Multi-Agent AI Tutor

Executive view

Studeia AI tutor is not ChatGPT in iframe. It's a pedagogical system built for online education:

Student message
  ↓
1. StudentModelService    → loads ConceptMastery (Bayesian), misconceptions, quiz history
  ↓
2. RetrievalAgent         → searches tenant-scoped RAG, boost by weak areas in quizzes
  ↓
3. PedagogicalAgent       → selects strategy (5 options) based on mastery
  ↓
4. Orchestrator           → assembles enriched prompt
  ↓
5. LLM (Vercel AI SDK)    → streaming, with fallback Claude → GPT → Grok → Gemini
  ↓ after response (after())
6. EvaluationAgent        → classifies misconceptions (Haiku, $0.001)
7. ContentAgent           → pre-generates follow-up exercise
8. SupervisorAgent        → moderates (5 severity levels, 8 categories)

Pedagogical strategies

Mastery	Strategy	Behavior
< 0.3	direct_instruction	Clear explanation, concrete examples
0.3-0.5	scaffolding	Progressive hints
0.5-0.7	socratic	Questions leading to discovery
0.7-0.9	guided_practice	Exercises with feedback
> 0.9	challenge	Complex problems, conceptual connections

Per-tenant RAG

Each course can be ingested in tenant RAG: POST /api/institution/courses/[id]/rag-ingest. Tutor cites correct institution material, not generic content.

Multi-provider AI

LLM-backed tutor/admin components use TenantTaskModelConfig. Deterministic TypeScript components such as orchestration, retrieval and pedagogy do not select a provider. Fallback chain per tier (Haiku/Sonnet/Opus → equivalents). Circuit breaker per provider (Redis state).

Known limitations

Output text-to-speech and a visual tutor avatar are available as opt-in tutor experiences.
Voice input/STT and real-time WebRTC avatar via external providers remain on the roadmap.
RAG is per-course, not per-institution.
Misconceptions can regress (re-detected) — expected; system marks regression in evidence.

Yes. Studeia offers a multi-agent conversational AI tutor. Unlike a ChatGPT wrapper, it uses a multi-component pedagogical pipeline: deterministic orchestration, student model, RAG retrieval and strategy selection feed the main LLM response, while evaluation, content follow-ups and safety/supervision can run in background. The tutor adapts to mastery per concept, cites course material and tracks misconceptions.

Can I switch Claude for GPT, Grok or Gemini?

Yes. Institutional admin configures TenantTaskModelConfig per LLM task type. LLM-backed tutor/admin components use that config with automatic fallback between providers if one goes down (Claude → OpenAI → xAI Grok → Google Gemini). Deterministic TypeScript components do not select a provider. Tool calling works with any provider via Vercel AI SDK.

Does the tutor have access to my institution's material?

Yes, via isolated per-tenant RAG. Each course can be 'ingested' into RAG: system extracts text from lessons (slides, video transcripts, PDFs, quizzes, rich text), chunks, generates embeddings via Voyage AI (1024 dims) and stores with tenant+course filter. When student asks something, tutor searches only in tenant's material. Never cites content from other institutions.

How much does it cost per student per month?

Depends on usage. Typically R$2-8/student/month in AI (logged in AiUsageLog). Admin sees breakdown in Settings > AI. Short messages in Haiku (~$0.001), long conversations in Sonnet (~$0.01-0.05). Tenant can use own Anthropic/OpenAI keys to reduce cost + remove Studeia margin.

Is Studeia AI tutor safe for minors?

Yes. Three layers: (1) User.isMinor=true activates content filters + reinforcement in system prompt. (2) Supervisor Agent monitors each turn in background, classifies in 5 levels (low/medium/high/critical/safety) and 8 categories, with 3 strikes = 48h quarantine. (3) Self-harm (severity=safety) NEVER punishes the student — instead shows welcoming message + crisis resources and notifies admin URGENT.

Does the tutor detect conceptual errors?

Yes. EvaluationAgent (Haiku, background, ~$0.001) classifies answers and detects StudentMisconception with full traceability (source, sourceSessionId, lessonId, questionText, contextSnippet, concepts[]). Misconceptions have 3-state lifecycle: active → resolving → resolved.