink 1s step-end infinite;
+.chat-markdown.streaming ::ng-deep {
+  > :last-child:not(ul, ol, table)::after,
+  > :is(ul, ol):last-child > li:last-child::after {
+    content: '';
+    display: inline-block;
+    width: 2px;
+    height: 1.05em;
+    margin-left: 2px;
+    background-color: currentColor;
+    vertical-align: -0.15em;
+    animation: blink 1s step-end infinite;
+  }
 }
 
 @keyframes blink {
-  from, to {
+  50% {
     opacity: 0;
   }
-  50% {
-    opacity: 1;
+}
+
+.chat-markdown ::ng-deep {
+  h1, h2, h3, h4, h5, h6 {
+    font-size: 1em;
+    font-weight: bold;
+  }
+
+  th, td {
+    padding: 0.15rem 0.4rem;
+    border: 1px solid var(--bs-border-color);
+  }
+
+  > :last-child {
+    margin-bottom: 0;
   }
 }

src-ui/src/app/components/chat/chat/chat.component.spec.ts

@@ -80,8 +80,16 @@ describe('ChatComponent', () => {
     expect(component.messages()[0].content).toBe('Hello')
     expect(component.loading()).toBe(true)
 
+    fixture.detectChanges()
+    const messagesEl = fixture.nativeElement.querySelector('.chat-messages')
+    expect(messagesEl.querySelector('.spinner-border')).not.toBeNull()
+    expect(messagesEl.textContent).toContain('Thinking...')
+
     mockStream$.next('Hi')
     expect(component.messages()[1].content).toBe('H')
+    fixture.detectChanges()
+    expect(messagesEl.querySelector('.spinner-border')).toBeNull()
+    expect(messagesEl.querySelector('.chat-markdown.streaming')).not.toBeNull()
     mockStream$.next('Hi there')
     // advance time to process the typewriter effect
     await jest.runAllTimersAsync()
@@ -96,6 +104,8 @@ describe('ChatComponent', () => {
     await fixture.whenStable()
     expect(component.loading()).toBe(false)
     expect(component.messages()[1].isStreaming).toBe(false)
+    fixture.detectChanges()
+    expect(messagesEl.querySelector('.chat-markdown.streaming')).toBeNull()
     expect(fixture.nativeElement.querySelector('#chatInput').disabled).toBe(
       false
     )
@@ -123,6 +133,10 @@ describe('ChatComponent', () => {
     mockStream$.next(
       `Hi there${CHAT_METADATA_DELIMITER}{"references":[{"id":42,"title":"Bread Recipe"}]}`
     )
+    fixture.detectChanges()
+    expect(fixture.nativeElement.querySelector('.chat-references')).toBeNull()
+
+    mockStream$.complete()
     jest.advanceTimersByTime(1000)
     fixture.detectChanges()
 
@@ -156,13 +170,96 @@ describe('ChatComponent', () => {
     component.input.set('Hello')
     component.sendMessage()
 
+    mockStream$.next('Partial')
     mockStream$.error('Error')
-    expect(component.messages()[1].content).toContain(
-      '⚠️ Error receiving response.'
+    expect(component.messages()[1].content).toBe(
+      'Partial\n\n⚠️ Error receiving response.'
     )
     expect(component.loading()).toBe(false)
   })
 
+  it('should keep streaming until the typewriter catches up', () => {
+    component.input.set('Hello')
+    component.sendMessage()
+
+    mockStream$.next('Hi there')
+    mockStream$.complete()
+    expect(component.messages()[1].isStreaming).toBe(true)
+    expect(component.loading()).toBe(true)
+
+    jest.advanceTimersByTime(1000)
+    expect(component.messages()[1].content).toBe('Hi there')
+    expect(component.messages()[1].isStreaming).toBe(false)
+    expect(component.loading()).toBe(false)
+  })
+
+  it('should show a slow response message while waiting', () => {
+    component.input.set('Hello')
+    component.sendMessage()
+    fixture.detectChanges()
+    const messagesEl = fixture.nativeElement.querySelector('.chat-messages')
+    expect(messagesEl.textContent).not.toContain('Still thinking...')
+
+    jest.advanceTimersByTime(8000)
+    fixture.detectChanges()
+    expect(messagesEl.textContent).toContain('Still thinking...')
+
+    mockStream$.next('Hi')
+    expect(component.slowResponse()).toBe(false)
+  })
+
+  it('should stop streaming and keep the partial response', () => {
+    component.input.set('Hello')
+    component.sendMessage()
+    fixture.detectChanges()
+    const stopButton = fixture.nativeElement.querySelector('.chat-input button')
+    expect(stopButton.getAttribute('aria-label')).toBe('Stop')
+
+    mockStream$.next('Hi there')
+    jest.advanceTimersByTime(20)
+    stopButton.click()
+
+    expect(mockStream$.observed).toBe(false)
+    expect(component.messages()[1].content).toBe('Hi there')
+    expect(component.messages()[1].isStreaming).toBe(false)
+    expect(component.loading()).toBe(false)
+
+    jest.advanceTimersByTime(1000)
+    expect(component.messages()[1].content).toBe('Hi there')
+  })
+
+  it('should drop the empty response when stopped before any content', () => {
+    component.input.set('Hello')
+    component.sendMessage()
+    component.stopStreaming()
+
+    expect(component.messages()).toHaveLength(1)
+    expect(component.loading()).toBe(false)
+    component.stopStreaming() // no-op once stopped
+    expect(component.messages()).toHaveLength(1)
+  })
+
+  it('should clear the conversation', () => {
+    component.input.set('Hello')
+    component.sendMessage()
+    mockStream$.next('Hi there')
+    fixture.detectChanges()
+
+    fixture.nativeElement.querySelector('button[title="Clear chat"]').click()
+    expect(component.messages()).toHaveLength(0)
+    expect(component.loading()).toBe(false)
+    expect(mockStream$.observed).toBe(false)
+
+    fixture.detectChanges()
+    jest.advanceTimersByTime(10)
+    expect(document.activeElement).toBe(
+      fixture.nativeElement.querySelector('#chatInput')
+    )
+    expect(fixture.nativeElement.textContent).toContain(
+      'Answers come from your documents'
+    )
+  })
+
   it('should enqueue typewriter chunks correctly', () => {
     const message = { content: '', role: 'assistant', isStreaming: true }
     component.enqueueTypewriter(null, message as any) // coverage for null
@@ -207,4 +304,33 @@ describe('ChatComponent', () => {
     component.searchInputKeyDown(event)
     expect(component.sendMessage).not.toHaveBeenCalled()
   })
+
+  it('should render markdown for assistant messages only', () => {
+    component.messages.set([
+      { role: 'user', content: '**user**' },
+      { role: 'assistant', content: '**assistant** <script>x</script>' },
+    ])
+    fixture.detectChanges()
+    const el: HTMLElement = fixture.nativeElement
+    expect(el.querySelector('.chat-markdown strong')?.textContent).toBe(
+      'assistant'
+    )
+    expect(el.querySelector('.chat-markdown script')).toBeNull()
+    expect(el.textContent).toContain('**user**')
+  })
+
+  it('should keep the destination hostname visible after sanitizing chat links', () => {
+    component.messages.set([
+      {
+        role: 'assistant',
+        content:
+          '[View original document](https://otherdomain.example/?data=secret)',
+      },
+    ])
+    fixture.detectChanges()
+    const el: HTMLElement = fixture.nativeElement
+    const link = el.querySelector('.chat-markdown a')
+    expect(link?.textContent).toBe('View original document')
+    expect(link?.nextElementSibling?.textContent).toBe('(otherdomain.example)')
+  })
 })

src-ui/src/app/components/chat/chat/chat.component.ts

@@ -10,12 +10,16 @@ import { FormsModule, ReactiveFormsModule } from '@angular/forms'
 import { NavigationEnd, Router, RouterModule } from '@angular/router'
 import { NgbDropdownModule } from '@ng-bootstrap/ng-bootstrap'
 import { NgxBootstrapIconsModule } from 'ngx-bootstrap-icons'
-import { filter, map } from 'rxjs'
+import { filter, map, Subscription } from 'rxjs'
+import { MarkdownPipe } from 'src/app/pipes/markdown.pipe'
 import {
   ChatMessage,
   ChatService,
   parseChatResponse,
 } from 'src/app/services/chat.service'
+import { pngxPopperOptions } from 'src/app/utils/popper-options'
+
+const SLOW_RESPONSE_DELAY_MS = 8000
 
 @Component({
   selector: 'pngx-chat',
@@ -25,6 +29,7 @@ import {
     RouterModule,
     NgxBootstrapIconsModule,
     NgbDropdownModule,
+    MarkdownPipe,
   ],
   templateUrl: './chat.component.html',
   styleUrl: './chat.component.scss',
@@ -34,6 +39,8 @@ export class ChatComponent implements OnInit {
   readonly loading = signal(false)
   readonly input = signal('')
   readonly documentId = signal<number>(undefined)
+  readonly slowResponse = signal(false)
+  public popperOptions = pngxPopperOptions
 
   private chatService: ChatService = inject(ChatService)
   private router: Router = inject(Router)
@@ -43,6 +50,11 @@ export class ChatComponent implements OnInit {
 
   private typewriterBuffer: string[] = []
   private typewriterActive = false
+  private typewriterTimeout: ReturnType<typeof setTimeout>
+  private streamSubscription: Subscription
+  private streamingMessage: ChatMessage
+  private streamFinished = false
+  private slowResponseTimeout: ReturnType<typeof setTimeout>
 
   public get placeholder(): string {
     return this.documentId()
@@ -81,47 +93,96 @@ export class ChatComponent implements OnInit {
     }
     this.messages.update((messages) => [...messages, assistantMessage])
     this.loading.set(true)
+    this.streamingMessage = assistantMessage
+    this.streamFinished = false
+    this.slowResponseTimeout = setTimeout(
+      () => this.slowResponse.set(true),
+      SLOW_RESPONSE_DELAY_MS
+    )
 
     let lastVisibleContent = ''
 
-    this.chatService.streamChat(this.documentId(), this.input()).subscribe({
-      next: (chunk) => {
-        const nextResponse = parseChatResponse(chunk)
+    this.streamSubscription = this.chatService
+      .streamChat(this.documentId(), this.input())
+      .subscribe({
+        next: (chunk) => {
+          this.clearSlowResponse()
+          const nextResponse = parseChatResponse(chunk)
 
-        if (nextResponse.content.length < lastVisibleContent.length) {
-          this.resetTypewriter(assistantMessage, nextResponse.content)
-          lastVisibleContent = nextResponse.content
-        } else {
-          const visibleDelta = nextResponse.content.substring(
-            lastVisibleContent.length
-          )
-          lastVisibleContent = nextResponse.content
-          this.enqueueTypewriter(visibleDelta, assistantMessage)
-        }
+          if (nextResponse.content.length < lastVisibleContent.length) {
+            this.resetTypewriter(assistantMessage, nextResponse.content)
+            lastVisibleContent = nextResponse.content
+          } else {
+            const visibleDelta = nextResponse.content.substring(
+              lastVisibleContent.length
+            )
+            lastVisibleContent = nextResponse.content
+            this.enqueueTypewriter(visibleDelta, assistantMessage)
+          }
 
