EduVis Schema

An open, curriculum-aware framework and knowledge representation for learning experiences.

🚀 Try the Live EduVis Studio IDE

EduVis describes the educational meaning of learning experiences — separating semantic pedagogical structure from raw presentation rendering. Inspired by the philosophy of Markdown, Mermaid, and Model Context Protocol (MCP), it aims to standardize how learning environments are modeled, compiled, and visualized.

EduVis consists of two complementary projects:

• EduVis Core: The stable educational Intermediate Representation (IR), compiler pipeline, schema validator, and reference SVG renderer.
• EduVis Studio: A browser-based educational IDE built on Pyodide/WASM that acts as the primary authoring, visualization, and debugging environment for educational content designers.

Inspired by the philosophy of separating meaning from rendering:

Separate meaning from rendering.

Who Is EduVis For?

Getting Started

Requirements: Python 3.10+

Option 1: Install from PyPI (Recommended for general use)

# Core framework only
pip install eduvis

# With Agentic Generation substrate (v1.3 with LangGraph & FastAPI)
pip install "eduvis[agent]"

Option 2: Clone and run locally with uv (Recommended for development)

git clone https://github.com/seehiong/eduvis
cd eduvis
uv sync --all-extras

Then prefix commands with uv run (or use the globally installed eduvis if installed via PyPI/pip):

# Validate showcase lessons
uv run eduvis validate showcase/lessons/negative-numbers-confidence-ladder-lesson.yaml
uv run eduvis validate showcase/features/adaptive-remediation-branching-lesson.yaml
uv run eduvis validate showcase/features/visual-elements-catalog-lesson.yaml
uv run eduvis validate showcase/features/assessment-schemas-lesson.yaml

# Render all showcase assets to showcase/assets/
uv run python scripts/build_showcase.py

# Or render a single showcase lesson to a custom directory
uv run eduvis render showcase/lessons/negative-numbers-confidence-ladder-lesson.yaml -o output/negatives/

# Utility commands
uv run eduvis docs --subjects math
uv run eduvis schema -o eduvis/schemas/

# Curriculum graph inspection
uv run eduvis graph inspect showcase/reference/showcase-curriculum.yaml
uv run eduvis graph prereqs showcase/reference/showcase-curriculum.yaml rational_numbers --transitive
uv run eduvis graph dependents showcase/reference/showcase-curriculum.yaml integers
uv run eduvis graph path showcase/reference/showcase-curriculum.yaml integers real_numbers

# Mastery projection
uv run eduvis mastery project showcase/reference/showcase-curriculum.yaml showcase/reference/sample-learner-state.yaml

# Study plan generation
uv run eduvis study-plan showcase/reference/showcase-curriculum.yaml showcase/reference/sample-learner-state.yaml --mode exam_prep --hours 2

# Schema migration (auto-detects current version, or override with --from-ver)
uv run eduvis migrate showcase/lessons/ --to-ver 1.0

# Graph-driven lesson generation
uv run eduvis generate lesson showcase/reference/showcase-curriculum.yaml integers negative_numbers -o new_lesson.yaml

Option 3: Install locally with pip

git clone https://github.com/seehiong/eduvis
cd eduvis
pip install -e .

Run the Tests

# Sync all dev and agent test dependencies
uv sync --all-extras

# Run full test suite (including core & agent tests)
uv run pytest tests/ -v

# Or run agent tests specifically
uv run pytest tests/test_agent.py -v

Verify the Agentic Pipeline End-to-End

Use scripts/verify_agent.py to prove the full 3-stage agent pipeline is working against your local Ollama instance:

# Requires: ollama serve (in a separate terminal) with at least one model pulled
python scripts/verify_agent.py

This script runs through three stages and prints a clear verdict for each:

Success criteria: Status: success, Is Valid: True, and LLM-sourced questions: 10 in the output. If Fallback questions: 10 appears instead, the LLM produced unusable JSON and the deterministic fallback fired — re-run once the model is warm in VRAM.

Code Quality and Maintainability Checks

To keep the codebase maintainable and clean, we enforce code quality and maintainability limits locally. These checks monitor:

1. Cyclomatic Complexity: Functions must have a McCabe complexity $\le 10$.
2. Function Length: Functions are restricted to a maximum of 50 statements.
3. Nesting Depth: Block nesting depth is restricted to at most 3 levels.
4. Duplication Detection: Code replication/duplication of 4 or more lines is flagged.
5. Module Size: Python files (modules) are limited to a maximum of 1000 lines.

