FAQ

Is my data safe?

In Local mode all your data lives in IndexedDB on your own device. No backend, no third-party service. Closing the browser tab doesn't delete it; clearing site data does. If you share the device, anyone with access to this browser profile can read it.

In Server mode the data lives in the SQLite database the FastAPI backend manages. API keys are encrypted at rest with Fernet using a secret you set via the ADAPTIVE_LEARNER_SECRET_KEY environment variable, or via secret_key: in ~/.config/adaptive-learner/secrets.yaml.

Neither mode sends telemetry, analytics, or your messages to any third party other than the AI provider you've chosen — and that one only sees the message content you'd expect (system prompt + your text + the AI's prior responses in the session).

Do I need an API key?

Yes for AI sessions. The app uses bring-your-own-key for all three supported providers: Anthropic Claude, OpenAI GPT, Google Gemini. Free-tier limits are usually enough to get started.

Three places to put the key (highest priority wins): an ADAPTIVE_LEARNER_<PROVIDER>_API_KEY env var, the ai.<provider>.api_key field in ~/.config/adaptive-learner/secrets.yaml, or the Settings UI. The UI shows the per-provider source so you always know where your key came from.

You can browse the Curriculum, take the Assessment, view your Dashboard, and even run the chat-history Import without an API key. The Session page + the analysis step + the AI-extraction features are the ones that need a key.

Can I use it offline?

Partially. The PWA service worker caches the static assets (HTML, JS, CSS, icons) so the app launches without internet. Past sessions and Dashboard data load from local storage too, so reading old material works fine.

Live sessions still need internet because the AI provider sits outside your browser. The Session page detects "offline" and shows a clear inline message rather than failing silently.

What does the method switch mean?

When three sessions in a row show your understanding stagnant AND your stress high, the app surfaces a banner: "Want to try [other method] for the next session?". The recommendation prefers your second-strongest method from the assessment that you haven't used recently.

It's a suggestion, not an order. You can dismiss the banner and continue with your current method; the banner reappears if the stagnation pattern continues. Method switches are recorded in the method_switches table and surface in the Progress page distribution.

What is auto-loop?

When a session reaches step 7 (Integrate) and the topic-transition evaluator judges the topic integrated AND recommends continuing, a fresh cycle starts automatically with a new subtopic. Up to 5 cycles per session (runaway protection). The chat history renders dashed-border "Cycle N" cards at each transition. The end-of-session rating dialog summarises the multi-cycle journey when cycle_count > 1.

Can I export my data?

Yes. Three export paths shipped:

• Backup: Settings → Backup → Create Backup. Downloads a timestamped JSON with every row from your account. API keys are stripped. Works in both storage modes.
• Progress / Session / Curriculum reports: Settings → Export. Markdown + PDF (browser print-to-PDF).
• Anki .apkg: review AI-extracted flashcards on the /anki page, accept the ones you like, click Export. The file works directly in Anki desktop.
• NotebookLM ZIP: from the Progress page, download a structured ZIP (summary + vocabulary + rules + errors + flashcards + sessions) formatted for NotebookLM's source upload.

What's the voice feature?

Three Web Speech API integrations (since v1.18.0):

• Text-to-Speech on AI replies + Assessment results — a ▶ button next to each speaks it aloud, language-matched.
• Speech-to-Text on the Session input — a 🎤 button captures your voice and populates the textarea with interim transcripts before send.
• Pronunciation Practice for language projects — visit /pronunciation, the AI generates a target phrase, you speak, and a judge AI scores similarity + suggests improvements.

Voice toggles live in Settings → Voice. The section hides itself in browsers that don't support the API.

What's the chat-history import?

The Import page (/import) accepts pasted or uploaded chat transcripts from ChatGPT, Claude.ai (both JSON bulk export and single-conversation Markdown export), Gemini, and arbitrary Markdown. The analyzer extracts your topic, weaknesses, error patterns, recommended method, vocabulary (for language conversations), and a suggested curriculum. One click seeds a Curriculum + starts a targeted session from the analysis.

The Claude.ai per-conversation Markdown export was the v1.19.0 audit case — the parser ships with full timestamp extraction + role boundary preservation for that format (BL-25 / BL-26 / BL-28 closed in v1.19.1).

Sync between devices?

Local-network bidirectional sync since v1.0.0. Settings → Sync → "Pair this device": scan the QR code on the other device's screen (rear camera), or paste the pairing URL. Once paired, push + pull buttons exchange data; conflicts go through an AI-merge resolver. 28 tables on the sync surface as of v1.19.0 (subjects + tags + study questions included).

How is this different from ChatGPT?

ChatGPT is a chat interface to a single model. Adaptive Learner is a structured learning system that uses an AI under the hood but adds:

1. A 6-method × 7-step matrix of bespoke system prompts.
2. Per-turn step evaluation — a second AI call judges readiness and may move you forward / back.
3. Auto-loop into new cycles when the topic is integrated.
4. A profile of your learning preferences from the 12-question assessment.
5. Long-term tracking — ProgressCommits, streak heatmap, XP, badges, time-per-step charts. ChatGPT forgets when you close the tab.
6. Provider freedom — Anthropic, OpenAI, or Gemini.
7. Local-first option — everything in your browser, nothing sent to a server (except your AI calls).

What if the AI goes wrong?

The system fails visibly:

• Wrong API key: the AI call returns a clear error message, surfaced inline in the chat.
• Provider down: same — the error renders the HTTP status from the provider's API.
• JSON parse failure from the evaluator: a deterministic +1 advance kicks in (capped at step 7), with fallback_used: true recorded so a future audit can spot models that struggle with the format.
• Streaming aborted mid-response: the partial reply is saved; the next message picks up from there.
• Stale or weird AI response: end the session, give it a low rating, retake. The method-switch heuristic will surface a different method if the pattern persists.