Plan Ai Conversation

Plan: an AI conversation partner, within what the learner knows

Source of truth: this wiki page.

Written 2026-10-05. A long-term plan.

What the owner asked

AI conversation partner (Duolingo Max, Babbel): chat limited to the learner's known vocabulary and current grammar.

long term future plan. will use user's API / local AI (downloaded within the app, something like gemma 4 E4B)

What exists

• Nothing that generates language. Every word the app shows comes from the decks.
• What the learner knows is in the review log: every card taught, and per skill, how well it is remembered. The grammar steps taught are the path's grammar units done. So "known vocabulary and current grammar" can be listed exactly.
• The app holds no secrets and makes no account. It reaches GitHub only for updates and, once planned, decks (ADR-0037).

Two ways to run the model

1. On the phone: Google's Gemma 4, released on 2 April 2026, has an E4B size built to run offline on phones.
   • The Flutter plugin flutter_edge_ai 2.0.0 (MIT) runs Gemma and other models on the device. It is the new name of flutter_gemma, renamed on 5 October 2026.
   • The model is a download of several gigabytes, offered inside the app, never bundled. Its size and licence terms are to check before choosing it.
   • It needs a recent phone, with enough memory.
2. The learner's own API key: for example Anthropic's Claude API, where claude-opus-5-5 is the current default model. There is no official Anthropic SDK for Dart, so the app would call its HTTP API directly.
   • The key stays on the phone, in the platform's secure storage. It is never sent anywhere but to that provider.
   • Messages leave the phone, which the app must say before the first chat.

What it takes

1. The learner's limits, as data:

   • the known words, with their forms;
   • the grammar steps taught;
   • the level reached (A1, A2, B1, from aaronified/fluenough#187).

   They are sent to the model as its instructions for each chat.

2. Keeping the model within them:

   • each reply is checked against the known words;
   • a reply with unknown words is asked again, or the words are marked and offered as cards (import-text.md);
   • grammar beyond the learner's steps is flagged, not hidden.

3. Scenarios taken from the path's themes, such as market or directions, so a chat practises what the unit taught.

4. Feedback on the learner's messages, which reuses explain-mistakes.md's rules where one applies.

5. Speech, optionally: the learner speaks and hears the replies, with the recogniser and voices the app already uses.

6. Privacy:

   • on the phone, nothing leaves it;
   • with an API key, the learner is told which provider receives what;
   • no chat is saved unless the learner keeps it.

7. Tests for the limits, the vocabulary check, and the prompts. They use a fake model, since real models cannot run in CI.

To decide

• On the phone, the API, or both, and which comes first.
• Which providers an API key may be for.
• Whether chats count as practice in the review log, and in which skills.
• The smallest phone the on-device model must run on.

Estimate

About 20–30 hours for a first version with one of the two ways. Confidence: low; this depends on the models available when it starts.

See also: Roadmap · Plans-Index · Home