Skip to content

Web application

Use the local web application to find documents, read any saved version, and export the records you select. You can also import files and email, organize them with tags, and review configured processing before it runs. Start it with:

docbank web

Docbank starts or reconnects to the vault's daemon and opens the browser. The daemon is the background process that owns the vault. The browser uses its authenticated HTTP API. It does not open the database or stored files directly. The daemon accepts connections only from the local machine.

Choose a task:

The browser has limited permissions. See Browser authentication for their scope and Current limits for workflows that require the CLI or another API client.

The Docbank web application showing a synthetic vault tree and the selected document's authority card.

Select a document to see its stable identity and content hash beside the table.

Browse the vault

Documents and Photos at the top of the sidebar switch workspaces. Photos shows Library, Map, Albums, People, and your album list in the sidebar. Documents/Photos switches keep your place until the session locks. Leaving an album resets its view. Open Library at /photos, the album index at /photos/albums, People at /photos/people, or an album at /photos/albums/{id}. Map requires tile configuration and a restart. See Browse photos.

Use the sidebar for All files, Searches, Productions when sets exist, Imports, and Trash. Searches, Productions, Imports, and Trash open full pages at /searches, /productions, /imports, and /trash. Back and Forward keep the page. Reload keeps the page with API-key sign-in; docbank web starts at the root. Export offers Download files… and Bates-numbered PDFs…. System groups Disk use, Running now, Backups, Protected folders, and Usage reporting. On a narrow screen, Open navigation reveals the same menu.

Folder breadcrumbs let you jump to any parent. Each file has an icon and a plain type name, so you do not need to recognize its media type.

  • Click a row once to inspect it. The authority card updates without opening or downloading the file.
  • Double-click a folder, select it and press Enter, or use Open folder in the authority card to navigate into it.
  • Use the back arrow to restore the previous folder or search view, including its inspected row and sort order.
  • Click Document, Type, Size, or Modified to sort. Click the active heading again to reverse its direction. Directories remain grouped ahead of files in folder views. Narrow windows hide Modified, then Type. Deep folder paths show the first folder and the last two, with the folders between them in the … menu.
  • Use refresh to reload the current directory by its ID. If another client renamed or moved that directory, the browser adopts its current canonical path.

The authority card is the details panel for the selected file or folder. A node ID identifies that entry even after a move or rename. Its revision is a change counter that Docbank uses to reject changes based on an outdated view.

The browser selects the first row after loading a folder. A selected directory shows its path and modification time. A selected file also shows its exact logical size. Open Details for the node ID, media type, current-version UUID, and SHA-256 content identity. The copy buttons copy the complete UUID or digest even when the card wraps it across lines. Every selected node also shows its assigned tag names. Hovering a tag shows its UUID and vault-wide assignment count. The tag stack expands in place when a node carries more than six tags.

The slide-out navigation menu on a narrow screen

Use keyboard shortcuts

Press ? or choose the keyboard button in the top bar to see every shortcut. Press / to focus search, j or k to inspect the next or previous loaded row, Space to check or uncheck the inspected file, Enter to open the inspected folder, and Escape to clear page selection. Row navigation stops at the loaded page boundary and never fetches another page.

In Photos, B adds selected photos to the last chosen album, or opens the album picker when no target is set.

The same help dialog can assign digits 1 through 9 to tag definitions. Bindings are separate for each vault and stay in this browser for the current daemon run. Reopening the page keeps them. Restarting the daemon or clearing browser data requires setting them again.

Pressing a configured digit toggles that tag on the inspected file only, regardless of checked documents. Docbank waits for the file's complete assigned-tag list and sends its displayed revision with the change. A stale revision remains visible as an error and is never retried against newer state.

App shortcuts pause while a text field, control, or dialog owns the keyboard.

Keyboard shortcuts and a configured tag hotkey

Select documents on this page

Check a document's checkbox to mark it without changing the document shown in the authority card. The bottom dock shows how many documents are checked.

  • Hold Shift while clicking another checkbox to select or clear a range in the displayed sort order. With the keyboard, focus a checkbox and press Shift+Space.
  • Use the header checkbox or Select visible documents to check the files currently shown. Folders remain available for navigation and are not selected.
  • Choose Clear selection or close the dock to clear the checked documents.

Selection covers only the loaded page, even when more results exist. Changing the folder, query, or tag clears it when the new rows arrive. If an error leaves the previous results on screen, their selection stays. Refreshing the same view keeps only checked documents that are still visible. Back navigation and ending the browser session also clear the selection.

Two documents selected on the current page while the authority card shows one document's details.

Export selected rows as CSV

Check file rows and choose Export page CSV. The download preserves their displayed order and includes the metadata already loaded, such as document IDs, versions, paths, sizes, timestamps, and hashes. Unavailable fields stay blank. It does not download document contents, fetch additional metadata, or include results beyond the loaded page. Formula-leading text is escaped for spreadsheet imports, and international filenames are preserved.

Checked documents and the page CSV export action

Edit a complete query

Enter text in the top bar and choose Plain words or Query syntax from its menu. Choose Advanced to open Filters and sort. Select the sort field, direction, and documents per page. More holds the processing profile selector, shown when more than one profile is available, and Filters JSON for structured filters. Search runs the top-bar text with these settings.

Saved and linked queries also show their expression, labeled Plain words or Query syntax. Syntax help appears for query syntax. The daemon validates each edited draft after a short pause. Errors appear beside the expression or filters. Focus query error selects the reported part of an expression. A newer edit cancels an older validation request.

Save… opens a new definition in Searches with the entire draft. Run beside a saved search opens its results. From an import in Imports, Search this import starts a draft scoped to that import. Escape, clicking outside, or toggling Advanced closes the modal. Edited drafts stay in the tab, with a dot on Advanced; reopening restores unfinished input too. Valid query edits also update the URL fragment. Cancel discards the draft and clears its fragment. Closing an unchanged new draft leaves no dot.

Validation does not execute a search. Once the draft is valid, Search creates a frozen result through the complete query endpoint, which uses word search. Clear the search box to return to ordinary search. See field-aware query syntax.

The complete query editor with a validated field and Boolean expression

Work with a frozen query

A completed run becomes the accepted frozen snapshot for the tab. Draft edits remain separate until you choose Search again. Choose 50, 100, or 250 documents per page before the run. The workspace shows the result count, Previous page and Next page controls when results span more than one page, and the node, content version, hash, size, revision, and tags observed when the snapshot was created. Later vault changes do not replace those rows. Clear search returns to the folder. Tag all… tags every result. Select rows to open the selection dock, which offers the same actions as the folder view. Live-folder keyboard shortcuts stay inactive while the snapshot is open.

Select a frozen row to inspect those original facts and the selected version. The same card loads current path, revision, tags, provenance, and audit status separately as live observations. A later rename, retag, or content replacement therefore does not rewrite the snapshot facts or substitute the new head for preview and download.

