Surgical Graph RAG plug-and-chug spatial scene reasoning

Portfolio demo

Plug in a scene. Ask spatial questions. Get graph-grounded answers.

The same typed-relation pipeline runs on a validated room and an unseen apartment import. It answers through a compact scene graph instead of handing every object to a model.

2Replica scenes in the showcase
346objects in the apartment plug-and-chug run
3wall-mounted ATTACHED_TO answers
G1-G9Phase 7 gate passed locally

Unseen scene

Replica apartment_0

Scene question visualization

Selected answer path

ATTACHED_TO narrows broad wall contact to mounted objects.

1. Import a scene

Replica object boxes and structural surfaces become a typed scene graph. Raw meshes stay out of the website.

2. Compile the question

"Attached to the wall" becomes ATTACHED_TO(?x, wall). Older Phase 6 artifacts keep their frozen deferral semantics.

3. Return evidence

The answer cites graph object ids and relation-specific edges, with caveats separated from benchmark claims.

Static deploy path: GitHub Pages publishes from /docs. The rendered PNG assets live in docs/assets/; source fixtures and reports stay in the repo. Apartment rows are plausibility labels, not a ground-truth accuracy benchmark.

View demo runner