To run these checks, ensure your development dependencies are synced:

uv sync --all-extras

Then run all quality checks with a single command:

uv run python scripts/run_checks.py

Alternatively, you can run the individual tools manually:

# Run Ruff for style, complexity, and function length
uv run ruff check eduvis agent tests scripts

# Run Pylint for nesting depth, module size, and duplication
uv run pylint eduvis agent tests scripts

[!TIP] Troubleshooting Pylint Crashes: If Pylint crashes with an astroid error or reports bogus wrong-import-position errors on every file (especially common on Python 3.12+), it usually means uv run is falling back to an outdated global installation. Force an update of your local virtual environment by running uv sync --all-extras to fix this.

To see every registered element type rendered to SVG in one pass:

uv run eduvis render showcase/reference/exhaustive-element-catalog.yaml -o output/exhaustive_catalog/

This produces one SVG per element type (test_number_line.svg, test_text_list.svg, test_math_grid.svg, test_solid_cube.svg, … test_solid_cylinder.svg) — useful for checking renderer output after code changes.

What EduVis Is

EduVis is not an SVG schema.

EduVis is a machine-readable instructional model.

It captures the structure of good tutoring — the same structure behind effective human instructors: no skipped steps, visual intuition before abstraction, confidence-building before challenge, and retrieval to lock it in long-term.

Crucially, it separates presentation (how to deliver), curriculum (what to learn), and student cognition (what is mastered) into orthogonal layers, allowing learning paths to adapt dynamically.

EduVis is to educational experiences what Markdown is to documents, Mermaid is to diagrams, and Model Context Protocol (MCP) is to context.

A specification can be rendered as SVG, PDF, slides, interactive lessons, or animated videos while preserving pedagogical intent.

The Problem

Most diagram libraries describe visuals. EduVis describes learning experiences.

A number line in a textbook, a number line used to discover a rule, and a number line shown during a recall exercise are pedagogically different objects. They happen to look the same. Today's tools treat them identically.

# What every library gives you
type: number_line
range: [-10, 10]
highlight: [-3, 5]

EduVis preserves the meaning that gets lost the moment most tools export to SVG:

id: explore_number_line
type: number_line
placement:
  lesson_phase: explore
  memory_role: anchor
  difficulty: starter
actions:
  conceptual:
    - compare: [-3, 5]
range: [-10, 10]

But more importantly, EduVis describes where this element sits inside a proven teaching pattern — something no diagram library models at all.

Specification Status & Tooling

This is not a theoretical schema. The placement model, element types, and LLM prompt vocabulary have been validated in real educational pipelines and are designed for production use.

The EduVis Studio is the primary IDE for educational design. EduVis Studio is not another YAML editor. It is a multi-projection educational IDE where text, graphs, diagnostics, and compiler views are synchronized representations of the same educational specification. Every projection is a synchronized view over the same educational specification; no projection owns the data.

A single educational specification is projected instantly into six interactive views, supported by a cross-cutting inspector:

             Educational Specification (Source of Truth)
                                │
      ┌─────────────────────────┼─────────────────────────┐
      ▼                         ▼                         ▼
  Curriculum View        Storyboard View        Assessment View
  Learner Projection      Compiler View        Specification View
                                │
 ───────────────────────────────┼───────────────────────────────
                                ▼
         Universal Relationship Explorer (Cross-cutting)

The Studio runs entirely client-side via WebAssembly (Pyodide), providing real-time curriculum graph visualization, lesson progression storyboards, diagnostic engines, and learner mastery projections. For design specifications, refer to docs/studio.md.

Running the Studio IDE Locally

To run and test the Studio workspace on your local machine:

1. Prepare Python WASM Assets:

   uv run python scripts/pack_studio.py
   

2. Start Development Server:

   cd studio
   npm install
   npm run dev
   

   Open http://localhost:5173 in your browser.

Running the EduVis Agent Service (v1.3 Agentic Generation)

The agent/ package provides a stateful generation toolchain powered by LangGraph, Pydantic IR, and local model inference (Ollama / Qwen3.5), using EduVis Core as a deterministic validator.

1. Start Local Ollama Model (Homelab): Ensure Ollama is running locally with your target Qwen3.5 model:

   ollama run qwen3.5:9b
   

2. Start the Agent FastAPI Server: Run from the repository root directory (eduvis/):

   python -m agent.server
   # Or using uv
   uv run python -m agent.server
   

   This launches the Agent API server at http://localhost:8000.

