[Contribution proposal]: Source-grounded interactive self-checks and visual chat responses

[CaseresMaxi]

Problem and desired outcome

When studying from notebook sources, learners need to test their understanding and sometimes reason about a diagram or a chart. Switching to another application loses the source context. I would like a source-grounded explanation to include an interactive self-check at the relevant point, with feedback after submitting answers, and genuine source figures where they help.

This proposal builds on the existing product conversations: #1150 (practice exams), #203 (content modes and artifacts), and #566 (the canonical image Discussion). I recognize that #1150 is still needs-design and that its author already has an implementation; this contribution should be coordinated with that work.

Working prototype

Draft PR: #1486. It remains a design-review prototype pending approved scope.

I have a self-hosted implementation on https://github.com/CaseresMaxi/open-notebook/tree/feature/practice-exams, offered as a draft for design review:

• Notebook practice exams generated from selected sources/notes, with deterministic grading where possible and AI feedback for open answers.
• Short interactive tests inside notebook and source-chat replies; answer keys stay out of normal chat payloads until submission, and attempts persist with the conversation.
• Image attachments in chat, source page previews and genuine crops with source/page provenance, and positioned figures in prose or exam questions.
• Optional generated illustrations on explicit request, clearly distinguished from source evidence. The current generation adapter uses an OpenAI-compatible Images API.
• Explicit requests activate image tools without requiring the composer toggle; ordinary requests for an exam produce a card, with bounded structured conversion when the model returns plain questions.

REST endpoints, migrations, frontend components, translations in 14 locales, tests, user docs and architectural records are included. Local validation: 1,082 backend tests and 229 frontend tests, lint/type checks, a Docker build, and actual browser submission/reload checks. This does not establish compatibility across the complete provider matrix.

Design decisions requested

1. Should an inline test be an artifact of a shared type under [Vision] Content Modes & Artifacts (study & visual outputs) #203, with attempts stored separately, rather than the dedicated exam/chat-quiz records used by this prototype?
2. Is an explicit request in chat the right initial trigger, and should unsolicited pedagogical self-checks be disabled or opt-in?
3. How should source figure locators, attachment persistence and citation provenance align with Discussion Images in Open Notebook: ingestion, display, and visual understanding #566 and the evidence-centered direction?
4. Is optional image generation in scope? The current OpenAI-compatible adapter needs a deliberate portability/product decision; source crops and quizzes can be reviewed independently.
5. Which small implementation slices would be useful to review first? Image attachments/source crops, inline tests, and standalone exams can be separated following maintainer guidance.

I am offering the code as concrete evidence for these decisions and am willing to reshape the storage/UI, narrow the scope, or split the contribution. Please route any accepted slices into approved Issues before this becomes a ready-for-review PR. I am not asking to bypass the existing needs-design process or close #1150 with this prototype.