Eva is a fully offline AI assistant for Android. The language model, speech recognition, text‑to‑speech, document search, maps, music and Wikipedia all run on the phone. No account, no cloud, nothing leaves the device.

⬇ Download the latest APK (arm64‑v8a) — or run binary/get-latest-apk.sh.

What Eva does

Eva is a chat assistant and an offline media/knowledge hub. The home screen has four tabs — Chat, Images, Music, Docs — and music keeps playing in the background with lock‑screen controls while you use the rest.

Features

💬 Chat & assistant

• On‑device chat with selectable models (LFM2.5, Qwen3, vision models), downloaded on demand or sideloaded from a folder.
• Grounded answers: Eva fuses three sources — your documents (hybrid vector
  • full‑text retrieval), the offline Wikipedia, and the model's own knowledge — and cites them. Tap a citation to open the source at the passage.
• Digital assistant: set Eva as the phone's assistant and invoke it with a power‑button hold — it listens, answers, and speaks the reply.
• Voice in/out: offline streaming speech‑to‑text (English) or the phone's recognizer (many languages), plus spoken replies.
• Vision: attach a photo and ask about it (with a vision model selected).
• Persistent chat history with a conversation drawer; reply‑length control; each answer shows how long it took.

🗺️ Offline maps & navigation

• Ask "where is X" / "show me X on a map" → an expandable map tile in chat that opens full‑screen with your live GPS dot and a distance/bearing readout.
• "route to X" draws a walking path from your position to the destination.
• Streets and satellite layers, cached on demand to a folder you choose, so places and routes you've viewed keep working offline. "Save this area offline" caches a region for later.

🎵 Music — a full offline player

• Browse Favourites (most played), Folders, and Artists.
• Folders are grouped by an LLM‑assigned genre → subgenre (so bands without a genre tag still get sorted — e.g. alternative rock, black metal, funk).
• A "+" on every track adds it to the playlist; tap the now‑playing bar to open a full player: seek bar, transport, shuffle/random, and the current queue (jump / remove). Save, load and delete named playlists.
• Background playback with lock‑screen / notification controls.

📚 Docs — an e‑reader

• Wikipedia: read the offline encyclopedia with search‑as‑you‑type (ranked by article popularity), a random article, the main page, and a recently‑viewed list.
• Folders: documents categorised by the LLM into genre → subcategory, with live progress, reusing the already‑extracted text.
• Favourites: recently read, starred, and often‑read documents.
• PDFs resume at the last page ("Page X of Y"); Word/EPUB/text render in full and resume your scroll position with a "% read".
• Index PDFs, Word, PowerPoint, Excel and EPUB — fully on‑device.

🖼️ Images — a gallery

• Search by the on‑device caption (a vision pass labels photos while charging) and by your own tags; filter by category (photos / screenshots / memes), group by day, and view GPS‑located photos on the map.

📻 Radio · ⏰ Reminders

• Online radio stations with play counts/favourites; "play radio …" from chat.
• Timed reminders — "remind me … in 15 minutes" / "at 7pm" (EN/PT) — fire as system notifications even when Eva is closed.

⚙️ Settings & storage

• Settings organised into per‑area panels (Appearance, Language model, Voice, Documents, Photos, Music, Radio, Wikipedia, Maps, Reminders, Storage).
• One unified storage folder for all data (models, Wikipedia, documents, map cache); choosing a new folder (e.g. an SD card) moves everything so nothing has to be re‑downloaded.

Everything stays on the device

Eva is built to work with no connectivity. The model, embeddings, speech, the document index (a sharded int8 vector index fused with full‑text search), the ZIM‑backed Wikipedia, GPS, cached map tiles and your media all live on the phone. Internet is only used for optional on‑demand downloads (a model, a Wikipedia edition, a map tile, lyrics, a radio stream) — never for your data.

Architecture

• Flutter app (flutter/example/lib/) talking over FFI to:
  • the Cactus inference engine (libcactus.so) for the LLM/VLM/embeddings,
  • libzim (libzimffi.so) for the offline Wikipedia.
• A background isolate runs inference; gated background passes (photo captioning, document & music categorisation) use the model while charging / idle and yield instantly to chat.
• Native libraries are content‑addressed and built by a Gradle hook (an NDK is only needed when engine sources change).

Repository layout

• flutter/example/ — the Eva Flutter app (lib/), its native FFI builds (native/, native_zim/) and on‑device integration tests.
• cactus/, cactus-engine/, cactus-graph/, cactus-kernels/ — the Cactus inference engine sources compiled into libcactus.so (see LICENSE).
• python/ — the host‑side transpiler that converts HuggingFace models into loadable bundles (used by CI; the phone only loads pre‑transpiled bundles).
• .github/workflows/android-release.yml — CI: transpiles missing model bundles and publishes the APK to the GitHub release.
• binary/ — convenience script to fetch the latest released APK.

Building

cd flutter/example
flutter build apk --release --target-platform android-arm64

The native library is built (or restored from a local cache) automatically by a Gradle hook; an Android NDK is only needed when engine sources actually change.

Credits

Built on the Cactus inference engine (Cactus Compute, Inc. — see LICENSE). Eva is maintained by Geogram.