OpExams

Accepts uploaded documents (PDFs, text files, Word docs) and uses prompt-based LLM generation to synthesize exam questions that directly reference and test comprehension of the source material. The system likely parses document content, chunks it into semantic segments, and feeds those segments to a generative model with a question-generation prompt template that specifies format, difficulty, and question type constraints.

Accepts a topic name or brief description and generates exam questions using the LLM's parametric knowledge without requiring uploaded documents. The system constructs a prompt that specifies the topic, desired question count, format, and difficulty level, then calls a generative model to produce questions. This approach relies on the model's training data rather than user-provided context.

Generates multiple-choice questions with configurable parameters: number of answer options (typically 3-5), difficulty level, and answer distribution. The system likely uses prompt templates that specify the desired format and constraints, then post-processes LLM output to ensure correct option count and valid answer key generation. May include logic to avoid obvious patterns (e.g., 'C' as correct answer for every question).

Generates open-ended short-answer questions (as opposed to multiple-choice) that require students to provide brief written responses. The system uses prompt templates that specify answer length constraints and expected response format, then generates questions with model-provided answer keys or rubrics. May include logic to generate acceptable answer variations to support flexible grading.

Exports generated questions in multiple formats (PDF, DOCX, potentially others) suitable for printing or learning management system (LMS) import. The system likely uses templating engines (e.g., Jinja2, Handlebars) to format questions into document structures, then leverages libraries like python-docx or similar to generate output files. May support customization of document layout, branding, and metadata.

Allows users to specify desired difficulty levels (e.g., easy, medium, hard, or numeric scale) for generated questions, and the system adjusts question complexity, vocabulary, and cognitive demand accordingly. Implementation likely uses prompt engineering with difficulty descriptors and examples, potentially with post-hoc validation to ensure generated questions match the specified difficulty. May track difficulty metadata in question objects.

Supports generating large numbers of questions in a single operation, potentially with progress tracking and asynchronous processing. The system likely queues generation requests, processes them in batches to optimize API calls to the underlying LLM, and provides status updates or completion notifications. May include rate-limiting and quota management for freemium tiers.

Provides a user interface for educators to manually edit, refine, or regenerate individual questions after initial generation. The system likely stores generated questions in an editable format, allows inline editing of question text and answer options, and may provide regeneration options to replace specific questions or options. May include version history or undo/redo functionality.

Detects and flags duplicate or highly similar questions within a generated set, potentially using semantic similarity metrics (e.g., cosine similarity of embeddings) or string-based matching. The system may automatically remove duplicates or present them to the user for manual review. Implementation likely uses embedding-based similarity or fuzzy string matching to identify near-duplicates that simple string comparison would miss.