Personalized Educational Tutoring
Dynamically select models based on student level, covering everything from basic Q&A to competition coaching. Use different tiers of models for questions of different difficulty levels, ensuring both explanation quality and the lowest possible per-session tutoring cost.
Tiered Model Selection
| Level | Scenario | Recommended Models |
|---|---|---|
| Basic Q&A | Concept explanations, homework hints | gpt-5.4-mini, gemini-3.5-flash |
| In-depth Explanation | Solution steps, error analysis | gpt-5.5, claude-sonnet-5 |
| Competition Coaching | Competition problems, proof problems | claude-opus-4-6-thinking, gpt-5.5-high |
| Photo-based Problem Search | Handwritten/printed problem recognition | Vision models, see Image Understanding |
Implementation Notes
- Use streaming output to present the explanation process word by word, giving students an experience close to live tutoring;
- Fix the teaching style in the system prompt (Socratic questioning, no direct answers, etc.), and preserve wrong-answer context across multi-turn conversations;
- Split tokens by student or class and set quotas; parents can view usage through the balance query API;
- Illustrated handouts can be paired with image generation, see Image and Video Generation Models.