-        assistantMessage.references = nextResponse.references
-        this.notifyMessagesChanged()
-      },
-      error: () => {
-        assistantMessage.content += '\n\n⚠️ Error receiving response.'
-        assistantMessage.isStreaming = false
-        this.notifyMessagesChanged()
-        this.loading.set(false)
-      },
-      complete: () => {
-        assistantMessage.isStreaming = false
-        this.notifyMessagesChanged()
-        this.loading.set(false)
-        this.scrollToBottom()
-      },
-    })
+          assistantMessage.references = nextResponse.references
+          this.notifyMessagesChanged()
+        },
+        error: () => {
+          this.flushTypewriter(assistantMessage)
+          assistantMessage.content += '\n\n⚠️ Error receiving response.'
+          this.finishStreaming(assistantMessage)
+        },
+        complete: () => {
+          // let the typewriter catch up before hiding the cursor
+          this.streamFinished = true
+          if (!this.typewriterActive) this.finishStreaming(assistantMessage)
+        },
+      })
 
     this.input.set('')
   }
 
-  private resetTypewriter(message: ChatMessage, content: string): void {
+  stopStreaming(): void {
+    const message = this.streamingMessage
+    if (!message?.isStreaming) return
+
+    this.streamSubscription?.unsubscribe()
+    this.flushTypewriter(message)
+    if (!message.content) {
+      this.messages.update((messages) => messages.filter((m) => m !== message))
+    }
+    this.finishStreaming(message)
+  }
+
+  clearChat(): void {
+    this.stopStreaming()
+    this.messages.set([])
+    setTimeout(() => this.chatInput?.nativeElement.focus())
+  }
+
+  private finishStreaming(message: ChatMessage): void {
+    this.clearSlowResponse()
+    this.streamFinished = false
+    message.isStreaming = false
+    this.notifyMessagesChanged()
+    this.loading.set(false)
+    this.scrollToBottom()
+  }
+
+  private clearSlowResponse(): void {
+    clearTimeout(this.slowResponseTimeout)
+    this.slowResponse.set(false)
+  }
+
+  private flushTypewriter(message: ChatMessage): void {
+    message.content += this.typewriterBuffer.join('')
+    this.cancelTypewriter()
+  }
+
+  private cancelTypewriter(): void {
+    clearTimeout(this.typewriterTimeout)
     this.typewriterBuffer = []
     this.typewriterActive = false
+  }
+
+  private resetTypewriter(message: ChatMessage, content: string): void {
+    this.cancelTypewriter()
     message.content = content
     this.notifyMessagesChanged()
     this.scrollToBottom()
@@ -141,6 +202,7 @@ export class ChatComponent implements OnInit {
   playTypewriter(message: ChatMessage): void {
     if (this.typewriterBuffer.length === 0) {
       this.typewriterActive = false
+      if (this.streamFinished) this.finishStreaming(message)
       return
     }
 
@@ -149,7 +211,8 @@ export class ChatComponent implements OnInit {
     this.notifyMessagesChanged()
     this.scrollToBottom()
 
-    setTimeout(() => this.playTypewriter(message), 10) // 10ms per character
+    // 10ms per character
+    this.typewriterTimeout = setTimeout(() => this.playTypewriter(message), 10)
   }
 
   private notifyMessagesChanged(): void {

src-ui/src/app/pipes/markdown.pipe.spec.ts

@@ -0,0 +1,90 @@
+import { MarkdownPipe } from './markdown.pipe'
+
+describe('MarkdownPipe', () => {
+  const pipe = new MarkdownPipe()
+
+  it('should return empty string for empty input', () => {
+    expect(pipe.transform(null)).toEqual('')
+    expect(pipe.transform(undefined)).toEqual('')
+    expect(pipe.transform('')).toEqual('')
+  })
+
+  it('should render basic markdown', () => {
+    const html = pipe.transform(
+      '**bold** _em_ `code`\n\n- one\n- two\n\n| a | b |\n|---|---|\n| 1 | 2 |'
+    )
+    expect(html).toContain('<strong>bold</strong>')
+    expect(html).toContain('<em>em</em>')
+    expect(html).toContain('<code>code</code>')
+    expect(html).toContain('<li>one</li>')
+    expect(html).toContain('<table>')
+  })
+
+  it('should escape raw html', () => {
+    const html = pipe.transform(
+      'hi <img src=x onerror="alert(1)"> <b>x</b>\n\n<script>alert(1)</script>'
+    )
+    expect(html).not.toContain('<img')
+    expect(html).not.toContain('<script')
+    expect(html).not.toContain('<b>')
+    expect(html).toContain('&lt;script&gt;')
+  })
+
+  it('should not render images', () => {
+    const html = pipe.transform(
+      '![secret](https://otherdomain.example/x.png?d=1)'
+    )
+    expect(html).not.toContain('<img')
+    expect(html).not.toContain('otherdomain.example')
+    expect(html).toContain('secret')
+  })
+
+  it('should render safe links with target and rel', () => {
+    const html = pipe.transform('[docs](https://docs.paperless-ngx.com "Docs")')
+    expect(html).toContain(
+      '<a target="_blank" rel="noopener noreferrer nofollow" href="https://docs.paperless-ngx.com" title="Docs">docs</a>'
+    )
+    expect(pipe.transform('[mail](mailto:a@b.c)')).toContain(
+      'href="mailto:a@b.c"'
+    )
+  })
+
+  it('should drop unsafe or relative links but keep their text', () => {
+    for (const href of [
+      'javascript:alert(1)',
+      'JaVaScRiPt:alert(1)',
+      'data:text/html,<script>alert(1)</script>',
+      'vbscript:msgbox',
+      '/api/documents/',
+    ]) {
+      const html = pipe.transform(`[click](${href})`)
+      expect(html).not.toContain('<a')
+      expect(html).toContain('click')
+    }
+  })
+
+  it('should show the actual hostname beside web links with misleading labels', () => {
+    for (const protocol of ['http', 'https']) {
+      const html = pipe.transform(
+        `[docs.paperless-ngx.com](${protocol}://docs.paperless-ngx.com@otherdomain.example/path?secret=1)`
+      )
+      expect(html).toContain(
+        '</a> <span class="text-muted small">(otherdomain.example)</span>'
+      )
+      expect(html).not.toContain('(docs.paperless-ngx.com)')
+    }
+  })
+
+  it('should omit the hostname label for mailto links', () => {
+    expect(pipe.transform('[mail](mailto:a@b.c)')).toBe(
+      '<p><a target="_blank" rel="noopener noreferrer nofollow" href="mailto:a@b.c">mail</a></p>\n'
+    )
+  })
+
+  it('should escape link titles', () => {
+    const html = pipe.transform(
+      '[x](https://a.example "a\\" onmouseover=\\"1")'
+    )
+    expect(html).not.toMatch(/"\s*onmouseover=/)
+  })
+})

src-ui/src/app/pipes/markdown.pipe.ts

@@ -0,0 +1,52 @@
+import { Pipe, PipeTransform } from '@angular/core'
+import { Marked, Renderer, Tokens } from 'marked'
+
+const ALLOWED_LINK_PROTOCOLS = new Set(['http:', 'https:', 'mailto:'])
+
+function allowedUrl(href: string): URL | null {
+  try {
+    const url = new URL(href)
+    return ALLOWED_LINK_PROTOCOLS.has(url.protocol) ? url : null
+  } catch {
+    return null
+  }
+}
+
+// Treat chat content as untrusted: no raw HTML, no images, and only
+// absolute links. All escaping is done by marked's default renderer, and
+// Angular's sanitizer runs on top of this as well.
+const renderer: Partial<Renderer> = {
+  html(this: Renderer, { raw, text }: Tokens.HTML | Tokens.Tag): string {
+    return this.text({ type: 'text', raw, text })
+  },
+  image(this: Renderer, { raw, text }: Tokens.Image): string {
+    return this.text({ type: 'text', raw, text })
+  },
+  link(this: Renderer, token: Tokens.Link): string {
+    const url = allowedUrl(token.href)
+    if (!url) return this.parser.parseInline(token.tokens)
+    const anchor = Renderer.prototype.link
+      .call(this, token)
+      .replace('<a ', '<a target="_blank" rel="noopener noreferrer nofollow" ')
+    if (!url.hostname) return anchor
+    const hostname = this.text({ type: 'text', raw: '', text: url.hostname })
+    return `${anchor} <span class="text-muted small">(${hostname})</span>`
+  },
+}
+
+const markdown = new Marked({
+  async: false,
+  gfm: true,
+  breaks: true,
+  renderer,
+})
+
+@Pipe({
+  name: 'markdown',
+})
+export class MarkdownPipe implements PipeTransform {
+  transform(value: string | null | undefined): string {
+    if (!value) return ''
+    return markdown.parse(value) as string
+  }
+}

src-ui/src/main.ts

@@ -32,6 +32,7 @@ import {
   arrowDownUp,
   arrowLeft,
   arrowRepeat,
+  arrowReturnLeft,
   arrowRight,
   arrowRightShort,
   arrowUpRight,
@@ -283,6 +284,7 @@ const icons = {
   arrowDownUp,
   arrowLeft,
   arrowRepeat,
+  arrowReturnLeft,
   arrowRight,
   arrowRightShort,
   arrowUpRight,

src-ui/tsconfig.json

@@ -15,7 +15,7 @@
     "target": "ES2022",
     "module": "es2020",
     "lib": [
-      "es2020",
+      "es2022",
       "dom"
     ],
     "useDefineForClassFields": false

src/documents/parsers.py

@@ -1,8 +1,8 @@
 from __future__ import annotations
 
 import logging
-import math
 import mimetypes
+import os
 import shutil
 import subprocess
 import tempfile
@@ -68,6 +68,58 @@ def get_supported_file_extensions() -> set[str]:
     return extensions
 
