Homework Ocr Intake

Return normalized JSON draft:

{
  "subject": "algebra|geometry|unknown",
  "grade": 7,
  "sourceLanguage": "ru",
  "detectedItems": [
    {
      "exercise": "№ 123",
      "topicHint": "линейные уравнения",
      "rawLine": "Алгебра: №123(а,б)",
      "statementPreview": "краткий фрагмент условия, если читается",
      "confidence": 0.86
    }
  ],
  "needsClarification": false,
  "clarificationQuestion": null
}

Workflow

1. OCR the full image.
2. Split by lines and detect subject markers (алг, геом, №, п., стр.).
3. Normalize exercise identifiers (123, 123а, 123(а,б)).
4. Assign per-line confidence.
5. If confidence is low (<0.65 for critical lines), switch to clarification mode.

Clarification mode

If OCR is uncertain, do not invent the assignment. Return a short RU-first clarification prompt, e.g.:

• "На фото не читается номер задания после '№ 241'. Пришли, пожалуйста, более чёткий снимок строки с номером."

Guardrails

• Never fabricate unreadable symbols or numbers.
• Never silently merge two different subject lines.
• Preserve rawLine for auditability.

No-plan-only handoff

• OCR stage must preserve enough structured items (exercise, subject, rawLine) for solver stage to produce real solved answers, not a study plan.
• If line contains numbered tasks (e.g. №455,457), keep exact numbers for textbook retrieval.