Apocalypse Mode is WebBrain's optional offline knowledge layer. It reads Wikipedia archives in the openZIM format used by Kiwix. It does not make the configured LLM available offline: generating an answer still requires a local model or a reachable model provider.
The feature is disabled by default. Enabling the packaged Wikipedia skill does not enable Apocalypse Mode, query the Kiwix catalog, or store article text. Open the ☢ Apocalypse Mode link beside Support in the Settings header to opt in.
On supported Chromium browsers, local LFM2.5-VL vision is an optional fallback. Apocalypse Mode never enables or downloads it. The dedicated Use local fallback control in Settings → Multimodal → Vision checks WebGPU support, records explicit consent, and starts caching the approximately 770 MB model from Hugging Face. The download continues in the background, but screenshot operations report its status and never wait for it. Wikipedia archives still require their own confirmation. The local text model defaults to LFM2.5 2.6B (about 1.55 GB), and that download starts automatically when Apocalypse Mode is enabled. Bonsai 27B is an opt-in second preset (about 3.8 GB, 16 GB+ RAM/VRAM recommended) and is never auto-downloaded. Disabling local vision preserves its cache and any configured remote vision provider.
Archive language is selected independently from WebBrain's interface language.
The management page reads Kiwix's current OPDS catalog and offers a language plus
one of two editions: Text only, which resolves to the complete nopic
edition, and With images, which resolves to the complete maxi edition.
Both are complete-encyclopedia (_all_) archives. selectWikipediaArchiveVariant
accepts nothing else, so curated subsets and compact mini editions are never
installed from the catalog even though classifyArchiveTier can still label them.
The required first download is separate and fixed: Simple English, complete, and
nopic. See isBasicWikipediaArchive. It is text-only, not an images edition.
Because both selectable editions are complete _all_ archives, both carry the
Kiwix full-text index: nopic omits images, not the index.
Before an install, WebBrain resolves the archive's Metalink and shows its exact byte size,
archive date, catalog publisher/source and license notice, integrity-piece
count, and the browser's reported free extension storage. The archive is
downloaded only after that confirmation. Existing .zim files are validated
and their embedded date/language/source/license metadata is shown before import.
When the current catalog or archive omits a license field, WebBrain says that it
was not declared instead of presenting the general Wikipedia notice as an exact
publisher declaration.
Kiwix publishes very different archive sizes. A complete language edition can require tens or hundreds of GiB, especially when images are included. Catalog values can change; the confirmation dialog is authoritative for the selected current entry.
- IndexedDB (
webbrain_apocalypse_mode) contains the opt-in setting, archive metadata, byte cursor, generation, retry state, and storage reference. - Archive bodies are kept in the extension's Origin Private File System (OPFS), not as multi-gigabyte IndexedDB values. Chromium browsers exposing the File System Access API can instead use a user-selected file target. Removing that archive from WebBrain retains the user-owned file; Firefox uses OPFS.
- Downloads use Metalink piece boundaries and verify each piece before writing it. The persisted cursor makes background-worker restarts resumable.
- A lease prevents two extension contexts from claiming the same piece.
- Pause and disable increment a generation so stale work cannot commit. Deletion removes metadata before bytes, so an in-flight request cannot resurrect the archive.
- Transient failures use bounded exponential backoff. Integrity failures never write the rejected piece and eventually require a manual retry.
- Catalog downloads continue in the background after the management page is closed. Reopen Apocalypse Mode to inspect progress or pause the download.
- The Chromium-only local vision model uses the browser's Transformers cache. After its explicitly requested download completes, its GPU allocations are released until WebBrain actually needs local screenshot analysis.
- An installed archive that later becomes unreadable because of corruption, eviction, or a revoked file grant moves from ready to an actionable error; WebBrain reports the read failure instead of misreporting an empty search.
- Update checks can be manual or automatic. The automatic policy performs a daily catalog network check, but installing a discovered replacement still requires confirmation. A newer archive never silently overwrites an older one; delete the older archive after verifying the replacement.
Imported archives are structurally checked and extension free space is reviewed before they are copied to OPFS. Closing the management page interrupts an active user-file import because browsers do not provide a durable file grant consistently. A stale import is marked failed, and its partial bytes are retained so the failure remains visible and recoverable. Delete the failed entry to reclaim that storage, then choose the file again to restart the import. Partial import bytes are removed automatically on explicit cancellation, quota exhaustion, or another write failure.
If a live Wikipedia tool request fails, the exact built-in Wikipedia skill can
search installed archives by canonical title and title prefix through the ZIM
URL index. Retrieval is behind a provider seam; createKiwixZimProvider() is
the default, and tests inject another provider without changing lifecycle or
tool-routing code. The ZIM provider follows redirects, decompresses
uncompressed and Zstandard clusters, selects a bounded passage around matching
query terms, and returns the resolved canonical Wikipedia URL plus embedded
archive language/date/source/license metadata. Local archive text uses the same
untrusted-result boundary as live third-party content.
WebBrain reads the documented openZIM structures directly, using the MIT-licensed
fzstd decoder, and that reader stays the only path used for reading articles.
Full-text search over a ZIM's Xapian index uses the GPL worker vendored under
src/*/vendor/libzim/. ZIM_XAPIAN_RUNTIME_BUNDLED is true. Catalog archives
that include an index (X/fulltext/xapian or the older
Z//fulltextIndex/xapian) answer with snippets. A conceptual query no longer
has to name the article title. See docs/offline-rag-licensing.md for the
decision, the release licensing strategy, and the reproducible -O2 source build.
Standalone retrieval detects the question language from script, multilingual query words, and the interface locale fallback. A matching installed archive is searched before newer archives in other languages. The user's language filters remain authoritative: an empty filter searches every installed language, while selected ISO 639-3 codes exclude other-language passages from the final evidence.
When the question language differs from the installed Wikipedia archive languages that will be searched, the already-downloaded WebGPU text model first produces short search keyword translations for those archive languages. This works both with the default "all installed languages" setting and explicit language filters. The translations are bounded, validated as data, and used only as per-archive Xapian queries; they are never stored as conversation messages. A failed or malformed translation causes no download and falls back to the original query. This improves cross-language lookup but remains dependent on the selected text model's language ability.
hasFullTextIndex() on the reader reports whether an archive carries an index at
all, checking both the current X/fulltext/xapian entry and the older
Z//fulltextIndex/xapian one. Every catalog edition has one. A manually imported
.zim may not, and libzim gives no way to tell the difference on its own: its
search() swallows the error and returns no results, which is indistinguishable
from a query that matched nothing. Probing the archive directly keeps that
distinction honest and lets an unindexed archive skip loading the runtime.
- OPFS quota and eviction policy are browser/profile specific. The pre-install estimate is informative, not a reservation.
- Chrome Manifest V3 background workers are ephemeral; persisted jobs and alarms resume piece downloads after the worker restarts.
- Firefox uses a persistent extension background page, but large storage quotas and OPFS behavior can still differ by version and device.
- Private/incognito profiles, profile clearing, extension removal, or browser storage eviction can remove archives.
- Very large archives may be impractical on mobile or low-storage devices.
Catalog metadata comes from the Kiwix OPDS catalog, the file format is documented by openZIM, and archive content remains subject to the license embedded by its publisher.
See Remote Downloads & Data Sources for download triggers, execution order, and integrity verification.