 
+def run_convert(
+    input_file,
+    output_file,
+    *,
+    density=None,
+    scale=None,
+    alpha=None,
+    strip=False,
+    trim=False,
+    type=None,
+    depth=None,
+    auto_orient=False,
+    use_cropbox=False,
+    extra=None,
+    logging_group=None,
+) -> None:
+    environment = os.environ.copy()
+    if settings.CONVERT_MEMORY_LIMIT:
+        # MAGICK_MEMORY_LIMIT sets the maximum amount of RAM the pixel cache can use.
+        # MAGICK_MAP_LIMIT sets the maximum amount of memory-mapped I/O allowed.
+        #
+        # For large-format documents  ImageMagick will hit the RAM limit and
+        # immediately try to "map" the remaining data. If MAGICK_MAP_LIMIT isn't
+        # also set, the process may trigger an OOM kill because the default
+        # system/policy map limit is often too restrictive for these massive bitmaps.
+        environment["MAGICK_MEMORY_LIMIT"] = settings.CONVERT_MEMORY_LIMIT
+        environment["MAGICK_MAP_LIMIT"] = settings.CONVERT_MEMORY_LIMIT
+    if settings.CONVERT_TMPDIR:
+        environment["MAGICK_TMPDIR"] = settings.CONVERT_TMPDIR
+
+    args = [settings.CONVERT_BINARY]
+    args += ["-density", str(density)] if density else []
+    args += ["-scale", str(scale)] if scale else []
+    args += ["-alpha", str(alpha)] if alpha else []
+    args += ["-strip"] if strip else []
+    args += ["-trim"] if trim else []
+    args += ["-type", str(type)] if type else []
+    args += ["-depth", str(depth)] if depth else []
+    args += ["-auto-orient"] if auto_orient else []
+    args += ["-define", "pdf:use-cropbox=true"] if use_cropbox else []
+    args += [str(input_file), str(output_file)]
+
+    logger.debug("Execute: " + " ".join(args), extra={"group": logging_group})
+
+    try:
+        run_subprocess(args, environment, logger)
+    except subprocess.CalledProcessError as e:
+        raise ParseError(f"Convert failed at {args}") from e
+    except Exception as e:  # pragma: no cover
+        raise ParseError("Unknown error running convert") from e
+
+
 def get_default_thumbnail() -> Path:
     """
     Returns the path to a generic thumbnail
@@ -75,168 +127,46 @@ def get_default_thumbnail() -> Path:
     return (Path(__file__).parent / "resources" / "document.webp").resolve()
 
 
-_THUMBNAIL_MAX_WIDTH = 500
-_THUMBNAIL_MAX_HEIGHT = 5000
-# Used only when the page geometry cannot be read
-_THUMBNAIL_FALLBACK_DPI = 150
-# Applied before supersampling, so tiny pages are not enlarged
-_THUMBNAIL_MAX_DPI = 300
-# Rendering at a multiple and downsampling keeps text crisper
-_THUMBNAIL_SUPERSAMPLE = 2
-
-
-def rasterize_pdf_page_to_png(
-    in_path: Path,
-    out_path: Path,
-    *,
-    dpi: int,
-    logging_group=None,
-) -> None:
-    """
-    Rasterizes the first page of a PDF to a PNG with pdftoppm.
-    """
-    # -singlefile drops the page number and -png appends ".png", so pass the
-    # path without its suffix
-    args = [
-        "pdftoppm",
-        "-f",
-        "1",
-        "-l",
-        "1",
-        "-r",
-        str(dpi),
-        "-png",
-        "-singlefile",
-        "-cropbox",
-        str(in_path),
-        str(out_path.with_suffix("")),
-    ]
-
-    logger.debug("Execute: " + " ".join(args), extra={"group": logging_group})
-
-    try:
-        run_subprocess(args, logger=logger)
-    except subprocess.CalledProcessError as e:
-        raise ParseError(f"pdftoppm failed at {args}") from e
-    except Exception as e:  # pragma: no cover
-        raise ParseError("Unknown error running pdftoppm") from e
-
-
-def encode_thumbnail_webp(
-    png_path: Path,
-    out_path: Path,
-    *,
-    supersample: int = 1,
-) -> None:
-    """
-    Flattens alpha onto white, undoes supersampling, shrinks to fit and saves as WebP.
-    """
-    from PIL import Image
-
-    try:
-        with Image.open(png_path) as im:
-            if im.mode in ("RGBA", "LA"):
-                flattened = Image.new("RGB", im.size, (255, 255, 255))
-                flattened.paste(im, mask=im.split()[-1])
-            else:
-                flattened = im.convert("RGB")
-
-        if supersample > 1:
-            flattened = flattened.resize(
-                (
-                    max(1, round(flattened.width / supersample)),
-                    max(1, round(flattened.height / supersample)),
-                ),
-                Image.Resampling.LANCZOS,
-            )
-
-        flattened.thumbnail((_THUMBNAIL_MAX_WIDTH, _THUMBNAIL_MAX_HEIGHT))
-        flattened.save(out_path, format="WEBP")
-    except (OSError, Image.DecompressionBombError) as e:
-        raise ParseError(f"Unable to encode thumbnail from {png_path}") from e
-
-
-def _compute_thumbnail_dpi(in_path: Path, logging_group=None) -> tuple[int, int]:
-    """
-    Returns (dpi, supersample). Unknown geometry is not supersampled, since
-    the render size cannot be bounded.
-    """
-    from paperless.parsers.utils import get_pdf_first_page_size_points
-
-    size = get_pdf_first_page_size_points(in_path)
-    if size is None:
-        logger.debug(
-            "Could not read PDF page size, using fallback DPI",
-            extra={"group": logging_group},
-        )
-        return _THUMBNAIL_FALLBACK_DPI, 1
-
-    width_pts, height_pts = size
-    dpi_for_width = _THUMBNAIL_MAX_WIDTH * 72 / width_pts
-    dpi_for_height = _THUMBNAIL_MAX_HEIGHT * 72 / height_pts
-    # Round up so the downsampled render is never a few pixels short of the
-    # target; the shrink-only clamp in encode_thumbnail_webp trims the excess.
-    dpi = max(
-        1,
-        math.ceil(min(_THUMBNAIL_MAX_DPI, dpi_for_width, dpi_for_height)),
-    )
-    # At the 1 DPI floor the page is already oversized, so do not supersample
-    return dpi, 1 if dpi == 1 else _THUMBNAIL_SUPERSAMPLE
-
-
-def _render_pdf_thumbnail(
-    in_path: Path,
-    png_path: Path,
-    out_path: Path,
-    logging_group=None,
-) -> None:
-    dpi, supersample = _compute_thumbnail_dpi(in_path, logging_group=logging_group)
-    rasterize_pdf_page_to_png(
-        in_path,
-        png_path,
-        dpi=dpi * supersample,
-        logging_group=logging_group,
-    )
-    encode_thumbnail_webp(png_path, out_path, supersample=supersample)
-
-
-def _repair_pdf_with_qpdf(in_path: Path, out_path: Path) -> None:
-    # qpdf exits 3 after a repair; --warning-exit-0 keeps that from failing
-    try:
-        shutil.copy(in_path, out_path)
-        run_subprocess(
-            ["qpdf", "--warning-exit-0", "--replace-input", str(out_path)],
-            logger=logger,
-        )
-    except (subprocess.CalledProcessError, OSError) as e:
-        raise ParseError(f"qpdf repair failed for {in_path}") from e
-
-
-def make_thumbnail_from_pdf_qpdf_fallback(
-    in_path: Path,
-    temp_dir: Path,
-    logging_group=None,
-) -> Path:
-    png_path = temp_dir / "page1_repaired.png"
-    out_path = temp_dir / "convert_qpdf.webp"
-    repaired_path = temp_dir / "repaired.pdf"
+def make_thumbnail_from_pdf_gs_fallback(in_path, temp_dir, logging_group=None) -> Path:
+    out_path: Path = Path(temp_dir) / "convert_gs.webp"
 
+    # if convert fails, fall back to extracting
+    # the first PDF page as a PNG using Ghostscript
     logger.warning(
-        "Thumbnail generation with pdftoppm failed, attempting qpdf repair and retry.",
+        "Thumbnail generation with ImageMagick failed, falling back "
+        "to ghostscript. Check your /etc/ImageMagick-x/policy.xml!",
         extra={"group": logging_group},
     )
+    # Ghostscript doesn't handle WebP outputs
+    gs_out_path: Path = Path(temp_dir) / "gs_out.png"
+    cmd = [settings.GS_BINARY, "-q", "-sDEVICE=pngalpha", "-o", gs_out_path, in_path]
 
     try:
-        _repair_pdf_with_qpdf(in_path, repaired_path)
-        _render_pdf_thumbnail(repaired_path, png_path, out_path, logging_group)
+        try:
+            run_subprocess(cmd, logger=logger)
+        except subprocess.CalledProcessError as e:
+            raise ParseError(f"Thumbnail (gs) failed at {cmd}") from e
+        # then run convert on the output from gs to make WebP
+        run_convert(
+            density=300,
+            scale="500x5000>",
+            alpha="remove",
+            strip=True,
+            trim=False,
+            auto_orient=True,
+            input_file=gs_out_path,
+            output_file=out_path,
+            logging_group=logging_group,
+        )
+
         return out_path
 
     except ParseError as e:
-        logger.error(f"Unable to make thumbnail after qpdf repair: {e}")
+        logger.error(f"Unable to make thumbnail with Ghostscript: {e}")
         # The caller might expect a generated thumbnail that can be moved,
         # so we need to copy it before it gets moved.
         # https://github.com/paperless-ngx/paperless-ngx/issues/3631
-        default_thumbnail_path = temp_dir / "document.webp"
+        default_thumbnail_path: Path = Path(temp_dir) / "document.webp"
         copy_file_with_basic_stats(get_default_thumbnail(), default_thumbnail_path)
         return default_thumbnail_path
 
@@ -245,18 +175,25 @@ def make_thumbnail_from_pdf(in_path: Path, temp_dir: Path, logging_group=None) -
     """
     The thumbnail of a PDF is just a 500px wide image of the first page.
     """
-    png_path: Path = temp_dir / "page1.png"
     out_path: Path = temp_dir / "convert.webp"
 
+    # Run convert to get a decent thumbnail
     try:
-        _render_pdf_thumbnail(in_path, png_path, out_path, logging_group)
-    except ParseError as e:
-        logger.error(f"Unable to make thumbnail with pdftoppm: {e}")
-        out_path = make_thumbnail_from_pdf_qpdf_fallback(
-            in_path,
-            temp_dir,
-            logging_group,
+        run_convert(
+            density=300,
+            scale="500x5000>",
+            alpha="remove",
+            strip=True,
+            trim=False,
+            auto_orient=True,
+            use_cropbox=True,
+            input_file=f"{in_path}[0]",
+            output_file=str(out_path),
+            logging_group=logging_group,
         )
+    except ParseError as e:
+        logger.error(f"Unable to make thumbnail with convert: {e}")
+        out_path = make_thumbnail_from_pdf_gs_fallback(in_path, temp_dir, logging_group)
 
     return out_path
 

src/documents/permissions.py

@@ -6,7 +6,9 @@ from django.contrib.auth.models import Permission
 from django.contrib.auth.models import User
 from django.contrib.contenttypes.models import ContentType
 from django.db.models import Case
+from django.db.models import CharField
 from django.db.models import Count
+from django.db.models import F
 from django.db.models import IntegerField
 from django.db.models import Model
 from django.db.models import Q
@@ -349,6 +351,7 @@ def permitted_object_ids(
     perm: str,
     *,
     include_deleted: bool = False,
+    parent_field: str | None = None,
 ) -> QuerySet[int]:
     """
     Generic version of ``permitted_document_ids`` for any model with an
@@ -357,6 +360,24 @@ def permitted_object_ids(
     soft-delete pattern (currently only ``Document``); for every other model
     it is accepted but has no effect, since those models have no soft-delete
     concept.
+
+    ``parent_field`` names a self-referencing foreign key whose target
+    authorizes the row (``Document.root_document``). A row with a parent is
+    visible exactly when its parent is, judged by the parent's owner and
+    grants, so the row's own owner and grants are ignored.
+
+    Guardian stores ``object_pk`` as a string, so the row key is cast to a
+    string and tested against the user's and groups' grants with a single
+    uncorrelated ``IN``. Postgres and SQLite build that set once. MariaDB
+    evaluates it as an index probe per row, which is cheap because the
+    lookups use guardian's unique indexes. Casting every ``object_pk`` to an
+    integer instead cannot use an index, and MariaDB cannot materialize it
+    inside the owner ``OR``, so it re-scans the user's grants for every row.
+    A correlated ``EXISTS`` per grant fixes MariaDB too, but Postgres and
+    SQLite re-run it for every row and end up slower than the original. The
+    user's groups are matched with an ``IN`` subquery rather than a join
+    through the membership table, which SQLite plans badly once the grant
+    tables grow.
     """
     has_soft_delete = hasattr(model, "global_objects")
     manager = (
@@ -364,8 +385,21 @@ def permitted_object_ids(
     )
     base_qs = manager.all().only("id", "owner")
 
