Compliance training only works if people actually understand the rules, but the source material is often long, dense, and hard to retain. This tool turns those documents into quick quizzes people can use to check what they know and reinforce what they've read. It pairs a language model with a searchable library built from your own documents, so every question and answer points back to the exact text it came from.
I built it around regulatory and legal documents, but the document handling is general enough to work with any written policy or procedure where getting the details right matters.
PythonFlaskOllamaChromaDBRAG
Problem
Most organizations have thick policy and procedure documents that people are expected to know,
but few read them closely. Reading the material once doesn't tell you whether someone actually understood it. The usual fix is a quiz someone writes by hand, which takes time to build, goes out of date as the documents change, and tends to test general knowledge rather than the specific rules that apply.
Approach
I built a pipeline that connects a language model to a searchable library of the actual documents.
When you ask for a quiz on a topic, it pulls the relevant passages, hands them to the model,
and asks it to write multiple-choice questions from that text only.
The documents are broken down in a way that keeps their structure (e.g., sections, headings, hierarchy)
so the model gets meaningful context instead of a wall of text,
and every question it produces comes with a reference back to the source file
and the passage the answer came from.
Outcome
The result is a way to turn dense documents into a checkable quiz in minutes instead of writing one by hand.
Since every answer links back to its source, teams get an easy way to confirm people understand the material,
and reviewers get something they can verify against the original material.
I built it around regulatory documents, but it works with any written policy or procedure.
Key Capabilities
Built from your own documents: Upload PDFs, Word files, text, or HTML and the tool builds a searchable library from them — the quiz is drawn from what's actually in those files, not from general knowledge.
Structure-aware document handling: Documents are broken down in a way that keeps sections, headings, and hierarchy intact, so the model gets meaningful context instead of a wall of text (with specific handling for legal and regulatory formats like the CFR and U.S. Code).
Grounded questions, not guesswork: The language model writes multiple-choice questions from the passages it retrieved, rather than pulling them from memory.
Every answer points to its source: Each question includes the source file and the passage the answer came from, shown alongside the results, so you can check it against the original.
Interactive and ready to use: Take the quiz in the browser, get scored, and review what you got right or wrong — with the supporting text for each question.