On-prem only
Models run on the leased appliance inside your network. Tokens for your corpus are not sent to a public cloud for inference.
Security
Isolation, walls, logs, open storage.
The same controls serve OEM pricing/config IP and firm privilege concerns. Architecture first. Policy second.
Payload filters
Retrieval is filtered by product line / region / role — or matter / client / ethical wall — before the model sees a chunk.
Query log
Every query and every retrieved chunk is logged. That log is the operational and ethics artifact.
Open & auditable
Plant or firm corpus on TrueNAS SCALE (open ZFS) over fast NFS. Not a closed NAS OS you cannot inspect for phone-home.
Drive layout
Spark internal drive ≠ your files.
On the Spark
- OS
- Model weights
On TrueNAS
- Your corpus and extracts
- Vector store data
- ZFS snapshots before updates
- RAID-Z redundancy
The inference device does not store your files. The Spark or Mac Studio holds the operating system and the model weights. The files stay on the storage system beside it.
Hardware & pricing
Same chassis for both doors. We maintain the box; you keep your own backups.