Quick Start
Get started with VBAgent in 5 minutes!
1. Initialize Workspace
This creates .vbagent.json with your preferences (subject, models, etc.).
2. Process Your First Image
This will: - Classify the question type - Extract LaTeX - Generate TikZ diagrams (if present) - Assess difficulty - Save everything with metadata
Output:
agentic/
├── scans/problem_1.tex (with metadata)
├── classifications/problem_1.json
├── diagrams/problem_1.json
├── difficulty/problem_1.json
└── tikz/problem_1.tex
3. Author a syllabus-scoped problem
Inspect the plan without making API calls:
vbagent author preflight \
--exam jee_main \
--subject physics \
--chapter kinematics \
--topic "Projectile Motion" \
--type mcq_sc \
--count 4
Then replace preflight with run and add --output agentic/authoring.
Only candidates that pass the independent solution, answer, syllabus,
difficulty, compile, review, and novelty gates are accepted.
4. Talk to the authoring workflow
Use the built-in transparent CLI host:
It shows each controller and tool input/output and asks locally before model execution begins. Or connect an external MCP host through stdio:
Then ask the host to plan a syllabus-scoped batch. It will inspect the catalog, show the resolved distribution and estimated model calls, request confirmation, and dispatch a durable worker. Closing the conversation does not discard the run.
5. Create a DPP Set
Creates a balanced 10-question practice set from your question bank.
Common Workflows
Scan Multiple Images
Generate Variants
vbagent variant --parent-spec-id ACCEPTED_SPEC_ID --type numerical --count 3 \
--output agentic/authoring
Variants require an accepted canonical parent and pass the same full acceptance pipeline as original problems.
Database Management
# Initialize database
vbagent db init ./questions
# Query questions
vbagent db query --topic Mechanics --difficulty medium