+    owner_field, key_field = "owner", "pk"
+    if parent_field is not None:
+        owner_field, key_field = "authorizing_owner", "authorizing_id"
+        base_qs = base_qs.annotate(
+            authorizing_id=Coalesce(f"{parent_field}_id", "id"),
+            authorizing_owner=Case(
+                When(**{f"{parent_field}_id__isnull": True}, then=F("owner_id")),
+                default=F(f"{parent_field}__owner_id"),
+                output_field=IntegerField(),
+            ),
+        )
+    unowned = Q(**{f"{owner_field}__isnull": True})
+
     if user is None or not getattr(user, "is_authenticated", False):
-        return base_qs.filter(owner__isnull=True).values_list("id", flat=True)
+        return base_qs.filter(unowned).values_list("id", flat=True)
 
     # Deactivated users get nothing, deactivated superusers included, so this
     # has to come before the superuser shortcut. guardian's
@@ -389,21 +423,26 @@ def permitted_object_ids(
         "permission__content_type": content_type,
     }
 
-    user_perm_ids = (
-        UserObjectPermission.objects.filter(user=user, **perm_filter)
-        .annotate(object_pk_int=Cast("object_pk", IntegerField()))
-        .values_list("object_pk_int", flat=True)
-    )
-    group_perm_ids = (
-        GroupObjectPermission.objects.filter(group__user=user, **perm_filter)
-        .annotate(object_pk_int=Cast("object_pk", IntegerField()))
-        .values_list("object_pk_int", flat=True)
-    )
-    permitted_ids = user_perm_ids.union(group_perm_ids)
+    # Both grant sets are compared to the row key as strings, exactly as
+    # guardian stores them, and are uncorrelated, so each engine can build the
+    # set once instead of probing per row.
+    user_keys = UserObjectPermission.objects.filter(
+        user=user,
+        **perm_filter,
+    ).values_list("object_pk", flat=True)
+    group_keys = GroupObjectPermission.objects.filter(
+        group_id__in=user.groups.values("id"),
+        **perm_filter,
+    ).values_list("object_pk", flat=True)
+    permitted_keys = user_keys.union(group_keys, all=True)
 
-    return base_qs.filter(
-        Q(owner=user) | Q(owner__isnull=True) | Q(id__in=permitted_ids),
-    ).values_list("id", flat=True)
+    return (
+        base_qs.annotate(permitted_key=Cast(key_field, CharField(max_length=64)))
+        .filter(
+            Q(**{owner_field: user.pk}) | unowned | Q(permitted_key__in=permitted_keys),
+        )
+        .values_list("id", flat=True)
+    )
 
 
 ModelT = TypeVar("ModelT", bound=Model)
@@ -471,30 +510,16 @@ def permitted_document_ids(
     ``include_deleted=True`` for callers that need to check permission on
     soft-deleted documents (e.g. trash restore). This intentionally avoids
     ``get_objects_for_user`` to keep the subquery small and index-friendly.
-    """
-    return permitted_object_ids(user, Document, perm, include_deleted=include_deleted)
 
-
-def documents_without_permitted_root(
-    documents: QuerySet[Document],
-    user: User | None,
-    *,
-    perm: str = "view_document",
-    include_deleted: bool = False,
-) -> QuerySet[Document]:
+    A version is authorized by its root document, so a version's own owner and
+    grants never matter.
     """
-    The documents the user lacks ``perm`` on. Versions are authorized by their
-    root document, so a version's own owner is ignored. A single query, without
-    loading the documents or joining the root.
-    """
-    return documents.annotate(
-        root_id=Coalesce("root_document_id", "id"),
-    ).exclude(
-        root_id__in=permitted_document_ids(
-            user,
-            perm=perm,
-            include_deleted=include_deleted,
-        ),
+    return permitted_object_ids(
+        user,
+        Document,
+        perm,
+        include_deleted=include_deleted,
+        parent_field="root_document",
     )
 
 

src/documents/tests/test_parsers.py

@@ -1,22 +1,11 @@
-import subprocess
 from collections.abc import Generator
-from pathlib import Path
 
-import pikepdf
 import pytest
-from PIL import Image
 from pytest_django.fixtures import Settings
-from pytest_mock import MockerFixture
 
