AeroGraph — Cached Demo

Hybrid graph-augmented retrieval over 2,000 NASA ASRS safety reports

This Space serves 10 precomputed showcase queries from our benchmark run. For the full live system (6 retrievers, dual-judge eval, graph explorer): see GitHub and app_full.py.

Paper finding: retrieval-precision rankings do not replicate on a disjoint held-out corpus (Kendall τ ≈ −0.07) — we name this Extractive Oracle Circularity and propose held-out swap as a falsifiability test for graph-RAG benchmarks.

Presets: