02, Private AI
Harbour
Document chat where the model and the index stay on hardware you control. Policies, contracts, and runbooks, without a default path to a vendor API.
Decision. Can someone answer from our own documents without the documents leaving the building.
How it is shaped
01
Estate
Map where documents live, who may see them, and which network the model is allowed to touch.
02
Index
Chunk, embed, and retrieve inside that estate. Answers that cannot cite do not ship.
03
Gate
A vendor model is a deliberate exception, logged, not the quiet default.
What runs
- Practice corpus stays on our own hardware.
- Refusal when retrieval is weak, instead of a fluent guess.
- You keep the weights path, the index, and the logs.
Stack, kept short
- On premises inference
- Local retrieval
- Citation
- Access logs
Use this when confidentiality is the product, not a footnote in a cloud demo.
Borrow this shape