Use Filters to narrow by import, tag, media family, extension, modified month, size, and duplicates. Imports appear under their names. The list leaves out values no result has and groups that cannot narrow anything. Choosing a facet or changing the query creates a new snapshot. It never splices live results into the accepted one.

Photos' date rail counts each included photo asset once by capture day. It creates no snapshot and does not block Documents queries. Choose a year or month to scroll there, including dates beyond the loaded grid. See Browse photos and Count photos by capture day.

Snapshot handles last for one daemon lifetime, up to 15 minutes idle and 30 minutes total. Locking the browser session or stopping the daemon revokes them. If paging reports that the snapshot is gone, run the complete query again and use only the new snapshot and its cursors.

A frozen query with fixed counts, facets, and document versions

Read archived email

Choose Email in the content inspector for an EML document whose selected version has decoded metadata. Read decoded headers, inspect the original ordered headers, and switch between HTML and plain-text body alternatives. Missing or invalid sender dates are labeled as such. Pending decoding and incomplete MIME inventories do not appear as complete messages.

HTML opens in an isolated frame. Sender scripts, forms, links, styles, and remote resources are disabled before display. Blocked or ambiguous images have visible placeholders. Only verified PNG and JPEG MIME parts in the body's nearest related group can display inline. Press Escape inside the frame to return to its entry button. Long messages scroll, and Show quoted text expands quoted content without changing the original or the full-body PDF source.

The viewer accepts HTML bodies up to 16 MiB. Its additional display limits are 100,000 markup candidates before parsing, 100,000 HTML nodes, 128 nesting levels, and 32 MiB of serialized output. Literal < characters also count toward the conservative markup limit.

Inline images are limited to 128 occurrences, 5 MiB and 8 million pixels per image, 8,192 pixels per side, and 16 MiB and 32 million pixels in aggregate. PNG/JPEG limits and unsupported image formats produce visible placeholders. Body limits report an unavailable operation instead of silently truncating the body.

Inspect attachment documents opens the attachment navigation described below. Inline display and attachment document publication remain separate. Download downloads that EML version unchanged. The existing PDF action downloads retained PDFs or uses the configured renderer, including the complete chosen body and quoted text. New rendering requires the pinned local Chromium/font configuration. An unavailable renderer reports its reason.

Archived email with decoded headers and its HTML body

Raw headers from the same retained message

Follow email attachments

Choose Attachments in the content inspector to list the documents attached to the selected version under Attachments and the messages it is attached to under Attached to. A document with neither shows No published attachments. Repeated filenames and identical bytes stay separate rows. Show more requests another page, up to 1,000 rows per group. An unavailable read reports an error, and an empty list then does not mean the message has no attachments. Rows without a published document show their outcome and cannot be opened.

Choose a row to open that related version in its own context. Later content replacement does not substitute the live head or its media type for that related version. Preview, text, and download keep the usual byte verification and current-access checks. Missing or denied targets report an error without removing the row.

Follow nested messages with Back to previous document, or choose Return to frozen document to restore focus to the original result. These steps do not rerun the query, change its snapshot membership, or check attachment rows. Ordinary live selections offer the same return path to the selected document.

Exact attachment documents linked to the selected email

Export a verified ZIP

In Photos, choose Export all N photos in the ⋯ menu for the current scope or Export selection for checked photos. Prepare JPEG or PNG copies with size, image metadata and GPS choices. Each preparation allows up to 16 photos, each at most 100 million pixels. The verified ZIP retains original filenames, vault paths and source hashes. See Export photos.

Choose Export selection in the selection dock for the checked documents. The toolbar's Export → Download files… opens the whole frozen query when one is active, or the documents on the current page otherwise. Folders are not export members. Exports support up to 100,000 documents and 50 GiB of role payloads.

Choose Download ZIP. By default the ZIP holds the original files. Open Options to also include retained text, page images, or email body PDFs, to share exact duplicate outputs, or to split the ZIP into volumes. Each role is Required, If available, or Skip. A missing required role stops the export before anything is built. An If available role records its unavailable status in the bundle.

Download ZIP copies and checks every frozen page, stores the selected versions and available role files, builds the archive, and hands it to your browser. Later edits to the live query, tags, or current content do not substitute new members or bytes. When some outputs are unavailable, the drawer stops on the frozen plan and lists them; choose Download anyway to continue. The ZIP filename applies only to the browser download, and bundle paths stay deterministic. Originals are not redacted or sanitized by annotation overlays.

The drawer offers the download only after it receives that job's verified archive receipt. It displays the final size and SHA-256 so you can check the downloaded file independently, and keeps a Download ZIP button in case the browser blocked the first download. The browser saves the archive directly. Docbank does not load the entire ZIP into browser memory. Check your browser's download list for local completion. New export releases the finished job for both photos and documents, and asks you to wait while a download is still running.

Closing the drawer stops its progress reader, not the server job. Reopen Export in the same browser session to reconnect to that job, or choose Cancel export to cancel it. A disconnected stream is not a completed export. Expired authority requires choosing Download ZIP again. Reloading or ending the browser session does not preserve the drawer's job handle.

A verified bundle ready to download from the reviewed frozen selection

Review and redact a PDF

Open a live PDF and choose Redact PDF in the reader, or use the same action in the Documents inspector. Mark rectangles or whole pages on the same reader, inspect verified output and sanitized text, and download one ZIP. Closing the review returns to the reader's page and zoom; unsaved masks require confirmation before discard. The original is unchanged. Reviews are temporary and require the configured Linux amd64 page runtime. See the browser PDF review workflow for coverage gaps, private review files, expiry and download limits.

Stamp selected pages with Bates labels

Open Export → Bates-numbered PDFs… and select a sealed package. Bates labels are sequential identifiers stamped on PDF pages, often used when exchanging records for review.

  1. Create or select a label namespace, including its prefix and number width.
  2. Choose Preview Bates labels and inspect the proposed labels and pages.
  3. Reserve the displayed range, then choose Start Bates export.
  4. Wait for Stamped PDF ready, then choose Download PDF.

The output contains the package's selected pages. The stored page map keeps their original source-page numbers. A number that was reserved is never issued again, even if the export was abandoned.

A completed Bates export with its reserved labels and verified PDF

Find a retained PDF by number

Open Productions in the Documents sidebar, enter the exact Bates number, and choose Find. Matching is case-sensitive and preserves spaces and leading zeroes. A warning identifies leading or trailing spaces without removing them. Bates numbers are unique across the vault, so an exact number identifies at most one retained PDF and matching page. A match opens automatically, showing the page, publication time, PDF size, delivery kind, Bates range, SHA-256 and allocation reference. Choose Download PDF to recover it, or Clear to return to the full production list.