-from documents.parsers import ParseError
-from documents.parsers import _compute_thumbnail_dpi
-from documents.parsers import encode_thumbnail_webp
 from documents.parsers import get_default_file_extension
-from documents.parsers import get_default_thumbnail
 from documents.parsers import get_supported_file_extensions
 from documents.parsers import is_file_ext_supported
-from documents.parsers import make_thumbnail_from_pdf
-from documents.parsers import rasterize_pdf_page_to_png
 from paperless.parsers.registry import get_parser_registry
 from paperless.parsers.registry import reset_parser_registry
 from paperless.parsers.tesseract import RasterisedDocumentParser
@@ -136,374 +125,3 @@ class TestParserAvailability:
         assert is_file_ext_supported(".pdf")
         assert not is_file_ext_supported(".hsdfh")
         assert not is_file_ext_supported("")
-
-
-class TestComputeThumbnailDpi:
-    @pytest.mark.parametrize(
-        ("size", "expected"),
-        [
-            pytest.param((612.0, 792.0), (59, 2), id="letter-width-bound"),
-            pytest.param((792.0, 612.0), (46, 2), id="landscape-rounded-up"),
-            pytest.param((612.0, 100000.0), (4, 2), id="tall-strip-height-bound"),
-            pytest.param((200.0, 300.0), (180, 2), id="small-page-width-bound"),
-            pytest.param((72.0, 72.0), (300, 2), id="tiny-page-capped-at-300"),
-            pytest.param(
-                (1000000.0, 1000000.0),
-                (1, 1),
-                id="huge-page-minimum-one-unsupersampled",
-            ),
-            pytest.param(None, (150, 1), id="unreadable-geometry-fallback"),
-        ],
-    )
-    def test_dpi_from_page_size(
-        self,
-        mocker: MockerFixture,
-        tmp_path: Path,
-        size: tuple[float, float] | None,
-        expected: tuple[int, int],
-    ) -> None:
-        """
-        GIVEN:
-            - A first page of the given size, or unreadable geometry
-        WHEN:
-            - The thumbnail DPI is computed
-        THEN:
-            - The expected DPI and supersample factor are returned
-        """
-        mocker.patch(
-            "paperless.parsers.utils.get_pdf_first_page_size_points",
-            return_value=size,
-        )
-        assert _compute_thumbnail_dpi(tmp_path / "doc.pdf") == expected
-
-
-class TestMakeThumbnailFromPdf:
-    @pytest.fixture
-    def work_dir(self, tmp_path: Path) -> Path:
-        path = tmp_path / "work"
-        path.mkdir()
-        return path
-
-    @staticmethod
-    def _write_blank_pdf(path: Path, page_size: tuple[int, int] = (612, 792)) -> Path:
-        pdf = pikepdf.new()
-        pdf.add_blank_page(page_size=page_size)
-        pdf.save(path, object_stream_mode=pikepdf.ObjectStreamMode.disable)
-        return path
-
-    @pytest.mark.parametrize(
-        ("size", "expected_dpi", "expected_supersample"),
-        [
-            pytest.param((612.0, 792.0), 118, 2, id="known-geometry-2x"),
-            pytest.param(None, 150, 1, id="unreadable-geometry-plain-fallback"),
-        ],
-    )
-    def test_render_dpi_requested(
-        self,
-        mocker: MockerFixture,
-        tmp_path: Path,
-        work_dir: Path,
-        size: tuple[float, float] | None,
-        expected_dpi: int,
-        expected_supersample: int,
-    ) -> None:
-        """
-        GIVEN:
-            - Readable or unreadable page geometry
-        WHEN:
-            - A thumbnail is made
-        THEN:
-            - Rasterize and encode get the matching DPI and supersample factor
-        """
-        mocker.patch(
-            "paperless.parsers.utils.get_pdf_first_page_size_points",
-            return_value=size,
-        )
-        rasterize = mocker.patch("documents.parsers.rasterize_pdf_page_to_png")
-        encode = mocker.patch("documents.parsers.encode_thumbnail_webp")
-
-        make_thumbnail_from_pdf(tmp_path / "in.pdf", work_dir)
-
-        assert rasterize.call_args.kwargs["dpi"] == expected_dpi
-        assert encode.call_args.kwargs["supersample"] == expected_supersample
-
-    @pytest.mark.parametrize(
-        ("page_size", "expected_width"),
-        [
-            pytest.param((612, 792), 500, id="letter"),
-            pytest.param((792, 612), 500, id="landscape-letter"),
-            pytest.param((595, 842), 500, id="a4"),
-            pytest.param((200, 300), 500, id="small-page"),
-            pytest.param((72, 72), 300, id="tiny-page-capped"),
-        ],
-    )
-    def test_thumbnail_width(
-        self,
-        tmp_path: Path,
-        work_dir: Path,
-        page_size: tuple[int, int],
-        expected_width: int,
-    ) -> None:
-        """
-        GIVEN:
-            - A PDF whose first page has the given size in points
-        WHEN:
-            - A thumbnail is made from it
-        THEN:
-            - The thumbnail has the expected width
-        """
-        pdf_path = self._write_blank_pdf(tmp_path / "in.pdf", page_size)
-
-        thumb = make_thumbnail_from_pdf(pdf_path, work_dir)
-
-        assert thumb == work_dir / "convert.webp"
-        with Image.open(thumb) as im:
-            assert im.format == "WEBP"
-            assert im.width == expected_width
-
-    @classmethod
-    def _write_pdf_without_xref(cls, path: Path) -> Path:
-        """
-        Cuts off the xref and trailer, which pdftoppm cannot recover from but qpdf can.
-        """
-        cls._write_blank_pdf(path)
-        data = path.read_bytes()
-        path.write_bytes(data[: data.rindex(b"\nxref")])
-        return path
-
-    def test_qpdf_repair_produces_real_thumbnail(
-        self,
-        tmp_path: Path,
-        work_dir: Path,
-    ) -> None:
-        """
-        GIVEN:
-            - A PDF with its xref table and trailer cut off
-        WHEN:
-            - A thumbnail is made from it
-        THEN:
-            - The thumbnail is rendered from a qpdf repaired copy
-            - The original file is unchanged
-        """
-        pdf_path = self._write_pdf_without_xref(tmp_path / "broken.pdf")
-        original_bytes = pdf_path.read_bytes()
-
-        with pytest.raises(ParseError):
-            rasterize_pdf_page_to_png(pdf_path, work_dir / "probe.png", dpi=50)
-
-        thumb = make_thumbnail_from_pdf(pdf_path, work_dir)
-
-        assert thumb == work_dir / "convert_qpdf.webp"
-        with Image.open(thumb) as im:
-            assert im.format == "WEBP"
-            assert im.width == 500
-        assert pdf_path.read_bytes() == original_bytes
-
-    @pytest.mark.parametrize(
-        "qpdf_error",
-        [
-            pytest.param(subprocess.CalledProcessError(2, "qpdf"), id="qpdf-fails"),
-            pytest.param(None, id="repaired-still-unrenderable"),
-        ],
-    )
-    def test_double_failure_uses_default_thumbnail(
-        self,
-        mocker: MockerFixture,
-        tmp_path: Path,
-        work_dir: Path,
-        qpdf_error: subprocess.CalledProcessError | None,
-    ) -> None:
-        """
-        GIVEN:
-            - A PDF that cannot be rendered, even after qpdf repair
-        WHEN:
-            - A thumbnail is made from it
-        THEN:
-            - A copy of the default thumbnail is returned
-        """
-        mocker.patch(
-            "documents.parsers.rasterize_pdf_page_to_png",
-            side_effect=ParseError("Does not compute."),
-        )
-        if qpdf_error is not None:
-            mocker.patch("documents.parsers.run_subprocess", side_effect=qpdf_error)
-        pdf_path = self._write_blank_pdf(tmp_path / "in.pdf")
-
-        thumb = make_thumbnail_from_pdf(pdf_path, work_dir)
-
-        assert thumb == work_dir / "document.webp"
-        assert thumb.read_bytes() == get_default_thumbnail().read_bytes()
-
-
-class TestRasterizePdfPageToPng:
-    @staticmethod
-    def _write_pdf(
-        path: Path,
-        *,
-        crop_box: tuple[float, float, float, float] | None = None,
-        rotate: int | None = None,
-    ) -> Path:
-        pdf = pikepdf.new()
-        pdf.add_blank_page(page_size=(144, 72))
-        pdf.add_blank_page(page_size=(300, 300))
-        page = pdf.pages[0]
-        if crop_box is not None:
-            page.obj.CropBox = pikepdf.Array(crop_box)
-        if rotate is not None:
-            page.obj.Rotate = rotate
-        pdf.save(path)
-        return path
-
-    @pytest.mark.parametrize(
-        ("crop_box", "rotate", "expected_size"),
-        [
-            pytest.param(None, None, (144, 72), id="plain"),
-            pytest.param(None, 90, (72, 144), id="rotated-90"),
-            pytest.param((0, 0, 72, 36), None, (72, 36), id="crop-box"),
-        ],
-    )
-    def test_renders_first_page_to_exact_path(
-        self,
-        tmp_path: Path,
-        crop_box: tuple[float, float, float, float] | None,
-        rotate: int | None,
-        expected_size: tuple[int, int],
-    ) -> None:
-        """
-        GIVEN:
-            - A two page PDF, first page optionally cropped or rotated
-        WHEN:
-            - The first page is rasterized at 72 DPI
-        THEN:
-            - Only out_path is written, sized to the first page's crop and rotation
-        """
-        pdf_path = self._write_pdf(
-            tmp_path / "in.pdf",
-            crop_box=crop_box,
-            rotate=rotate,
-        )
-        out_dir = tmp_path / "out"
-        out_dir.mkdir()
-        out_path = out_dir / "page1.png"
-
-        rasterize_pdf_page_to_png(pdf_path, out_path, dpi=72)
-
-        assert list(out_dir.iterdir()) == [out_path]
-        with Image.open(out_path) as im:
-            assert im.format == "PNG"
-            assert im.size == expected_size
-
-    def test_failure_raises_parse_error(self, tmp_path: Path) -> None:
-        """
-        GIVEN:
-            - A file that is not a PDF
-        WHEN:
-            - Rasterization is attempted
-        THEN:
-            - A ParseError is raised
-        """
-        bad = tmp_path / "bad.pdf"
-        bad.write_bytes(b"not a pdf")
-
-        with pytest.raises(ParseError):
-            rasterize_pdf_page_to_png(bad, tmp_path / "page1.png", dpi=72)
-
-
-class TestEncodeThumbnailWebp:
-    @pytest.mark.parametrize(
-        ("mode", "color"),
-        [
-            pytest.param("RGBA", (0, 0, 0, 0), id="rgba"),
-            pytest.param("LA", (0, 0), id="la"),
-        ],
-    )
-    def test_alpha_flattened_onto_white(
-        self,
-        tmp_path: Path,
-        mode: str,
-        color: tuple[int, ...],
-    ) -> None:
-        """
-        GIVEN:
-            - A fully transparent PNG with an alpha channel
-        WHEN:
-            - It is encoded as a thumbnail
-        THEN:
-            - The WebP output is RGB with the transparency flattened to white
-        """
-        png_path = tmp_path / "in.png"
-        Image.new(mode, (20, 10), color).save(png_path)
-        out_path = tmp_path / "out.webp"
-
-        encode_thumbnail_webp(png_path, out_path)
-
-        with Image.open(out_path) as im:
-            assert im.format == "WEBP"
-            assert im.mode == "RGB"
-            assert im.size == (20, 10)
-            red, green, blue = im.getpixel((10, 5))
-            assert min(red, green, blue) >= 250
-
-    @pytest.mark.parametrize(
-        ("in_size", "supersample", "expected_size"),
-        [
-            pytest.param((1000, 2000), 1, (500, 1000), id="too-wide-shrunk"),
-            pytest.param((100, 10000), 1, (50, 5000), id="too-tall-shrunk"),
-            pytest.param((100, 200), 1, (100, 200), id="small-not-enlarged"),
-            pytest.param((1000, 1400), 2, (500, 700), id="2x-halved"),
-            pytest.param((1001, 1401), 2, (500, 700), id="2x-odd-rounded"),
-            pytest.param((600, 800), 2, (300, 400), id="2x-small-not-enlarged"),
-            pytest.param((900, 1200), 2, (450, 600), id="2x-below-clamp"),
-        ],
-    )
-    def test_size_clamped_and_downsampled(
-        self,
-        tmp_path: Path,
-        in_size: tuple[int, int],
-        supersample: int,
-        expected_size: tuple[int, int],
-    ) -> None:
-        """
-        GIVEN:
-            - A rendered image and its supersample factor
-        WHEN:
-            - It is encoded as a thumbnail with that factor
-        THEN:
-            - It is downsampled, fit within 500x5000 and never enlarged
-        """
-        png_path = tmp_path / "in.png"
-        Image.new("RGB", in_size, (255, 255, 255)).save(png_path)
-        out_path = tmp_path / "out.webp"
-
-        encode_thumbnail_webp(png_path, out_path, supersample=supersample)
-
-        with Image.open(out_path) as im:
-            assert im.size == expected_size
-
-    @pytest.mark.parametrize(
-        "error",
-        [
-            pytest.param(OSError("broken image"), id="os-error"),
-            pytest.param(Image.DecompressionBombError("too large"), id="bomb"),
-        ],
-    )
-    def test_decode_failure_raises_parse_error(
-        self,
-        tmp_path: Path,
-        mocker: MockerFixture,
-        error: Exception,
-    ) -> None:
-        """
-        GIVEN:
-            - Opening the rendered image fails
-        WHEN:
-            - It is encoded as a thumbnail
-        THEN:
-            - A ParseError is raised
-        """
-        png_path = tmp_path / "in.png"
-        Image.new("RGB", (10, 10)).save(png_path)
-        mocker.patch("PIL.Image.open", side_effect=error)
-
-        with pytest.raises(ParseError):
-            encode_thumbnail_webp(png_path, tmp_path / "out.webp")

src/documents/tests/test_permission_filtering_security.py

@@ -32,6 +32,8 @@ from paperless_testing.permissions import grant_global
 from paperless_testing.permissions import grant_object
 
 if TYPE_CHECKING:
+    from django.contrib.auth.models import User
+
     from paperless_testing.dirs import PaperlessDirs
 
 
@@ -178,6 +180,296 @@ class TestPermittedDocumentIdsIncludeDeleted:
         )
 
