How it works
One database, fenced by the database. One retrieval path, inside the fence. Any model on top.
The five-minute technical walkthrough for the people who have to sign off: how three tiers' documents live in one Postgres without ever meeting, how a dealer's question is answered only from what that dealer may see, and why swapping the AI model changes nothing about the product.
Ready to put your library on the platform, or want an engineer-to-engineer review of the design? Both start the same way.
The five things the video says, in writing
- 1
Tagged
Every chunk, KPI, and message row carries the id of the company that owns it — a column on the row, not a folder or an app-side filter.
- 2
Fenced
Postgres enforces row-level security on every table. A request runs as one company; other companies' rows do not exist to it. A guessed id returns 404.
- 3
Approved
Distributor and manufacturer libraries open only through the approval chain, recomputed on every question and never copied. Revoke a dealer and it is gone on their next question.
- 4
Retrieved
The model never searches the database. Hybrid vector + keyword search runs over the permitted rows only, and the top chunks come back with their owner and filename.
- 5
Answered
The prompt is the copilot's methodology, the shop's KPIs, and only what retrieval returned. The dealer's own model — Claude, OpenAI, or Grok — writes the answer and cites every claim.
Put your library where your dealers will actually read it.
Tell us about your network and we will take it from there — including a design review with your IT or security lead if you want one.