Recovery works after restart without the review files or a configured page renderer. Downloads check file integrity against the retained record; they do not repeat the mask and approved-text checks. See the numbered delivery verification scopes for a full check with your review files. The browser returns the PDF, not the render ZIP, text sidecar or review inputs.

Assign and remove tags

  1. Select a file or folder.
  2. Choose Add tag, or Manage when it already has tags.
  3. Type a tag name, then select a match or Create “name” to create and assign it. Remove an assigned tag with its × button.

The dialog opens with the cursor in the tag field and lists assigned tags separately from available tags. It loads the first 1,000 definitions in name order, then displays related names in groups. The picker uses the same color for a tag wherever it appears. Use the arrow keys to navigate its options, Enter to select, and Escape to close it.

Every change is bound to the node ID and revision shown in the dialog. After success, the browser updates the node revision and the tag's vault-wide assignment count. It also refreshes the tag catalog and permanent-audit status. If another person, agent, or CLI command changed the node first, the dialog keeps the failed decision visible and asks you to refresh rather than applying it to newer state.

Tag selected documents

Use the document checkboxes to select files on the displayed page, then choose Tag selected… in the selection dock. Pick a tag, choose Add or Remove, and press Apply to run one atomic, revision-checked operation on at most 1,000 documents. Type a new name and choose Create “name” to create the tag, then apply it to every selected document. Folders and undisplayed query results are not included in page selection.

If the result is uncertain, the dialog locks and the browser refreshes the view. Close the dialog and check the documents' tags before trying again. Applying the same addition or removal again is safe. A stale selection requires you to close the dialog and select again. After confirmation, the browser reloads current observations. The retained receipt may describe an earlier successful operation. See selected-set tagging for retention and backup behavior.

The tag dialog shows how many selected documents already have the chosen tag

Tag a frozen query

In a frozen snapshot, the results header shows the result count and Tag all…, and the selection dock shows Tag selected… for the checked rows on the visible page. Both open the same dialog: choose or create a tag, pick Add or Remove, and read the line under it, such as Adds “Reviewed” to 128 documents., before you apply. Tag changes can't be undone automatically. Remove affects the whole selection, including documents that already had the tag before an addition.

Tag selected… applies one atomic, revision-checked change to at most 1,000 documents and reports how many changed. If the result is uncertain, the dialog locks and the view refreshes. Close it and check the documents' tags before trying again. Applying the same addition or removal again is safe. A stale document means you close the dialog and select again.

Tag all… → Continue captures every result, up to 250,000 documents, and prepares revision-checked batches of at most 1,000. The captured population does not expand when an untagged query starts receiving the tag. Completed changes appear as overlays while the frozen rows and facet membership remain unchanged. New actions use confirmed tag revisions only when the receipts connect back to the frozen document revision. Other document changes still make it stale. Once prepared, an action keeps its original requests and expected revisions.

Apply runs the batches in the same dialog, with progress and Pause. Pressing Apply is the explicit confirmation. Review documents lists the vault and action identities and, in pages of 50, node IDs, content versions, hashes, sizes, and expected revisions. Cancel before anything was sent drops the prepared action.

The browser journals the action in IndexedDB for that origin. When a run stops early, for example after Finish later, a lost response, or a closed tab, the results header shows that it didn't finish, with Resume and Discard…. Another tab can resume it, but only the original operation identities, requests, and expected revisions are retried. If a response is lost after the daemon commits, Retry recovers the retained receipt without applying the change twice. Starting a new Tag all… replaces a finished run. While an earlier run is unfinished, Continue closes the dialog and shows that run's Resume and Discard… notice instead of capturing the results again.

A daemon restart creates a fresh browser origin and session, so choose Tag all… → Resume from file… and select the saved recovery file to continue. Import validates and stages the action but does not run it. The vault must match, and the new session still requires Apply. A stale document revision blocks the action. Docbank does not silently refresh its target or partially apply that batch.

Without a recovery file, run Tag all… again. Tags that already landed are left unchanged.

Recovery files contain instructions supplied by their creator. Checksums detect inconsistent files, but they do not authenticate the creator. Review the documents before applying an imported action. Receipts inside a file do not update the snapshot's observed revisions. Only receipts returned by the daemon during this session do that.

Recovery files contain private document identities, content-version identities, hashes, sizes, and revisions. They omit paths, names, query text, and browser credentials, but still need the same protection as vault metadata.

Discarding removes the browser journal and does not roll back changes that already committed. If a tag was deleted, the dialog still lets you save the recovery file or discard the run. Discard… in the results header also clears an unreadable journal for the current vault, so corrupt browser data cannot block later actions.

Manage tag definitions

Open the tag filter and choose Manage tags to create, rename, or delete the vault's shared tag definitions. Creating a tag allocates a new UUID. Renaming keeps that UUID while advancing the definition revision and every assigned node's metadata authority. A stale rename remains visible instead of overwriting a definition changed by another person or agent.

Deleting an unused tag happens in one click. Deleting a tag that is on files asks first and says how many files lose it. Deletion removes the definition and all of its assignments, but never deletes a document or its stored content. Reusing the same name later creates a different identity. After a rename or deletion, the browser reloads the active folder, search, or tag view so node revisions and assignment counts stay current.

Tags use a color derived from their ID. A rename keeps that color. Slash-separated names create display groups automatically: matter/acme/reviewed appears as reviewed under matter/acme. The full name remains available in the tooltip and to assistive technology. Groups do not create parent tags or change assignment rules.

The catalog shows the first 1,000 name-sorted definitions and discloses the complete count. Use docbank tag, the paginated HTTP API, or an embedded client for exhaustive definition management and assignments outside the displayed selection.

Export dated search counts

Open Searches → Run report to create a CSV of search counts for selected sources and date ranges. Review unresolved dates before downloading the CSV or evidence ZIP. Recent reports keeps the latest 100 requests and outcomes. Use as draft runs a request against current data without its old date choices. See Search exports for the full workflow, count definitions, and expiry limits.

Search counts for one selected version with CSV and frozen evidence downloads

Saved queries and highlights

Open the Searches page at /searches to run, rename, or delete a saved search. The page also holds Run report and Recent reports. Create searches from Advanced → Save…. Rename opens the name and description fields; Edit as JSON reveals the complete definition. Choose Save changes to update it or Save as new to keep a copy. The list is paginated in groups of 100. New highlight list creates reusable highlights.

Keep query draft keeps the complete query in the tab and URL fragment. The fragment contains query text and filters, so treat copied URLs as private. It does not contain browser-session credentials. Saving a definition does not replace the kept draft or change the URL. Cancel in advanced search, locking the tab, or an expired or rejected session clears the draft and its URL fragment. Drafts do not carry over when you open a new session with docbank web. Save a named definition before reloading or leaving the session if you need it later.