 
+@pytest.mark.django_db
+class TestPermittedDocumentIdsVersions:
+    """
+    A version is authorized by its root document: the version's own owner and
+    grants never matter.
+    """
+
+    @pytest.mark.parametrize(
+        ("root_owner", "version_owner", "expected_visible"),
+        [
+            pytest.param(
+                "other",
+                "nobody",
+                False,
+                id="unowned-version-of-private-root",
+            ),
+            pytest.param("other", "user", False, id="own-version-of-private-root"),
+            pytest.param("user", "other", True, id="foreign-version-of-own-root"),
+            pytest.param("user", "nobody", True, id="unowned-version-of-own-root"),
+            pytest.param("nobody", "other", True, id="private-version-of-unowned-root"),
+        ],
+    )
+    def test_version_follows_root_owner(
+        self,
+        root_owner: str,
+        version_owner: str,
+        *,
+        expected_visible: bool,
+    ) -> None:
+        """
+        GIVEN:
+            - A root document and a version with differing owners
+        WHEN:
+            - The permitted document ids are resolved for the user
+        THEN:
+            - The version is visible exactly when its root is
+        """
+        user = UserFactory()
+        owners = {"user": user, "other": UserFactory(), "nobody": None}
+        root = DocumentFactory(owner=owners[root_owner])
+        version = DocumentFactory(root_document=root, owner=owners[version_owner])
+
+        visible = set(permitted_document_ids(user))
+
+        assert (version.pk in visible) is expected_visible
+        assert (root.pk in visible) is expected_visible
+
+    @staticmethod
+    def grantee(user: User, kind: str) -> User | Group:
+        """The user itself, or a new group the user belongs to."""
+        if kind == "user":
+            return user
+        group = Group.objects.create(name="shared")
+        user.groups.add(group)
+        return group
+
+    @pytest.mark.parametrize(
+        "grantee_kind",
+        [pytest.param("user", id="user"), pytest.param("group", id="group")],
+    )
+    def test_grant_on_root_applies_to_version(self, grantee_kind: str) -> None:
+        """
+        GIVEN:
+            - A private root document shared with a user or one of their groups
+            - A version of it owned by someone else
+        WHEN:
+            - The permitted document ids are resolved for the user
+        THEN:
+            - Both the root and the version are visible
+            - A user without the grant sees neither
+        """
+        user = UserFactory()
+        stranger = UserFactory()
+        root = DocumentFactory(owner=UserFactory())
+        version = DocumentFactory(root_document=root, owner=UserFactory())
+        grant_object(self.grantee(user, grantee_kind), root, "view_document")
+
+        assert_visible_document_ids(
+            permitted_document_ids(user),
+            expected_visible=[root.pk, version.pk],
+            expected_hidden=[],
+        )
+        assert_visible_document_ids(
+            permitted_document_ids(stranger),
+            expected_visible=[],
+            expected_hidden=[root.pk, version.pk],
+        )
+
+    @pytest.mark.parametrize(
+        "grantee_kind",
+        [pytest.param("user", id="user"), pytest.param("group", id="group")],
+    )
+    def test_grant_on_version_is_ignored(self, grantee_kind: str) -> None:
+        """
+        GIVEN:
+            - A private root document
+            - A version with an explicit grant for the user or one of their groups
+        WHEN:
+            - The permitted document ids are resolved for the user
+        THEN:
+            - Neither the root nor the version is visible
+        """
+        user = UserFactory()
+        root = DocumentFactory(owner=UserFactory())
+        version = DocumentFactory(root_document=root, owner=UserFactory())
+        grant_object(self.grantee(user, grantee_kind), version, "view_document")
+
+        assert_visible_document_ids(
+            permitted_document_ids(user),
+            expected_visible=[],
+            expected_hidden=[root.pk, version.pk],
+        )
+
+    def test_grant_on_one_root_does_not_reach_another_roots_version(self) -> None:
+        """
+        GIVEN:
+            - Two private roots, each with a version
+            - The user may view only the first root
+        WHEN:
+            - The permitted document ids are resolved for the user
+        THEN:
+            - Only the first root and its version are visible
+        """
+        user = UserFactory()
+        first = DocumentFactory(owner=UserFactory())
+        first_version = DocumentFactory(root_document=first, owner=UserFactory())
+        second = DocumentFactory(owner=UserFactory())
+        second_version = DocumentFactory(root_document=second, owner=user)
+        grant_object(user, first, "view_document")
+
+        assert_visible_document_ids(
+            permitted_document_ids(user),
+            expected_visible=[first.pk, first_version.pk],
+            expected_hidden=[second.pk, second_version.pk],
+        )
+
+    def test_user_in_several_groups(self) -> None:
+        """
+        GIVEN:
+            - A user in two groups
+            - Two private roots shared with one group each, and a third shared with nobody
+            - A version of each root
+        WHEN:
+            - The permitted document ids are resolved for the user
+        THEN:
+            - The two shared roots and their versions are visible
+            - The third root and its version are not
+        """
+        user = UserFactory()
+        groups = [Group.objects.create(name=f"group{i}") for i in range(2)]
+        user.groups.add(*groups)
+        shared = [DocumentFactory(owner=UserFactory()) for _ in groups]
+        for root, group in zip(shared, groups, strict=True):
+            grant_object(group, root, "view_document")
+        unshared = DocumentFactory(owner=UserFactory())
+        shared_versions = [
+            DocumentFactory(root_document=root, owner=UserFactory()) for root in shared
+        ]
+        unshared_version = DocumentFactory(root_document=unshared, owner=None)
+
+        assert_visible_document_ids(
+            permitted_document_ids(user),
+            expected_visible=[
+                *(root.pk for root in shared),
+                *(version.pk for version in shared_versions),
+            ],
+            expected_hidden=[unshared.pk, unshared_version.pk],
+        )
+
+    def test_permission_is_resolved_through_the_root(self) -> None:
+        """
+        GIVEN:
+            - A private root document where the user may view and change
+        WHEN:
+            - The permitted ids are resolved for view, change and delete
+        THEN:
+            - The version is visible for view and change only
+        """
+        user = UserFactory()
+        root = DocumentFactory(owner=UserFactory())
+        version = DocumentFactory(root_document=root, owner=UserFactory())
+        grant_object(user, root, "view_document", "change_document")
+
+        assert version.pk in set(permitted_document_ids(user))
+        assert version.pk in set(permitted_document_ids(user, perm="change_document"))
+        assert version.pk in set(
+            permitted_document_ids(user, perm="documents.change_document"),
+        )
+        assert version.pk not in set(
+            permitted_document_ids(user, perm="delete_document"),
+        )
+
+    def test_anonymous_sees_versions_of_unowned_roots_only(self) -> None:
+        """
+        GIVEN:
+            - A version owned by nobody under a private root
+            - A version owned by someone under an unowned root
+        WHEN:
+            - The permitted document ids are resolved for an anonymous user
+        THEN:
+            - Only the version of the unowned root is visible
+        """
+        private_root = DocumentFactory(owner=UserFactory())
+        private_version = DocumentFactory(root_document=private_root, owner=None)
+        open_root = DocumentFactory(owner=None)
+        open_version = DocumentFactory(root_document=open_root, owner=UserFactory())
+
+        assert_visible_document_ids(
+            permitted_document_ids(AnonymousUser()),
+            expected_visible=[open_root.pk, open_version.pk],
+            expected_hidden=[private_root.pk, private_version.pk],
+        )
+
+    def test_deleted_versions_follow_their_deleted_root(self) -> None:
+        """
+        GIVEN:
+            - A soft-deleted root document and its version, which deleting the
+              root soft-deletes too; the version is owned by someone else
+        WHEN:
+            - The permitted document ids are resolved with and without deleted
+              documents
+        THEN:
+            - Nothing is visible by default
+            - With deleted documents included, the version is visible to the
+              root's owner and not to the version's own owner
+        """
+        owner = UserFactory()
+        version_owner = UserFactory()
+        root = DocumentFactory(owner=owner)
+        version = DocumentFactory(root_document=root, owner=version_owner)
+        root.delete()
+
+        assert not {root.pk, version.pk} & set(permitted_document_ids(owner))
+        assert_visible_document_ids(
+            permitted_document_ids(owner, include_deleted=True),
+            expected_visible=[root.pk, version.pk],
+            expected_hidden=[],
+        )
+        assert_visible_document_ids(
+            permitted_document_ids(version_owner, include_deleted=True),
+            expected_visible=[],
+            expected_hidden=[root.pk, version.pk],
+        )
+
+    @pytest.mark.parametrize(
+        "is_superuser",
+        [
+            pytest.param(False, id="regular-user"),
+            pytest.param(True, id="superuser"),
+        ],
+    )
+    def test_inactive_user_sees_no_versions(self, *, is_superuser: bool) -> None:
+        """
+        GIVEN:
+            - An inactive user, possibly a superuser, who owns a root and its version
+        WHEN:
+            - The permitted document ids are resolved for them
+        THEN:
+            - Nothing is visible
+        """
+        user = UserFactory(is_active=False, is_superuser=is_superuser)
+        root = DocumentFactory(owner=user)
+        version = DocumentFactory(root_document=root, owner=user)
+
+        assert_visible_document_ids(
+            permitted_document_ids(user),
+            expected_visible=[],
+            expected_hidden=[root.pk, version.pk],
+        )
+
+    def test_superuser_sees_all_versions(self) -> None:
+        """
+        GIVEN:
+            - A private root owned by someone else, with a version
+        WHEN:
+            - The permitted document ids are resolved for a superuser
+        THEN:
+            - Both the root and the version are visible
+        """
+        superuser = UserFactory(superuser=True)
+        root = DocumentFactory(owner=UserFactory())
+        version = DocumentFactory(root_document=root, owner=UserFactory())
+
+        assert_visible_document_ids(
+            permitted_document_ids(superuser),
+            expected_visible=[root.pk, version.pk],
+            expected_hidden=[],
+        )
+
+
 @pytest.mark.django_db
 class TestAiChatAllDocumentsPermissionBoundary:
     """

src/documents/views.py

@@ -174,7 +174,6 @@ from documents.permissions import TrashPermissions
 from documents.permissions import ViewDocumentsPermissions
 from documents.permissions import annotate_document_count_by_ids
 from documents.permissions import annotate_document_count_for_related_queryset
-from documents.permissions import documents_without_permitted_root
 from documents.permissions import get_document_count_filter_for_user
 from documents.permissions import get_objects_for_user_owner_aware
 from documents.permissions import has_global_statistics_permission
@@ -2117,7 +2116,7 @@ class DocumentViewSet(
         documents = Document.objects.filter(pk__in=document_ids)
         if (
             request.user is not None
-            and documents_without_permitted_root(documents, request.user).exists()
+            and documents.exclude(id__in=permitted_document_ids(request.user)).exists()
         ):
             return HttpResponseForbidden("Insufficient permissions")
 
@@ -3009,12 +3008,8 @@ class DocumentOperationPermissionMixin(PassUserMixin, DocumentSelectionMixin):
             user.has_perm(
                 "documents.change_document",
             )
-            and not Document.global_objects.filter(
-                pk__in=[doc.pk for doc in root_docs],
-            )
-            .exclude(
-                pk__in=permitted_document_ids(user, perm="change_document"),
-            )
+            and not Document.global_objects.filter(pk__in=documents)
+            .exclude(pk__in=permitted_document_ids(user, perm="change_document"))
             .exists()
         )
 
@@ -3627,7 +3622,7 @@ class SelectionDataView(DocumentSelectionMixin, GenericAPIView[Any]):
         documents = Document.objects.filter(pk__in=ids)
         if (
             documents.count() != len(ids)
-            or documents_without_permitted_root(documents, request.user).exists()
+            or documents.exclude(id__in=permitted_document_ids(request.user)).exists()
         ):
             return HttpResponseForbidden("Insufficient permissions")
 
@@ -4124,21 +4119,16 @@ class BulkDownloadView(DocumentSelectionMixin, GenericAPIView[Any]):
             validated_data=serializer.validated_data,
         )
         documents = Document.objects.filter(pk__in=ids)
-        versioned_documents = []
         compression = serializer.validated_data.get("compression")
         content = serializer.validated_data.get("content")
         follow_filename_format = serializer.validated_data.get("follow_formatting")
 
