Skip to content

mod.ai-grader — AI grading with rubrics

Community · AI category (can be disabled)

What it does

AI-assisted grading for the open-ended questions in mod.assessments (SHORT_ANSWER, LONG_ANSWER). The instructor defines a rubric per question (instructions + 1-8 weighted criteria); the AI reads the rubric and the student's answer and proposes a score with a per-criterion justification plus overall feedback. The score is never published on its own: the instructor reviews and applies (or discards) the suggestion. The module records AI–human agreement metrics.

How it works

  • The criteria weights must add up to the question's points (validated when the rubric is saved).
  • There is one suggestion per answer: regenerating with force replaces the previous one; without force, the cached one is served without calling the model again.
  • "Apply" on a suggestion only marks it as applied for auditing — the real score is set through the manual grading in assessments (POST attempts/:id/grade), keeping a single grading path.
  • Suggestions are persisted so they can be reviewed later without spending tokens.

Configuration

Enabling the module. Under /admin/configuracion?tab=modules (the "Modules" tab). The module contributes the "Grading" entry to the Instructor group of the sidebar (/formador/correcciones). If the module is disabled, the manual grading page keeps working: AI suggestions simply do not load.

AI provider (BYOK). It only uses the AI Gateway's chat purpose: the "Chat (AI tutor)" block under /admin/ia/providers ("Integrations and API → AI providers" in the menu) or, failing that, the global default DEFAULT_AI_CHAT_*. Chat providers the code accepts today: OpenAI, Anthropic (Claude), Google Gemini, OpenRouter, Mistral AI, Groq and Ollama (self-hosted). Gateway and key-encryption details in Configuration → AI. With no provider configured (or if the provider fails), the suggestions request returns 502 AI_GRADER_PROVIDER_ERROR; manual grading is unaffected.

The chat BYOK block that powers grading suggestions

Rubrics. There is no rubric editing screen yet: they are managed through the API (GET/PUT/DELETE /modules/ai-grader/questions/:questionId/rubric, instructor role and above). Validations on save: instructions from 10 to 2000 characters, 1 to 8 criteria, each criterion with a name, a description and an integer weight (1-100), and the weights summing to the question's points.

No variables of its own, no per-tenant settings and no Enterprise license requirement.

Step by step

Preparation (instructor, through the API)

  1. Create the rubric for each open-ended question of the quiz with PUT /modules/ai-grader/questions/:questionId/rubric, body { "instructions": "...", "criteria": [{ "name": "...", "description": "...", "weight": n }] }. Questions with no rubric are skipped when generating suggestions (that is not an error).

Grading (instructor, in the panel)

  1. Open "Grading" in the menu (Instructor group): the "Pending gradings" page lists attempts with open answers waiting for review, with the "Pending" badge and "N answers to grade".

    The instructor's queue of pending gradings

  2. Enter an attempt (/formador/correcciones/[id]): you will see each question with the "Student's answer" and, for the open-ended ones, the "Points (max N)" and "Feedback (optional)" fields.

  3. In the "AI grading assistant" block, click "Suggest grades with AI". One suggestion is generated per open question with a rubric; the notice sums it up: "N suggestions generated.", "N questions without a rubric." and "Tokens: N.".

  4. Each suggestion appears under its question with the "AI suggestion" badge, the proposed score ("X/Y pts · provider"), the overall feedback and the per-criterion breakdown with its justification.

    The attempt detail with the AI suggestion broken down per criterion

  5. Click "Apply to the form" to copy the score and feedback into the grading fields — you can edit them before submitting. Applying only marks the suggestion for auditing. "Re-generate" forces a new model call and replaces the suggestion; without regenerating, re-entering the attempt serves the persisted one without spending tokens.

  6. Click "Submit grades": the grade is issued by mod.assessments (the single grading path), the student sees the result and the lesson is marked complete if they passed.

Dependencies

Optional: mod.assessments (the source of attempts, answers and questions).

Data model

mod_ai_grader_rubric (one per question: instructions + weighted criteria as JSON) · mod_ai_grader_suggestion (proposed score, per-criterion breakdown, provider/model, application traces).

API

Prefix /modules/ai-grader (instructor and above): rubric management and the suggestion cycle. Details in Reference → Payments, classroom and AI.

Events

Emits: ai-grader.suggestion.generated, ai-grader.suggestion.failed. It consumes none.