Searches retain the search definition. Run creates a new result set using current documents. Keeping or saving the draft alone does not run it or change the current live results. This browser run does not produce the durable saved-run receipt that the authenticated HTTP API exposes. Unknown fields are rejected, not silently dropped.

Highlight sets hold 1–64 unique literal terms, each up to 256 Unicode characters, with lowercase #rrggbb colors. They cannot run as queries. Choose one in a document's Text tab to apply its ordered colors without changing the saved definition. Neither result rows nor document bodies are stored in browser preferences.

Edits and deletions use the definition's inspected revision. A stale response remains visible without automatically retrying against newer state. Reload the definitions and reopen the item before deciding again. Deletion requires a separate confirmation naming the definition. It never deletes documents or discards the kept query draft.

A named query retaining its complete expression, filters, and sort

A reusable highlight set with literal terms and colors

Move a node to recoverable trash

Choose Move to trash on the selected live file or folder: the trash button for a file, or the folder's actions menu. The item leaves the list at once and a notice names it, with the item count for a folder. A folder and its live descendants move to trash together. Choose Undo in the notice to restore it.

The daemon applies the mutation only if that node revision is still current. If another person, agent, or CLI command changed the node first, the notice shows the error and the item stays listed. A successful receipt removes the selection and refreshes the current folder, search, or tag view in place.

This action is recoverable. It does not empty trash, garbage collect content, reclaim packed space, or erase permanent audited history.

Restore from recoverable trash

Choose the Trash page at /trash in the Documents sidebar to inspect the newest 1,000 independently restorable roots. Each entry shows its name, kind, node ID, revision, trash time, and logical size for files. A folder entry represents the complete subtree that left the live tree in that trash operation.

Choose Restore to return the item at once. Docbank returns the same node and its content to its original live parent when that directory still exists. If the parent is unavailable, restore falls back to the vault root. If the chosen name is already occupied, Docbank adds its normal collision suffix. A notice shows the actual canonical folder, names a missing original folder when this browser trashed the item, and counts the files of a photo group. Choose Show to open the folder.

If the revision changed, the notice shows the error and the item stays on the page. After a successful restore, the browser removes the item from the trash page and reloads the current view.

The browser cannot empty trash. Permanent tree-metadata deletion, subsequent garbage collection, and packed-space reclamation remain CLI or master-authenticated API operations.

Upload verified documents

  1. Browse to the destination folder.
  2. Choose Add → Upload files or Add → Upload folder.
  3. Select files or a folder from this device, then choose Upload in the drawer.
  4. Check each file's result in the upload drawer.

Upload is available only in a live folder, not in search or tag results. The drawer shows the destination path once in its heading.

Docbank makes two bounded-memory passes over each selected file. The first pass computes the browser's declared SHA-256 while showing hashing progress. The second streams the bytes with visible progress over a dedicated upload channel. Before the upload button becomes available, the daemon proves that channel with a random secret issued through the ownership-pinned CLI handoff. File bytes never use an ordinary reconnectable browser HTTP request.

The channel is bound to one browser session and one daemon lifetime. If it breaks, the page permanently disables upload rather than reconnecting to whatever process now owns the loopback port. Run docbank web again to obtain a newly proved channel. The daemon independently computes the hash and size and grants node/blob authority only when both match. The browser compares that receipt again before reporting Added or Already present, then refreshes the destination by ID.

Files are independent queue entries: one rejection does not make another success ambiguous, and failed entries can be retried. Cancellation can race with a daemon commit whose receipt did not reach the browser, so the drawer labels that item Unconfirmed, refreshes the destination, and directs the operator to retry. Uploads are idempotent, so the retry converges on the stored result. Name/content collisions keep the ordinary ingest suffix behavior rather than overwriting a document. Selecting a local file never changes or removes the source.

The folder picker keeps each file's relative path. The browser creates the matching subfolders under the destination (reusing any that already exist) before it uploads the files inside them. Server-filesystem ingest, watched-inbox configuration, and replacing an existing document remain CLI or authenticated API workflows.

Import a mailbox

  1. Browse to the destination folder and choose Add → Import mailbox.
  2. Choose an MBOX file or Google Takeout ZIP, then Upload and preview.
  3. Review the detected mailbox entries and choose Import messages.
  4. Follow the imported and rejected counts until the job finishes.

Each message occurrence becomes its own email document. Repeated messages remain separate occurrences, and attachments become child documents. An interrupted import resumes after its last saved message. See mailbox import for CLI status and receipts.

A completed Takeout import with three separate message occurrences

Import load files

  1. Browse to the destination folder.
  2. Choose Add → Import load files.
  3. Choose a ZIP that holds the DAT or CSV load file, an optional OPT or LFP page map, and the referenced original documents, images, and supplied text.
  4. Pick the metadata profile, page-map profile, and source encoding, then choose Upload and preview.
  5. Resolve blocking diagnostics, name the package and sending party, then choose Import package.

The ZIP uses the authenticated upload channel and is stored as a sealed container before preflight reads it. The drawer shows imported, gap, and total counts, plus the operation ID, as the job runs.

Keep supported records and report gaps allows an import to keep records when referenced files become unavailable after preflight. Blocking preflight diagnostics must still be resolved. Without this option, the worker verifies every declared file before creating document entries. Index package-supplied text makes the sender's text files searchable.

Browser source metadata omits fields marked sensitive by the retained mapping; compact Map-projection rows include latitude and longitude when map tiles are configured. Member lists and raw timeline inputs require the daemon API key.

Closing the drawer stops its progress updates. Lock web session revokes the session and cancels unfinished imports it owns. Cancellation keeps committed records and removes unfinished document entries as described in the package command reference.

A load-file ZIP preview before importing its documents and supplied text

Process a version and read its text

Processing extracts a document's text and stores it. With a configured embedding service, it also prepares the document for search by meaning. Configure a processing profile first. A new vault has none.

  1. Select a document and choose Extract text beside its name. This action appears when a configured profile includes text extraction. The processing dialog checks which profiles accept the file. Profiles that only prepare search vectors show Index for search instead.
  2. Read where the file goes. A local profile says it runs on this computer; a network or hosted profile names the service and its address, and says whether the file name goes too. Details lists the exact plan, service, deployment, and fingerprints.
  3. The first time you use a profile, tick Allow "profile" to process documents in this vault. That permission covers this operator and profile configuration for every document, not just this file, until it is revoked or expires. Later runs need only the button.
  4. When the run finishes, choose Read text to read the extracted text Docbank stored for that version.

To withdraw permission, open Details and choose Revoke all processing permissions. It applies to every profile.

The processing plan discloses local extraction and a configured network embedding service

A finished text extraction with its text ready to read

Verified retained Markdown for the selected source version

Opening the plan or reading coverage does not start processing. A restored vault needs fresh consent before network processing. The processing guide explains consent and provider requirements. To inspect results without the browser, see CLI coverage and text windows.

