anthonym21/coda-gqa-l-neural-db-demo

CoDA-GQA-L: Stateful Neural Database

This Space shows the "stateful neural database" workflow:

1. 1.ingest a document into a fixed-size bounded-attention state,
2. 2.save that neural state,
3. 3.reload it later,
4. 4.query it without re-reading the original document.

Demo profile

• —Built for public play on ZeroGPU
• —Bundled example documents can be prepared into reusable saved states
• —Custom ingestion is intentionally capped to short documents
• —Suggested questions are attached to the prepared sample states

Important note about storage

This Space is configured to work without paid persistent storage. That keeps it cheap, but it also means the local state library is ephemeral. If the Space restarts, the bundled examples may need to be prepared again.

If you later want prepared states to survive restarts, switch the library path to persistent storage or a Hub-backed dataset store.

Optional launch polish

Set DEMO_MEDIA_URL or add assets/walkthrough.mp4 to show a short ingest -> save -> reload -> query clip at the top of the app.