-        permitted_ids = set(permitted_document_ids(request.user))
-        for document in documents:
-            root_doc = get_root_document(document)
-            if root_doc.pk not in permitted_ids:
-                return HttpResponseForbidden("Insufficient permissions")
-            versioned_documents.append(
-                get_latest_version_for_root(
-                    root_doc,
-                ),
-            )
+        if documents.exclude(id__in=permitted_document_ids(request.user)).exists():
+            return HttpResponseForbidden("Insufficient permissions")
+        versioned_documents = [
+            get_latest_version_for_root(get_root_document(document))
+            for document in documents
+        ]
 
         if content == "both":
             strategy_class = OriginalAndArchiveStrategy
@@ -4811,7 +4801,7 @@ class ShareLinkBundleViewSet(PassUserMixin, ModelViewSet[ShareLinkBundle]):
             )
 
         denied_id = (
-            documents_without_permitted_root(documents_qs, request.user)
+            documents_qs.exclude(id__in=permitted_document_ids(request.user))
             .order_by("pk")
             .values_list("pk", flat=True)
             .first()
@@ -5657,11 +5647,12 @@ class TrashView(ListModelMixin, PassUserMixin):
             if doc_ids is not None
             else self.filter_queryset(self.get_queryset()).all()
         )
-        if documents_without_permitted_root(
-            docs,
-            request.user,
-            perm="delete_document",
-            include_deleted=True,
+        if docs.exclude(
+            id__in=permitted_document_ids(
+                request.user,
+                perm="delete_document",
+                include_deleted=True,
+            ),
         ).exists():
             return HttpResponseForbidden("Insufficient permissions")
         action = serializer.validated_data.get("action")

src/locale/en_US/LC_MESSAGES/django.po

@@ -2,7 +2,7 @@ msgid ""
 msgstr ""
 "Project-Id-Version: paperless-ngx\n"
 "Report-Msgid-Bugs-To: \n"
-"POT-Creation-Date: 2026-10-08 18:39+0000\n"
+"POT-Creation-Date: 2026-10-10 22:04+0000\n"
 "PO-Revision-Date: 2022-02-17 04:17\n"
 "Last-Translator: \n"
 "Language-Team: English\n"
@@ -1653,7 +1653,7 @@ msgid "workflow runs"
 msgstr ""
 
 #: documents/serialisers.py:516 documents/serialisers.py:873
-#: documents/serialisers.py:2903 documents/views.py:344 documents/views.py:2751
+#: documents/serialisers.py:2903 documents/views.py:343 documents/views.py:2750
 #: paperless_mail/serialisers.py:156
 msgid "Insufficient permissions."
 msgstr ""
@@ -1694,7 +1694,7 @@ msgstr ""
 msgid "Duplicate document identifiers are not allowed."
 msgstr ""
 
-#: documents/serialisers.py:2989 documents/views.py:4807
+#: documents/serialisers.py:2989 documents/views.py:4797
 #, python-format
 msgid "Documents not found: %(ids)s"
 msgstr ""
@@ -1945,40 +1945,40 @@ msgstr ""
 msgid ", "
 msgstr ""
 
-#: documents/views.py:337 documents/views.py:2748
+#: documents/views.py:336 documents/views.py:2747
 msgid "Invalid more_like_id"
 msgstr ""
 
-#: documents/views.py:1683
+#: documents/views.py:1682
 msgid "Invalid AI configuration."
 msgstr ""
 
-#: documents/views.py:1694
+#: documents/views.py:1693
 msgid "AI backend request timed out."
 msgstr ""
 
-#: documents/views.py:1706
+#: documents/views.py:1705
 msgid "AI backend rejected the request. Check logs for details."
 msgstr ""
 
-#: documents/views.py:2573 documents/views.py:2889
+#: documents/views.py:2572 documents/views.py:2888
 msgid "Specify only one of text, title_search, query, or more_like_id."
 msgstr ""
 
-#: documents/views.py:4823
+#: documents/views.py:4813
 #, python-format
 msgid "Insufficient permissions to share document %(id)s."
 msgstr ""
 
-#: documents/views.py:4870
+#: documents/views.py:4860
 msgid "Bundle is already being processed."
 msgstr ""
 
-#: documents/views.py:4934
+#: documents/views.py:4924
 msgid "The share link bundle is still being prepared. Please try again later."
 msgstr ""
 
-#: documents/views.py:4948
+#: documents/views.py:4938
 msgid "The share link bundle is unavailable."
 msgstr ""
 

src/paperless/parsers/utils.py

@@ -265,52 +265,6 @@ def get_page_count_for_pdf(
         return None
 
 
-def get_pdf_first_page_size_points(
-    path: Path,
-    log: logging.Logger | None = None,
-) -> tuple[float, float] | None:
-    """Return the first page's (width, height) in PDF points, post-rotation.
-
-    Uses the CropBox (MediaBox if absent), which must match pdftoppm's
-    ``-cropbox`` or the computed DPI targets the wrong box.
-
-    Swaps width and height for 90/270 rotation. ``page.rotation`` resolves
-    inherited and negative ``/Rotate`` values, a raw lookup does not.
-
-    Parameters
-    ----------
-    path:
-        Absolute path to the PDF file.
-    log:
-        Logger for warnings.  Falls back to the module-level logger when omitted.
-
-    Returns
-    -------
-    tuple[float, float] | None
-        ``(width_points, height_points)``, or ``None`` if the file cannot be
-        opened, has no pages, or the page box is degenerate.
-    """
-    import pikepdf
-
-    _log = log or logger
-    try:
-        with pikepdf.Pdf.open(path) as pdf:
-            if len(pdf.pages) == 0:
-                return None
-            page = pdf.pages[0]
-            llx, lly, urx, ury = (float(v) for v in page.cropbox)
-            width = abs(urx - llx)
-            height = abs(ury - lly)
-            if width <= 0 or height <= 0:
-                return None
-            if page.rotation in (90, 270):
-                width, height = height, width
-            return width, height
-    except Exception as e:
-        _log.warning("Could not determine PDF page size for %s: %s", path, e)
-        return None
-
-
 def extract_pdf_metadata(
     document_path: Path,
     log: logging.Logger | None = None,

src/paperless/settings/__init__.py

@@ -986,6 +986,8 @@ GNUPG_HOME = os.getenv("HOME", "/tmp")
 
 # Convert is part of the ImageMagick package
 CONVERT_BINARY = os.getenv("PAPERLESS_CONVERT_BINARY", "convert")
+CONVERT_TMPDIR = os.getenv("PAPERLESS_CONVERT_TMPDIR")
+CONVERT_MEMORY_LIMIT = os.getenv("PAPERLESS_CONVERT_MEMORY_LIMIT")
 
 GS_BINARY = os.getenv("PAPERLESS_GS_BINARY", "gs")
 

src/paperless/tests/parsers/test_tesseract_parser.py

@@ -15,10 +15,9 @@ from typing import TYPE_CHECKING
 
 import pytest
 from ocrmypdf import SubprocessOutputError
-from PIL import Image
 
 from documents.parsers import ParseError
-from documents.parsers import rasterize_pdf_page_to_png
+from documents.parsers import run_convert
 from paperless.models import ModeChoices
 from paperless.parsers import ParserProtocol
 from paperless.parsers.tesseract import RasterisedDocumentParser
@@ -281,71 +280,24 @@ class TestGetThumbnail:
         )
         assert thumb.is_file()
 
-    @pytest.mark.parametrize(
-        ("filename", "expected_height"),
-        [
-            pytest.param("simple-digital.pdf", 647, id="portrait-letter"),
-            pytest.param("rotated.pdf", 386, id="landscape"),
-        ],
-    )
-    def test_thumbnail_is_correct_format_and_size(
-        self,
-        tesseract_parser: RasterisedDocumentParser,
-        tesseract_samples_dir: Path,
-        filename: str,
-        expected_height: int,
-    ) -> None:
-        """
-        GIVEN:
-            - A portrait or landscape PDF
-        WHEN:
-            - A thumbnail is generated
-        THEN:
-            - A 500px wide WebP keeping the page's aspect ratio
-        """
-        thumb = tesseract_parser.get_thumbnail(
-            tesseract_samples_dir / filename,
-            "application/pdf",
-        )
-        with Image.open(thumb) as im:
-            assert im.format == "WEBP"
-            assert im.width == 500
-            assert im.height == pytest.approx(expected_height, abs=2)
-
-    def test_thumbnail_fallback_on_pdftoppm_error(
+    def test_thumbnail_fallback_on_convert_error(
         self,
         mocker: MockerFixture,
         tesseract_parser: RasterisedDocumentParser,
         tesseract_samples_dir: Path,
     ) -> None:
-        """
-        GIVEN:
-            - Rasterizing the original PDF fails
-        WHEN:
-            - A thumbnail is generated
-        THEN:
-            - The PDF is repaired with qpdf and a real thumbnail is rendered
-        """
-        original = tesseract_samples_dir / "simple-digital.pdf"
-
-        def _fail_on_original(in_path: Path, out_path: Path, **kwargs) -> None:
-            if in_path == original:
+        def _raise_on_pdf(input_file, output_file, **kwargs) -> None:
+            if ".pdf" in str(input_file):
                 raise ParseError("Does not compute.")
-            rasterize_pdf_page_to_png(in_path, out_path, **kwargs)
+            run_convert(input_file=input_file, output_file=output_file, **kwargs)
 
-        rasterize = mocker.patch(
-            "documents.parsers.rasterize_pdf_page_to_png",
-            side_effect=_fail_on_original,
+        mocker.patch("documents.parsers.run_convert", side_effect=_raise_on_pdf)
+
+        thumb = tesseract_parser.get_thumbnail(
+            tesseract_samples_dir / "simple-digital.pdf",
+            "application/pdf",
         )
-
-        thumb = tesseract_parser.get_thumbnail(original, "application/pdf")
-
-        assert rasterize.call_count == 2
         assert thumb.is_file()
-        assert thumb.name == "convert_qpdf.webp"
-        with Image.open(thumb) as im:
-            assert im.format == "WEBP"
-            assert im.width == 500
 
     def test_thumbnail_encrypted_pdf(
         self,

src/paperless/tests/test_parser_utils.py

@@ -6,10 +6,8 @@ import codecs
 from pathlib import Path
 from typing import TYPE_CHECKING
 
-import pikepdf
 import pytest
 
-from paperless.parsers.utils import get_pdf_first_page_size_points
 from paperless.parsers.utils import is_tagged_pdf
 from paperless.parsers.utils import pdf_born_digital_text
 from paperless.parsers.utils import post_process_text
@@ -72,149 +70,6 @@ class TestIsTaggedPdf:
         assert is_tagged_pdf(bad) is False
 
