AEROSPEC-AI

AeroSpec AI is a demonstration RAG application for asking questions about a small corpus of technical connector documentation. It returns focused answers with document and page citations, exposes the source PDFs in the frontend, and includes deterministic unit conversion and a measured evaluation view. The current UI, corpus, and query workflow are primarily English. This is a portfolio prototype rather than a certified engineering system.

The architecture consists of a React/Vite frontend, a FastAPI backend, PDF retrieval with a Chroma vector store, and grounded OpenAI generation with validated citations. The repository includes backend and frontend code, five sample technical PDFs in data/sample_docs, evaluation benchmarks and scripts, and documentation. The frontend Documents page opens the same five PDFs stored in data/sample_docs; the files are not duplicated under frontend.

Authentication uses a shared demo username and bcrypt password hash for reviewer access. Successful login creates an opaque bearer session that expires after four hours. Authenticated reviewers share a quota of 20 successful /ask requests per four-hour UTC-aligned window. Intentional abstentions consume quota; technical backend failures do not. Authentication sessions and quota counters are in memory, reset on process restart, and are not shared across instances.

The system can be run locally with Python 3.12 and Node.js 24 or newer, or via Docker. The static frontend is deployed to Vercel and the Dockerized backend to Render. The benchmark is small and manually curated. Retrieval quality is measured separately, but evaluation results apply only to this bounded dataset. The system is designed to reduce and measure unsupported answers, not guarantee their absence.