When a processing profile stores search vectors, each file row shows a Find similar action. Choose it to compare its stored embeddings with the file versions loaded in the current view. This calls no provider. If embeddings are missing, build them with a processing run, which has its own review and consent. See Find similar documents.

Related documents found from stored embeddings in the loaded view

Download verified content

The inspector puts the preview above details, tags, and History. Drag the inspector's left edge, or focus it and press the arrow keys, to resize it. Double-click the edge to reset. Your browser remembers the width. Click a path, version ID, SHA-256, or ID to copy its full value. Download saves the selected version. Extract text appears beside it when a configured profile includes text extraction; otherwise, configured profiles offer Index for search. The processing dialog checks which profiles accept the file. Files have a direct Move to trash button. Folders have an Open button, and unprotected folders have Move to trash in ⋯.

Protected items show a Protected chip that opens Audit history. For an unprotected folder, choose Protect permanently… in ⋯ to copy its command-line preview command. The command previews retained history and vault-wide metadata before you explicitly enable irreversible protection.

Double-click a PDF or Markdown file, press Enter on its row, or choose Open in the inspector to read it in a large viewer. Escape or Close returns to the selected row. This also works in search results and preserves the version selected by a frozen snapshot.

PDFs open in EmbedPDF, with search, zoom, page navigation, thumbnails, text selection and printing. Download saves the selected original. Redact PDF switches live PDFs to review in the same workspace; frozen search results stay read-only. Its rendering engine and fonts ship with Docbank; viewing a PDF needs no optional server runtime or external asset download. PDF previews are limited to 64 MiB. Markdown displays headings, lists, tables and code blocks; embedded HTML is sanitized and form controls are removed before display. Links open in a separate tab. Other eligible UTF-8 text appears as plain text. Text previews are limited to 16 MiB.

The inspector also previews bounded PNG, JPEG and WebP images. Animated WebP images may animate in the browser. Docbank verifies the selected version UUID, size, media type and SHA-256 before passing content to a viewer. Unsupported or oversized files remain available through Download. Reclicking a selected item retains its preview and metadata; browsing keeps the preview area a consistent height while the next file loads.

Selecting a PDF in the inspector reads its retained page images without loading the whole PDF or starting a browser PDF engine. Open the reader to load the verified original. PNGs with retained page geometry also use the retained-page preview. Choose Render page when an image is missing; this needs the optional local page runtime. Other PNGs use the verified original-image preview, including when that runtime is unavailable.

Choose Text to read the verified text rendition for the selected version and processing profile. The browser binds the source, profile, generation, attachment, build, artifact, length, and digest before decoding the text. A check icon shows verification. Find searches within the displayed text and shows its matches in place of query and saved highlights; match counts and arrows appear when there is a query and matches, respectively. The Text tab remains available for plain-text files, even when Preview shows the same text. Processing failures, verified-empty text, and unavailable historical renditions remain distinct. Eligible original UTF-8 text is a fallback for that same version when no readable rendition is available. Without a processing profile, Set up opens the processing guide. With a profile, Process… opens the processing plan.

Query text operands are highlighted from the resolved query, not by splitting the expression in the browser. Negated and structured operands are excluded. The browser refuses query highlights if a saved definition revision no longer matches the accepted snapshot. Find and saved highlight sets share bounded, non-overlapping marks and previous/next match controls. All content stays inert.

Previous and next document controls move through the accepted frozen snapshot, including across its existing page cursors. They stop at the snapshot boundary and do not replace an expired snapshot with live results. Duplicates lists at most 16 live documents with the selected content hash. Opening one creates a separately labeled live context. Return restores the frozen selection.

Choose Download on a file to retrieve its selected version. A live row selects its current head. A frozen query row keeps the version captured by that snapshot even when the live head has since changed.

Docbank first copies the object from loose or packed storage into owner-private daemon staging while the browser shows verified byte progress. The selected node revision, version UUID, SHA-256 identity, and size must still agree before that work starts. The current node revision authorizes access to a retained historical version but never changes which bytes were selected. A concurrent replacement, move, or trash operation therefore asks you to refresh instead of silently downloading a different document.

After the complete content passes verification, Docbank issues a one-use download ticket. The browser starts its save only after that check succeeds. Cancellation or a verification failure removes a pending private staging file and publishes nothing to the browser. Changing the selected source or locking the session also cancels pending previews and revokes published image URLs. A ticket identifies only its prepared file, expires after two minutes, and cannot call any other Docbank route. It is not the vault API key or the browser session.

Preparation streams on the daemon and does not buffer the object in browser memory. It temporarily needs local free space equal to the document's logical size. Docbank reports that verified bytes were handed to the browser, not that the browser or operating system completed its final save. Use docbank document get when automation needs a durable receipt after private staging, file sync, atomic publication, and parent-directory sync.

Inspect immutable versions

Choose All versions in a file's History section to inspect retained versions without losing the current folder, search results, or selected document. The newest version appears first. Each entry says whether the content was created, replaced, or restored from a prior version, with its recorded time and size, and has a Download button. Details shows the version ID, SHA-256, media type, and the version it was restored from.

Open a link with #node=<node ID>&version=<version UUID> to inspect that exact version, including versions older than the newest 1,000 listed in the drawer. For drawers opened from these links, selecting a version updates the address, and browser Back and Forward restore the selected version. Missing, purged, or mismatched links show an unavailable message.

Selecting an entry shows its complete record: full version UUID, SHA-256 content identity, exact byte count, media type, canonical timestamp, introducing operation UUID, and the source version for a revert. The current head is marked. A revert is a new immutable version and does not erase or relabel the earlier one.

Trashed files carry a Trash label. Restore the file before downloading a version.

Choose Download in the complete-version panel to retrieve that version through the same private staging, progress, terminal verification, and one-use handoff as the current-content action. The version UUID, owning node, SHA-256 identity, size, and current node revision must all agree before preparation begins. Historical versions do not keep a filename timeline, so the browser uses the live document's current name. Version comparison remains a CLI or authenticated API workflow.

The drawer reads at most the newest 1,000 versions and says when older history exists. Use docbank document versions list, its pagination flags, or the authenticated HTTP API for exhaustive automation. This view does not compare bytes, revert, or prune history.

Understand where a document came from

The History section shows the five most recent versions and imports together, newest first. Choose All imports to see the document's complete import history. Each entry shows how the document arrived, its copyable source path or filename, and the import date. Corrected entries show Superseded.

Open Import details for the import ID, record hash, document ID, source reference, original modification time when supplied, and any replaced record. Click a value to copy it. The drawer uses the document state returned with the history, so a concurrent move displays its current path and a trashed document is labeled as trashed.

An original reference records the source at import time. It does not guarantee that the source still exists. It may be a portable relative path from a watched inbox, a local path recorded by an ordinary ingest, or an opaque reference supplied by an embedded application. The browser does not open, validate, change, or retain that external source.