 
-class TestGetPdfFirstPageSizePoints:
-    @staticmethod
-    def _write_pdf(
-        path: Path,
-        *,
-        media_box: tuple[float, float, float, float] = (0, 0, 600, 800),
-        crop_box: tuple[float, float, float, float] | None = None,
-        page_rotate: int | None = None,
-        inherited_rotate: int | None = None,
-    ) -> Path:
-        pdf = pikepdf.new()
-        pdf.add_blank_page(page_size=(media_box[2], media_box[3]))
-        page = pdf.pages[0]
-        page.obj.MediaBox = pikepdf.Array(media_box)
-        if crop_box is not None:
-            page.obj.CropBox = pikepdf.Array(crop_box)
-        if page_rotate is not None:
-            page.obj.Rotate = page_rotate
-        if inherited_rotate is not None:
-            pdf.Root.Pages.Rotate = inherited_rotate
-        pdf.save(path)
-        return path
-
-    def test_letter_sample(self) -> None:
-        """
-        GIVEN:
-            - A US Letter sample PDF with no CropBox and no rotation
-        WHEN:
-            - The first page size is requested
-        THEN:
-            - The MediaBox size in points is returned
-        """
-        assert get_pdf_first_page_size_points(SAMPLES / "simple-digital.pdf") == (
-            612.0,
-            792.0,
-        )
-
-    @pytest.mark.parametrize(
-        ("rotate", "expected"),
-        [
-            pytest.param(0, (600.0, 800.0), id="rotate-0"),
-            pytest.param(90, (800.0, 600.0), id="rotate-90"),
-            pytest.param(180, (600.0, 800.0), id="rotate-180"),
-            pytest.param(270, (800.0, 600.0), id="rotate-270"),
-            pytest.param(-90, (800.0, 600.0), id="rotate-negative-90"),
-        ],
-    )
-    def test_page_rotation_swaps_dimensions(
-        self,
-        tmp_path: Path,
-        rotate: int,
-        expected: tuple[float, float],
-    ) -> None:
-        """
-        GIVEN:
-            - A portrait PDF page with /Rotate set directly on the page
-        WHEN:
-            - The first page size is requested
-        THEN:
-            - Width and height are swapped for quarter-turn rotations only
-        """
-        pdf_path = self._write_pdf(tmp_path / "rotated.pdf", page_rotate=rotate)
-        assert get_pdf_first_page_size_points(pdf_path) == expected
-
-    def test_inherited_rotation_swaps_dimensions(self, tmp_path: Path) -> None:
-        """
-        GIVEN:
-            - A page inheriting /Rotate 90 from the /Pages node
-        WHEN:
-            - The first page size is requested
-        THEN:
-            - Width and height are swapped
-        """
-        pdf_path = self._write_pdf(tmp_path / "inherited.pdf", inherited_rotate=90)
-        assert get_pdf_first_page_size_points(pdf_path) == (800.0, 600.0)
-
-    def test_crop_box_preferred_over_media_box(self, tmp_path: Path) -> None:
-        """
-        GIVEN:
-            - A PDF page with a CropBox smaller than its MediaBox
-        WHEN:
-            - The first page size is requested
-        THEN:
-            - The CropBox size is returned
-        """
-        pdf_path = self._write_pdf(
-            tmp_path / "cropped.pdf",
-            crop_box=(50, 100, 350, 500),
-        )
-        assert get_pdf_first_page_size_points(pdf_path) == (300.0, 400.0)
-
-    def test_degenerate_box_returns_none(self, tmp_path: Path) -> None:
-        """
-        GIVEN:
-            - A PDF page whose box has zero width
-        WHEN:
-            - The first page size is requested
-        THEN:
-            - None is returned
-        """
-        pdf_path = self._write_pdf(
-            tmp_path / "degenerate.pdf",
-            media_box=(0, 0, 600, 800),
-            crop_box=(100, 0, 100, 800),
-        )
-        assert get_pdf_first_page_size_points(pdf_path) is None
-
-    def test_nonexistent_path_returns_none(self) -> None:
-        """
-        GIVEN:
-            - A path that does not exist
-        WHEN:
-            - The first page size is requested
-        THEN:
-            - None is returned and nothing is raised
-        """
-        assert get_pdf_first_page_size_points(Path("/nonexistent/file.pdf")) is None
-
-    def test_corrupt_pdf_returns_none(self, tmp_path: Path) -> None:
-        """
-        GIVEN:
-            - A file that is not a PDF
-        WHEN:
-            - The first page size is requested
-        THEN:
-            - None is returned and nothing is raised
-        """
-        bad = tmp_path / "bad.pdf"
-        bad.write_bytes(b"not a pdf")
-        assert get_pdf_first_page_size_points(bad) is None
-
-    def test_encrypted_pdf_returns_none(self) -> None:
-        """
-        GIVEN:
-            - A password protected PDF
-        WHEN:
-            - The first page size is requested
-        THEN:
-            - None is returned and nothing is raised
-        """
-        assert get_pdf_first_page_size_points(SAMPLES / "encrypted.pdf") is None
-
-
 class TestPostProcessText:
     @pytest.mark.parametrize(
         ("source", "expected"),

src/paperless_ai/ai_classifier.py

@@ -4,8 +4,7 @@ from django.conf import settings
 from django.contrib.auth.models import User
 
 from documents.models import Document
-from documents.permissions import permitted_object_ids
-from documents.permissions import restrict_queryset_to_visible
+from documents.permissions import permitted_document_ids
 from documents.permissions import user_is_unrestricted
 from paperless.config import AIConfig
 from paperless_ai.base_model import ClassificationSuggestions
@@ -54,8 +53,8 @@ def _fulltext_similar_documents(
     active superuser - see user_is_unrestricted) is normalized to ``None``
     before calling, since the backend's permission filter has no superuser
     short-circuit of its own. Results are re-checked with
-    restrict_queryset_to_visible() since Tantivy's indexed permission fields
-    lag the DB via async reindexing.
+    permitted_document_ids() since Tantivy's indexed permission fields lag the
+    DB via async reindexing and judge a version by its own owner.
     """
     from documents.search import get_backend
 
@@ -69,10 +68,9 @@ def _fulltext_similar_documents(
     )
     if not unrestricted:
         allowed_ids = set(
-            restrict_queryset_to_visible(
-                Document.objects.filter(pk__in=similar_ids),
-                user,
-                "view_document",
+            Document.objects.filter(
+                pk__in=similar_ids,
+                id__in=permitted_document_ids(user),
             ).values_list("pk", flat=True),
         )
         similar_ids = [doc_id for doc_id in similar_ids if doc_id in allowed_ids]
@@ -200,13 +198,13 @@ def get_taxonomy_context(
             # quadratic scan in the vector store at best, and past ~32,763
             # documents a hard sqlite3.OperationalError (SQLite's
             # bound-parameter limit) at worst.
-            # permitted_object_ids() has its own superuser shortcut that would
+            # permitted_document_ids() has its own superuser shortcut that would
             # return every Document's id anyway, so this changes nothing about
             # which documents are considered -- only how we get there.
             visible_document_ids = (
                 None
                 if user_is_unrestricted(user)
-                else list(permitted_object_ids(user, Document, "view_document"))
+                else list(permitted_document_ids(user))
             )
             nodes = retrieve_similar_nodes(
                 document,

src/paperless_ai/tests/test_ai_classifier.py

@@ -552,7 +552,7 @@ class TestGetTaxonomyContextVisibility:
             return_value=[],
         )
         mock_permitted = mocker.patch(
-            "paperless_ai.ai_classifier.permitted_object_ids",
+            "paperless_ai.ai_classifier.permitted_document_ids",
         )
         user = UserFactory.create(is_superuser=True)
 
@@ -582,7 +582,7 @@ class TestGetTaxonomyContextVisibility:
             return_value=[],
         )
         mock_permitted = mocker.patch(
-            "paperless_ai.ai_classifier.permitted_object_ids",
+            "paperless_ai.ai_classifier.permitted_document_ids",
         )
 
         get_taxonomy_context(document, None)
@@ -611,16 +611,55 @@ class TestGetTaxonomyContextVisibility:
             return_value=[],
         )
         mock_permitted = mocker.patch(
-            "paperless_ai.ai_classifier.permitted_object_ids",
+            "paperless_ai.ai_classifier.permitted_document_ids",
             return_value=[1, 2, 3],
         )
         user = UserFactory.create(is_superuser=False)
 
         get_taxonomy_context(document, user)
 
-        mock_permitted.assert_called_once_with(user, Document, "view_document")
+        mock_permitted.assert_called_once_with(user)
         assert mock_retrieve.call_args.kwargs["document_ids"] == [1, 2, 3]
 
+    @pytest.mark.django_db
+    @override_settings(LLM_EMBEDDING_BACKEND="huggingface")
+    def test_version_of_private_root_is_not_visible(
+        self,
+        mocker: pytest_mock.MockerFixture,
+    ) -> None:
+        """
+        GIVEN:
+            - A private root document owned by someone else
+            - A version of it whose own owner is unset, as when the root
+              changed hands after the version was created
+        WHEN:
+            - get_taxonomy_context() is called for a non-superuser
+        THEN:
+            - Neither the root nor the version is in the visible ids passed to
+              retrieve_similar_nodes(), since a version follows its root
+        """
+        owner = UserFactory.create()
+        viewer = UserFactory.create(is_superuser=False)
+        root = DocumentFactory.create(content="private", owner=owner)
+        version = DocumentFactory.create(
+            content="private",
+            owner=None,
+            root_document=root,
+            version_index=1,
+        )
+        source = DocumentFactory.create(content="Some content", owner=viewer)
+        mock_retrieve = mocker.patch(
+            "paperless_ai.ai_classifier.retrieve_similar_nodes",
+            return_value=[],
+        )
+
+        get_taxonomy_context(source, viewer)
+
+        visible = mock_retrieve.call_args.kwargs["document_ids"]
+        assert source.pk in visible
+        assert root.pk not in visible
+        assert version.pk not in visible
+
 
 @pytest.mark.django_db
 class TestFulltextSimilarDocuments:
@@ -803,7 +842,7 @@ class TestFulltextSimilarDocuments:
             - _fulltext_similar_documents() is called with that user
         THEN:
             - Only the still-permitted document is returned - the DB
-              re-check via restrict_queryset_to_visible() must catch the
+              re-check via permitted_document_ids() must catch the
               document Tantivy's stale index still thinks is visible
         """
         owner = UserFactory.create()
@@ -834,6 +873,45 @@ class TestFulltextSimilarDocuments:
 
         assert [s["document_id"] for s in result] == [permitted.pk]
 
+    def test_excludes_version_of_private_root_for_regular_user(
+        self,
+        fulltext_backend: TantivyBackend,
+    ) -> None:
+        """
+        GIVEN:
+            - A regular user and a private root owned by someone else
+            - A version of that root with no owner of its own, which the
+              Tantivy index therefore treats as visible to everyone
+        WHEN:
+            - _fulltext_similar_documents() is called with that user
+        THEN:
+            - The version is not returned, since the DB re-check judges it by
+              its root
+        """
+        owner = UserFactory.create()
+        viewer = UserFactory.create(is_superuser=False)
+        source = DocumentFactory.create(
+            content="shared content phrase",
+            owner=viewer,
+        )
+        root = DocumentFactory.create(
+            content="shared content phrase",
+            owner=owner,
+        )
+        version = DocumentFactory.create(
+            content="shared content phrase",
+            owner=None,
+            root_document=root,
+            version_index=1,
+        )
+        fulltext_backend.add_or_update(source)
+        fulltext_backend.add_or_update(root)
+        fulltext_backend.add_or_update(version)
+
+        result = _fulltext_similar_documents(source, user=viewer, top_k=5)
+
+        assert version.pk not in [s["document_id"] for s in result]
+
 
 @pytest.mark.django_db
 @override_settings(LLM_EMBEDDING_BACKEND="huggingface")