Back to work
Repository Analysis · Retrieval · Trust checks

Codebase Investigator

A repository analysis tool for public GitHub projects. It answers plain-English questions using retrieved source evidence, cited files and line ranges, session history, and a separate audit pass.

PythonFastAPIOpenAI embeddingsTF-IDFMMR rerankingCLI and browser UI
Open live app View source repository View public case-study repo
Codebase Investigator browser UI screenshot

Problem

Developers often need to understand unfamiliar repositories quickly. The useful version needs citations and checks that tie claims back to real code instead of vague summaries.

What I built

  • Public GitHub repository ingestion and indexing.
  • Hybrid retrieval using exact matching, embeddings, review scan, duplicate merge, and MMR reranking.
  • Answer generation grounded in selected evidence.
  • File and line citations with programmatic validation.
  • Separate model audit for unsupported claims, risky fixes, contradictions, and reasoning gaps.
  • CLI, browser UI, API endpoint, and session memory for follow-up questions.

Architecture

GitHub URL to clone and index, then source chunks move through hybrid retrieval, reranking, answer synthesis, citation validation, trust check, and session memory.

Why it matters

The important part is traceability: answers point back to files, line ranges, retrieved evidence, and a separate check for unsupported claims.