The drawer reads at most the newest 1,000 facts and says when older provenance exists. Use docbank document provenance, its pagination flags, or the authenticated HTTP API for exhaustive automation. Provenance correction and external-content pinning are not browser workflows.

Browse tags and search text

Enter a word or phrase in the search box and press Enter. Results can match a live document name or verified extracted text. The results heading shows the count, each row shows the name with its folder beneath, and the Match column identifies which one. Previous and Next in the inspector step through the results. Name matches keep their API relevance ranking and appear before content-only matches until you choose a column sort.

Use the tag selector in the browser toolbar without a text query to browse up to 1,000 live items carrying that tag. The result is read in one metadata snapshot, so its node state and complete virtual paths cannot mix concurrent moves, trash operations, or content updates across pages. Trashed assignments are omitted from this live view and disclosed in its count. Use docbank tag nodes for the exhaustive paginated list of live and trashed assignments.

With text in the search box, the same selector requires that tag in addition to the name or content match. The selector displays tag names, but both workflows are bound to the tag's UUID, so a later rename does not change which definition was selected. Changing the selector reruns the current browse or search.

The search box searches words in names and document text. When a processing profile has a query embedding binding, it also searches by meaning automatically. The menu selects Plain words or Query syntax, matching Photos. Query syntax accepts expressions such as name:invoice AND tag:urgent and runs through the complete query endpoint.

Rerank results is opt-in when a processing profile supports it. Base rows appear before reranking settles. Degraded reranking keeps those rows and shows its cause. Other processing failures fall back to names and text with a visible explanation.

The Docbank web application showing extracted-text search results in a synthetic vault.

Search results show each document's folder under its name and offer the same authority card as ordinary folder browsing.

Clear the search box to return to the selected tag's assignment view, or to the current directory when All tags is selected. The selector loads at most the first 1,000 name-sorted tag definitions and discloses when more exist. Use docbank tag list for the tag catalog. Use docbank search for directory, media-type, or modification-time filters, structured JSON, or another result limit. See Searching for limits and truncation.

Read permanent audited history

The details panel shows a green Protected badge beside the name when the selected node belongs to a permanent audit scope. Ordinary deletion, version pruning, garbage collection, and repacking cannot erase that node's history.

Click Protected to open the audit timeline without losing the current folder, search results, or selection. The timeline is newest first and explains the primary change for each event: live and retained-trash paths, content-version transitions, tag definitions and assignments, or provenance. Select an event to inspect its complete event, operation, scope, and node identities; canonical timestamp and origin; before/after revision, path, and version state; and typed tag, provenance, or photo payload. Photo states show authored values and confirmed field names before and after edits and undo, including confirmation of an empty caption.

The first page contains at most 50 events. Load older events follows the API's append-stable cursor, so new activity cannot shift or duplicate the history already being inspected. Protection status remains authoritative even when a page contains no events. The web application does not infer protection from an empty or non-empty timeline.

Verify permanent audit evidence

Choose System → Protected folders to run the vault-wide permanent audit verifier. The verifier does more than read stored status. Docbank replays canonical audit history against the current node, version, membership, topology, tag, and provenance projections, then reads and recomputes SHA-256 for every unique blob retained by protected history.

A successful result reports the protected and verified file totals and the protected size. Details holds the vault and allocation-lineage identities, the operation high-water mark, the allocation entry count and head, and each scope's terminal entry count and chain head. Copy buttons preserve the complete identities even when they wrap. A vault with nothing protected shows the docbank audit enable command instead of a failure. Metadata, missing-content, corruption, and unreadable-content problems stay visible with the affected hash.

This drawer runs only a fresh proof of current authority. It does not accept a previous evidence bundle, enroll a scope, or change protected state. Record docbank audit verify --json outside the vault and later use docbank audit verify --expected when that external copy must act as a rollback trust anchor. Closing the drawer cancels its active request. Maintenance contention or interruption remains visible, and you can retry.

Inspect background work

Choose System → Running now to see work in progress, paused workers, and failures. Active work appears first, with counts when known and its pause, resume, worker count, and cancellation controls beside it. When workers are waiting for work, a quiet Nothing running line sits above the worker list. The list below shows every worker and its status, and offers pause controls for idle workers. Queued operations stay in that list until they start.

A worker waiting for work shows Idle, and the header counts only work in progress. Watched inboxes show their configured tags. A missing tag appears beside the configured tags, separately from any worker error.

Pause or resume photo imports, visual previews, and the storage lanes place, evacuate, repair, and salvage. Pause takes effect at the next safe item boundary. Visual preview concurrency accepts 1 through 4. A paused lane keeps its Resume control after its last operation disappears. The drawer reads lane settings with the job list and uses PUT to change them. The drawer refreshes every two seconds while open. If another client changes controls first, review the refreshed settings and try again. Controls survive daemon restarts. The list includes the newest 1,000 storage operations. Finished storage operations clear within an hour after their retention ends unless they still own pending cleanup. Worker records disappear when the daemon restarts. Browser errors omit backend paths; use the CLI to inspect details.

Inspect physical storage

Choose System → Disk use to see how the current vault occupies managed blob storage. Tiles show the total on disk and the number of stored files. When packs hold dead payload, Reclaimable shows how much and what share of the pack payload it is. Locations has one row per storage location with its kind (filesystem or S3), stored size, object count, how many objects have no other copy, and its observed online, unavailable, fenced, or unbound state. When an unhealthy location is the only authority for an object, the drawer says how many objects currently have no readable alternative. It never exposes binding paths, endpoints, buckets, credential profiles, or ownership epochs.

Storage separates four related quantities:

  • Loose files is the physical inventory of individual raw or zstd files. It can include untracked files or redundant loose copies of packed objects, so it is not an authority count.
  • Packed is authoritative content stored in immutable pack files. The view shows both its logical raw size and its stored size.
  • Pack files is the complete stored payload of every pack, including live and logically dead entries.
  • Waiting to compact is logically dead payload that still occupies those pack files.

Waiting-to-compact bytes have not been reclaimed. This distinction matters when a GC report has removed unreachable catalog mappings but the vault's disk usage has not fallen by the same amount. The percentage beside Reclaimable shows how much of the current pack payload is dead, not a compression ratio or a promise that every pack is immediately eligible for compaction.

This drawer is read-only and refreshes from the daemon's current catalog. It cannot pack, garbage-collect, or repack content. Use docbank storage status for structured or scripted inspection, see Multi-store storage for repair and placement, and run docbank storage repack when you intend to rewrite eligible sparse packs and retire their old files.

