Skip to content

IntPortal for Android 0.1.8 beta 1

Pre-release
Pre-release

Choose a tag to compare

@cGradying cGradying released this 26 Oct 12:13
· 299 commits to 9b4ce73dd55bb2b5d0e7aeb155ef876c4b11743b since this release

Install this one if you have 0.1.5, 0.1.6 or 0.1.7. Those APKs crash as soon as they open (R8 stripped a method Tauri needs). This build fixes it and installs over them as an update.

What's new

  • IntAssis: Qwen3 1.7B is the default chat model. If you already have Gemma installed, you'll see an Update IntAssis button. It downloads a 1.1 GB chat model and 310 MB for note search, so use Wi-Fi.
  • Note search: uses EmbeddingGemma-2.
  • Refusals: LAR refuses off-topic, prompt-leak, homework and harmful questions in its own voice. Self-harm replies point to the NCMH crisis hotline: 1553 or 0917-899-8727.
  • Chat: LAR talks like a person, and chit-chat is answered instantly.
  • Due dates: answers cover overdue items, eMabini and this week.
  • Reminders: the app asks for notification permission.
  • Download sizes: shown item by item.
  • Filipino: questions in Filipino get Filipino answers (eval: Qwen 18/24, Gemma 21/24).
  • No invented answers: says "I don't have that" instead of making things up (near-miss questions: 0 of 12 invented with retrieval).
  • Cards: cards from notes or a PDF honour the count you ask for, with a preview.

Known gaps

  • Qwen's Filipino is garbled in about 1 reply in 4.
  • Chat source chips are not restored after relaunch (#180).
  • Not yet run on a real phone with Qwen + EmbeddingGemma-2, so speed and memory are unverified. Check on a real iPhone and Android phone before wide rollout.
  • The in-app model download was only simulated by placing the files.
  • The engine is pinned to an unreleased upstream llama.cpp commit (cb98c630).
  • Emulator results (Android 15 arm64, 4 GB RAM):
    • The release APK installs and opens to sign-in.
    • On a debug build of the same code: demo sign-in works, and the reminder notification-permission prompt shows.
    • The in-app IntAssis download works for real: 1.42 GB in about 16 minutes, both files at full size.
    • IntAssis did not answer. Loading Qwen + EmbeddingGemma-2 ran for over 4 minutes, Android killed other apps for memory, and IntPortal showed "isn't responding".
    • The debug build's AI engine is unoptimized, so this doesn't measure release speed. It does show memory is tight on a 4 GB phone. Treat IntAssis as untested until it's checked on a real phone.
  • Phones with 4 GB RAM may stall when IntAssis first answers (being fixed). Until then, stay on Light, or pick the lighter Gemma 3 1B chat model, on those phones.
  • Today's LAR card says "Nothing due" while eMabini lists an overdue item right below it (#187).

Signed with the same key as earlier betas (SHA-256 5b1449…bf8027).