The ingenuity
The visible map is only the last mile.
The hard part was making a large, mixed-speed homelab behave like one dependable research instrument. Work is broken into bounded claims, leased to whichever node is ready, and returned with receipts. If a machine disappears, its lease expires and the tract safely re-enters the water.
Retrieval stays separate from prose. Structured facts, source permissions, embeddings, and generated synopses remain distinguishable, so an answer can be ranked without pretending every derived sentence is raw data.
- 01Normalize tract facts and attach their provenance before any model sees them.
- 02Generate local summaries on mismatched workers through a pull-based queue.
- 03Embed accepted text into the vector store, with license fences carried alongside it.
- 04Resolve a map question into geographic filters, retrieval results, and a grounded response.