Inspect backup snapshots

Choose System → Backups to inspect the recovery points in the repository selected by [backup] repo in config.toml. The drawer shows the repository path and repository ID so you can tell which repository you are inspecting. Snapshots appear newest first with their tag, creation time, full or incremental relationship, file count, logical content bytes, and the bytes newly added by that capture. A copy button on each row copies the snapshot ID.

An initialized repository with no snapshots is different from an unconfigured repository, and the browser reports those states separately, each with the command that fixes it. The browser does not accept an arbitrary server path, initialize a repository, create a snapshot, verify content, restore a vault, or delete retention history.

Seeing a manifest in this list is not proof that its referenced bytes are still readable. Run docbank backup verify to check repository integrity, and periodically restore into a separate vault to rehearse the complete recovery path.

Browse import history

Choose Imports in the sidebar to open the page at /imports. Each row shows its label or source folder name, import time, document count, and size. Select an import to see its source paths, each with a copy button, browse its current live members, and open a document by its node identity. Counts are current membership, rather than a historical import total.

Labels belong to the import group, not its documents. Rename or clear a label under its inspected revision. If another client changes it first, the page keeps your draft and reports the conflict. Reload the label before deciding whether to save again. Label changes are unavailable once permanent audit authority has been enabled.

Collection and member lists show at most 100 entries each. Empty groups, failed reads, and truncated lists are reported separately. Use the paginated HTTP API to browse beyond that limit. This is direct member browsing, not a collection-filtered text search.

Choose Quality for document distributions, duplicate content counts, zero-byte files, media/extension mismatches, and text coverage. With multiple processing profiles, choose one before reading coverage. Without a configured profile, coverage is unavailable. That does not mean zero failures or complete processing. A configured policy does not mean an extraction adapter is running.

Coverage distinguishes complete, partial, failed, unprocessed, and activated output with no text. Retained searchable output takes precedence over a failed retry. The existing worker rejects blank provider output as a failed attempt. Quality reads are bounded to 250,000 members, a 64 MiB census, and five seconds. An exceeded limit returns an error instead of a partial summary.

Select extensions, media types, or media families and choose New query to open a collection-scoped draft. Multiple selected values use OR. This does not change live results or execute a search. Concentrations describe common values, not document defects.

An import collection with a saved label and its document members

An import collection showing file types and processing coverage

Sign in on a server or container

When an operator configures server.api_key, the Open your Docbank page accepts that key and exchanges it for the same limited browser credentials issued by docbank web. The input clears after submission. Docbank never stores the key in browser storage, returns it in a response, or uses it for subsequent browser requests. Each tab keeps its scoped credentials in memory. Sign in again after a reload or daemon restart. If a later request finds the session expired or revoked, Docbank clears that tab's selection and open dialogs and returns it to sign-in.

The Open your Docbank page asking for an API key

  1. Set a fixed API port and a strong API key in $DOCBANK_HOME/config.toml. With public_origin, the key must be at least 32 characters; generate one with openssl rand -hex 32. Protect this file as a credential. Configure the exact public origin:
[server]
api_port = 7777
api_key = "output-of-openssl-rand-hex-32"
idle_timeout = "0"

[web]
public_origin = "https://archive.example.test"
allowed_hosts = []
session_lifetime = "24h"
  1. Keep the daemon on loopback. Terminate HTTPS at a protected reverse proxy that preserves the public Host and forwards WebSocket upgrades to the API port. Docbank takes authority from configuration, never Forwarded or X-Forwarded-* headers. Every request, including API-key requests, health checks, and static pages, must use the public authority, a concrete loopback backend authority, or an exact entry in [web] allowed_hosts. Docbank also admits these authorities through the server Host allowlist. Entries are host[:port] values; schemes and wildcards are rejected.
  2. Open the public origin and enter the API key. Use HTTPS for remote access. An HTTP origin is accepted on loopback; non-loopback HTTP requires trust_private_network = true, which explicitly trusts the private network and its proxy with the key and browser credentials.

Without public_origin, configured-key login is available at the daemon's fresh loopback web origin, and key login adds no Host restriction beyond the server's allowed_hosts. The usual docbank web fragment handoff also works with ephemeral API keys and keeps its existing upload proof and permissions.

Serve Docbank under a path

Set base_path when other applications share the public origin:

[web]
public_origin = "https://library.example.test"
base_path = "/archive"

Restart the daemon, then open https://library.example.test/archive/. Keep public_origin free of paths. The proxy must preserve /archive in upstream requests, preserve the public Host, and forward WebSocket upgrades. For example, inside an existing Caddy HTTPS site:

@docbank path /archive /archive/*
handle @docbank {
    reverse_proxy 127.0.0.1:7777
}

Use handle, which preserves the path. handle_path would strip the prefix. The daemon redirects /archive to /archive/; nested pages, browser API requests, uploads, downloads, and assets stay beneath that prefix. Existing CLI API endpoints remain available at the backend root, and docbank web includes the path in its loopback launch URL. The proxy only needs to expose the configured prefix.

Login cookies are scoped to the configured path and have distinct names for each origin and path. This lets separate Docbank instances share an origin. Photo-import resume cookies use the mounted checkpoint endpoint, and pending Bates exports in the same browser tab are kept separately for each mount. As with any applications on the same origin, their scripts share browser origin authority; a path is not a browser security boundary.

Browser sessions

Key login sets a host-only, HttpOnly, SameSite=Strict instance cookie; HTTPS also sets Secure. Each tab receives its own token and upload proof. The cookie alone cannot read the vault. Requests need the tab token and cookie; mutations also need the exact public Origin. Cross-site requests are rejected. Credential requests reject redirects.

session_lifetime accepts 1 minute through 90 days. Omitting it or setting "0" uses 24 hours. Expiry is absolute. Failed sign-ins are not locked out; the key length requirement keeps guessing impractical. A daemon allows up to 512 active key-login sessions. Sessions are held only in memory and are excluded from vault backups.

Choose Sign out to revoke the current tab. Login and logout do not rotate or delete the shared cookie or invalidate other tabs. An operator can list and revoke individual key-login sessions without seeing their credentials:

docbank web sessions
docbank web sessions --json
docbank web sessions revoke SESSION_ID

Session IDs are non-secret digests. These CLI commands use the daemon's ownership-proven master connection; browser credentials cannot manage other sessions. Expiry, revocation and daemon shutdown cancel the session's uploads, exports and prepared reads through the existing browser resource lifecycle.

Browser authentication

When Docbank opens the browser, it writes a small launch page beside the owner-private daemon runtime record and passes only that credential-free local file path to the operating system. Before doing so, the ownership-pinned CLI asks the daemon to exchange its master API authority for a random, daemon-lifetime browser session. The master key stays on that pinned connection and never enters browser storage, a URL, or a child-process argument.

The daemon serves that session from a second listener with a cryptographically random docbank-<random>.docbank.localhost hostname and a newly selected loopback port. This browser origin is independent of the configured API port and unique to one daemon lifetime. A process that later captures either port therefore cannot leave a service worker or cached script waiting for a future browser session.