3. Verify Server Health:

   curl http://localhost:8000/health
   

⚡ 5-Minute Agent Walkthrough: From Prompt to Validated Spec

Instead of asking an LLM to directly invent complex YAML specifications, the EduVis Agent separates proposals from deterministic compilation:

Suppose a teacher requests:

"Generate a 60-mark Secondary 1 Math practice paper on Negative Numbers with 10 questions and 20% diagnostic questions."

USER PROMPT
   │
   ▼
[Step 1: Intent Interpreter]   ──►  LLM maps prompt into structured GenerationIntent IR
   │
   ▼
[Step 2: Context Enricher]     ──►  EduVis Core injects prerequisite DAGs & concept rules
   │
   ▼
[Step 3: Candidate Generator]   ──►  Local LLM proposes candidate question items
   │
   ▼
[Step 4: EduVis Validator]     ──►  Deterministic check (Pass / Retry with Critique Feedback)
   │
   ▼
[Step 5: Paper Compiler]       ──►  Compiles final validated assessment_paper specification

Try It via REST API:

curl -X POST http://localhost:8000/api/generate/paper \
  -H "Content-Type: application/json" \
  -d '{
        "prompt": "Create a 60-mark Secondary 1 paper on negative numbers",
        "curriculum_yaml": "concepts:\n  - code: negative_numbers\n    name: Negative Numbers\nskills:\n  - code: order_integers\n    concept: negative_numbers\nmisconceptions: []\ndependencies: []",
        "target_marks": 60
      }'

Output:

{
  "status": "success",
  "intent": {
    "subject": "mathematics",
    "level": "Secondary 1",
    "topic": "negative_numbers",
    "total_marks": 60,
    "question_count": 10,
    "objective_distribution": { "AO1_recall": 0.4, "AO2_analysis": 0.4, "AO3_synthesis": 0.2 }
  },
  "candidate_questions": [...],
  "validation_errors": []
}

📖 For full state-diagrams, Pydantic IR field definitions, and architecture specs, read docs/agentic_generation.md.

Running the Standalone Live Editor (Static Playground)

For a quick, zero-install client-side playground, run the HTTP server from the repository root (or navigate into showcase/ directly):

Option A: From Repository Root

uv run python -m http.server 8000 --directory showcase

Option B: From showcase/ Directory

cd showcase
uv run python -m http.server 8000

Then open http://localhost:8000/editor.html in your browser.

[!NOTE] Unlike the React-based EduVis Studio, the Showcase Live Editor is a lightweight single-page tool. It runs in-browser by dynamically loading Pyodide and mounting your local Python files to sync engine and rendering logic changes on the fly.

Documentation Reference Map

EduVis is fully documented in the docs/ directory. Please refer to these resources for detailed specifications:

• docs/philosophy.md: Core design principles, primitive guidelines, and architecture philosophy.
• docs/architecture.md: The four decoupled layers (Knowledge, Assessment, Learner State, Presentation) and schema classifications.
• docs/compiler.md: The v1.0 compiler pipeline execution context, Planners, Assemblers, and CLI compile stages.
• docs/five_pillars.md: Full reference specification for the Five Pillars (Elements, Actions, Relationships, Placement, Progression) and geometric solid shapes.
• docs/engines.md: Detailed algorithms and mathematical formulas for Spaced Repetition (SM-2), Study Plan Priority Scoring, Remediation Trace Paths, and Greedy Paper Assembly.
• docs/ecosystem.md: Ecosystem framing, comparison matrix, and out-of-scope boundaries.
• docs/llm_system_prompt.md: Structured prompt vocabulary for injecting EduVis schema rules directly into LLM prompts.
• docs/studio.md: Design goals, multi-projection workspace, bidirectional editing, and local Pyodide/WASM compiler architecture for EduVis Studio.
• docs/agentic_generation.md: Agentic generation architecture, GenerationIntent IR, Educational Generation Graph, and FastAPI service for EduVis v1.3.
• docs/roadmap.md: Project priority guidelines, version timelines (v1.2 - v1.5), and strategic extensibility planning.

Project Structure

eduvis/ (repository root)
├── eduvis/                  ← Python package source (core models, validator, compiler, renderers)
├── agent/                   ← EduVis Agent service (LangGraph workflows, Pydantic IR, FastAPI server)
├── studio/                  ← EduVis Studio web application (React, TypeScript, Pyodide WASM)
├── docs/                    ← Conceptual, architectural, and studio documentation files
├── showcase/                ← Reference lesson files, feature catalogs, and learner state configurations
├── scripts/                 ← Build, packaging (pack_studio), and maintenance scripts
└── tests/                   ← Unit and integration test suite

License

This project is licensed under the Apache License 2.0. See LICENSE for details.