Gap Inc.: a first internal GPT with retrieval, access control and an evaluation harness

A general-purpose model knows the public internet. It does not know a retailer's policies, its supplier contracts, its merchandising procedures or which of its business units is allowed to read which of them. Gap Inc. wanted its first internal GPT to close that gap without sending company documents somewhere they should not go.

The engagement covered a private in-house model, retrieval over internal documents, access control per business unit, and the evaluation harness that decided whether it was good enough to launch.

The model answers from company documents through retrieval, with sources shown, and the permissions of the person asking are carried through from the systems the documents live in. A question from one business unit is answered only from what that unit may see. The model runs inside the company's boundary, behind one routing layer that logs every call.

Before it reached users it was measured against an evaluation harness built from real questions and expected answers, and the same harness runs after every change to the model, the prompts or the documents. A change that makes the answers worse is caught before anyone sees it.

Because the knowledge, the permissions and the evaluation sit outside the model, the model itself can be swapped as better ones arrive. The first internal GPT is designed to be the first of several, not a single build that ages with its model.