Photo import recovery uses a persistent HttpOnly, SameSite=Strict cookie on docbank.localhost to locate the saved selection in the same vault after a restart. The cookie grants no API access. Each checkpoint request still needs the current browser session, matching Host and Origin. Other browser data stays with its per-run origin.

The launch page carries the scoped session and its random upload-proof secret in a URL fragment. Browsers do not include fragments in the initial HTTP request. The application removes them from the address bar and holds them only in page memory. Ordinary requests send X-Docbank-Web-Session. The daemon accepts that credential for the following operations:

Permission Limit
Read the tree, nodes, search results, tags, versions, and provenance Uses the same result limits and document IDs as the ordinary API.
Read audit status and history Does not enroll or change a permanent scope.
Verify permanent audit history Does not run general metadata/content verification or backup verification.
Read background jobs and physical storage status Can pause or resume supported lanes, set visual preview concurrency, and cancel individual storage operations.
List backup snapshots Uses only the repository already configured for this daemon.
List trash Returns a bounded list of restorable roots.
Prepare, preview, or cancel an exact-version download Writes only a private temporary file, enforces preview MIME and size limits, and issues one expiring ticket for that file.
Plan, preview, run, cancel, and download exports Owns exact sources, frozen plans, jobs, and verified tickets within this browser session. Download naming cannot select a server destination.
Read verified text and exact-content duplicate context Revalidates the selected source and rendition identities. Duplicate context is bounded to 16 live-current references.
Read retained email metadata and MIME parts Source version, decoder generation, part, role, byte size, and SHA-256 are checked before display. Browser reads cannot start decoding.
Read page geometry and images, request or cancel page rendering Binds the source revision, version, page, frame, and renderer recipe. The inspector requests one page at a time.
Move to trash or restore Requires the selected node ID and its current revision.
Create a folder (POST /api/v1/nodes with no query string) Makes directories only, for folder uploads. File creation stays on the upload channel.
Add or remove a tag assignment Requires the selected node ID and its current revision.
Create, rename, or delete a tag definition Rename and delete require the inspected tag revision. Deletion reports the removed assignment count.
Read and manage saved query or highlight definitions Edit and delete require the saved definition's revision. Permanent audit history blocks these writes.
Read collections and their members Returns bounded lists of live import membership.
List, create, rename, star, set covers, duplicate, or delete photo albums; add or remove members Existing-album writes require the inspected revision. Membership requests accept up to 1,000 IDs per request or a complete live query without a total member cap. Removing members or deleting albums keeps the photos. See Albums.
Prepare, edit, or undo photo metadata (POST /api/v1/photos/authored/prepare, /api/v1/photos/authored/edit, /api/v1/photos/authored/{receipt_id}/undo) Edits bind up to 1,000 live files to their asset, file and node revisions. Each edit target may also supply content_version_id, which fences the asset's displayed version whichever member is edited. A replaced version or a different display file returns HTTP 412 stale_revision. Fields and keywords share one transaction and receipt. Undo requires matching revisions; hidden photos also require an unlocked Hidden view.
Search, rename, or merge people; read and draw photo faces; name or clear a face Rename and merge require the person revisions and apply only to people the browser can list. See People in Photos. Naming a face can create its person. Standalone person creation requires the API key. Face edits require the face revision. Drawing creates a box. The browser can mark an existing person or create a person by name in a whole photo and clear whole-photo marks. Person links open a filtered Library or, from unlocked Hidden, a filtered Hidden view. Hidden face reads and edits require an unlock.
Set or clear a collection label Requires the inspected collection-label revision.
Create and page a frozen query snapshot Uses complete validated query intent. Handles remain bound to this session and daemon lifetime.
Apply a tag to an exact frozen population Requires captured node revisions and confirmation. Retries preserve the original operation identities.

Use the bookmark button to manage saved queries and highlights or Advanced to validate and run a complete query.

Upload uses a separate WebSocket: a connection that never reconnects during the session. The browser must prove it holds the upload secret before sending any file bytes. Upload can create only file nodes beneath the live directory selected in the browser. Folder uploads also create the subfolders they need through POST /api/v1/nodes, which makes directories only. The daemon independently checks the caller-declared hash and size, as it does for other remote writers.

These permissions do not allow emptying trash, document mutations other than the revision-checked tag workflows above, audit enrollment, backup creation, restore, maintenance, configuration changes, or access to general API endpoints. A browser session never receives the master API key. See the HTTP API for the route and credential contracts.

The lock button revokes the session in daemon memory and clears the page. Every remaining fragment browser session and its dedicated browser origin disappear when that daemon stops. Run docbank web again to create a fresh origin and session against the ownership-proven daemon.

Closing the browser tab does not stop the daemon or revoke other sessions. Use the lock button when the current tab should lose access immediately, and use docbank daemon stop to stop the daemon. All browser sessions end when the daemon stops.

The launch file remains beneath $DOCBANK_HOME/web-launch/ with the same owner-only Unix permissions or Windows DACL as the runtime record. It is runtime state, excluded from snapshots, replaced by the next launch, and removed when the daemon stops.

Use docbank web --no-browser only when another local program must open the URL. That output contains the live browser credentials. Do not put it in shell history, logs, screenshots, issue trackers, or chat.

Current limits

Folder, tag, search, and recoverable-trash views show at most 1,000 rows and say when more exist. Use CLI or HTTP pagination to list complete folders, tags, and trash. Search has no continuation cursor. Narrow incomplete results as described in Searching. It uses the same name and verified-text matching rules as docbank search. Results initially preserve the API's relevance ranking. Choosing Document, Size, or Modified changes to that column order. Document compares the complete paths shown in search results rather than only their basenames.

Refreshing a folder resolves its node ID, current canonical path, and children in one metadata snapshot, so a concurrent CLI or agent move cannot leave the browser constructing child paths beneath an obsolete name.

Choose Add → Upload folder to upload a folder and its subfolders while preserving their structure.

The current web application does not compare versions, edit, revert, prune, move live nodes between folders, empty trash, enroll audit scopes, or run maintenance, backup creation, backup verification, general metadata/content verification, or restore operations. Frozen-query tag actions are capped at 250,000 documents. Use an authenticated API client for larger or different bulk workflows.

Original-file previews support UTF-8 or ASCII text up to 16 MiB and PNG, JPEG, and WebP images up to 32 MiB. Animated WebP images may animate in the browser. PDF and PNG page rendering accepts sources up to 64 MiB within the page runtime limits. Empty or larger sources offer verified download without a page preview. Other document, image, audio, video, and archive formats use verified download.

If a page reports that its browser session or upload channel expired, ended, or was rejected, run docbank web again. Neither credential nor the upload channel survives daemon restart, and the previous random .localhost origin is not reused.