Documentation

How CodeSage works, from repository ingestion to scored examination.

How it works

Connect your repository

Paste any public or private GitHub repository URL. CodeSage clones it, parses the codebase with Tree-sitter, and embeds every code chunk via NV-Embed-QA into a Qdrant vector store.

Choose an examination mode

Pick from Viva Voce (academic defense), Technical Interview (FAANG-style), or Code Review (senior engineer feedback). Each mode calibrates the examiner's focus and scoring rubric.

Answer questions

CodeSage generates questions that reference your actual code — specific file paths, function names, and line numbers. Answer as many as you want, skip freely, end when ready.

Get scored and reviewed

The Nemotron-340B reward model grades each answer on accuracy, depth, and awareness. You get a dimension breakdown, per-question feedback, and an overall score out of 100.

Study and improve

A personalized study guide maps your weak areas to specific concepts and code sections worth reviewing. Track your scores over multiple sessions to see real improvement.

AI Models

CodeSage uses 5 specialized NVIDIA NIM models, each optimized for its task in the examination pipeline.

NV-Embed-QAEmbedding

Converts code chunks into semantic vectors for retrieval. Optimized for question-answering tasks.

NV-Rerank-QAReranking

Reranks retrieved code chunks by relevance to the question. Ensures the examiner sees the most relevant context.

Llama-3.3-70B-InstructExaminer

Generates viva-ready questions from code context. Probes architecture decisions, trade-offs, and implementation details.

Nemotron-340B-RewardScorer

Evaluates answers against rubrics on accuracy, depth, and code awareness. Produces objective, reproducible scores.

Llama-Guard-3-8BSafety

Runs alongside the examiner to ensure all generated questions stay within appropriate academic scope.

Examination Modes

Viva Voce

Simulates academic viva panels. Questions probe your rationale, alternatives considered, and design intent. Ideal for thesis defenses, project evaluations, and course assessments. Difficulty ranges from easy recall to expert-level architectural reasoning.

Technical Interview

Replicates FAANG-style system design and code review rounds. Surfaces scalability gaps, edge cases, and algorithmic trade-offs in your actual codebase. Designed for interview preparation and self-assessment against industry standards.

Code Review

Generates senior-engineer-level feedback on architecture, coupling, test coverage, and documentation. Questions are framed as issues you should be able to explain or justify. Useful for pre-submission self-review and improving code quality.

Scoring

Each answer is graded by Nemotron-340B on three dimensions:

Accuracy (0–100)

Is the answer factually correct? Does it demonstrate understanding of the code?

Depth (0–100)

Does the answer go beyond surface-level explanation? Are trade-offs discussed?

Code Awareness (0–100)

Does the answer reference specific files, functions, and line numbers from the codebase?

The overall score is a weighted average of these three dimensions. Scores above 80 indicate strong understanding. Scores below 60 suggest areas that need review.