VaultExam AI: Mock Tests from Your Notes, Without Uploading Them Anywhere

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

VaultExam AI is a 100% local, privacy-first tool that turns a chapter (In any format) into a ready-to-print question paper + answer key and lets you ask questions of the same document with page-cited answers.

Who would pay for it: independent coaching institutes, test-prep tutors, university professors and corporate trainers. Their curated notes and question banks are their intellectual property, and they don't want to upload them to a cloud API. They also spend hours every week typing up mock tests, DPPs and Set A/B variations by hand.

What it does:

• Upload Notes in any format. It is chunked and embedded locally, and the file is deleted from disk straight after ingestion.
• Ask the document: answers come only from your notes, with Page citations.
• Generate an exam: choose the institute name, exam type (Coaching Mock Test, University Mid-Sem, Corporate SOP Quiz), difficulty (Recall, Application, Analytical/Edge Cases) and the number of MCQs, short answers and long answers.
• Set A/B/C: each set draws on different parts of the document and avoids repeating earlier questions.
• Answer key included: MCQ answers with explanations, model answers for short and long questions, marking-scheme key points, and the source page for every question so a teacher can verify it.
• Print or save as PDF straight from the browser.

Demo

Code

https://github.com/ronitsolanki-01/VaultExam.git

How I Built It

Everything runs on the user's own machine. No cloud calls, no per-token fees.

• Backend: Java 21, Spring Boot 4.1.1 and Spring AI 2.0.1
• LLM: an open-weight model (qwen3.5:4b) served by local Ollama
• Embeddings: nomic-embed-text via Ollama
• Vector store: PostgreSQL with PGVector (pgvector/pgvector:pg16 via Docker Compose), with metadata filtering by document_id
• Ingestion: Spring AI's TikaDocumentReader and PagePdfDocumentReader keeps page number on every chunk, then TokenTextSplitter chunks it
• Frontend: one Next.js page (TypeScript, Tailwind), with window.print() for output

How the exam generator works: it retrieves a sample of chunks from the document, asks the model for MCQs, short answers and long answers in separate structured-output calls, then validates each result (question count, exactly four options, valid answer index) and retries if the local model returns bad JSON. Each set uses a different seeded sample of the document, plus a "don't repeat these questions" list.

Things that bit me on a small local model:

• Reasoning models like qwen3.5 can spend their whole token budget "thinking" and return an empty answer, which breaks JSON parsing. I fixed it by disabling thinking in the Ollama options.
• Ollama's default context window silently truncates long RAG prompts, so I set a larger num-ctx and tuned how many chunks each call retrieves.
• Small models sometimes return a bare JSON array instead of an object, so the parsing and validation code handles that and retries.

Why Does Open Innovation Matter?

A closed API would have made this product impossible to sell to its target customers. A coaching institute's notes are its business, and "send your question bank to a third party" is a dealbreaker.

Open-weight models let the whole pipeline run inside the customer's own machine. That also changes the economics: generating a hundred papers costs electricity, not tokens, so a tutor can regenerate Set B at 11pm without watching a bill. And because the stack is open, institutes can swap the model for a bigger one as their hardware allows.