Glossary

What is Enterprise RAG?

Enterprise RAG refers to retrieval-augmented generation run inside an organization, against its own documents and records, under its own access rules. The retrieval mechanism is the same as any other RAG system. What differs is everything around it: who is allowed to see which passage, how often the source changes, and whether an answer given last quarter can still be accounted for.

Three constraints separate it from a demo. Permissions have to survive retrieval, so a passage a user cannot open in the source system must not reach them through the model either. The corpus keeps moving, so an index built last month answers from a version of the business that no longer exists. And evaluation cannot be a spot check, because the same question asked twice should not return two different answers without an explanation.

There is an earlier constraint that appears before any of this. Some of the records an organization would most want to index cannot be sent out in their original form at all, so they are left out of the corpus and the system answers from whatever remains. Masking the sensitive fields lets the documents through, but it also removes the table structure and the relationships between business identifiers that made those records worth retrieving. Deciding what enters the index, in what form, is part of designing enterprise RAG rather than a step that comes after it.

Frequently asked questions

What is enterprise RAG?

Retrieval-augmented generation run inside an organization, over its own documents and records, under its own access rules. The retrieval mechanism is standard; the permissions, freshness and audit requirements around it are what make it different.

How is enterprise RAG different from regular RAG?

Regular RAG is judged on whether the answer is good. Enterprise RAG is also judged on whether the person was allowed to see the source, whether the index reflects current data, and whether the same question returns a consistent answer later.

Why do enterprise RAG pilots stall after the demo?

Usually not on retrieval quality. They stall on permissions that do not carry through to the model, on a corpus that has moved since indexing, and on evaluation that cannot be repeated. These only appear once real users and real documents are involved.

What about documents that cannot leave the environment?

They are often excluded from the index, so the system answers from a partial corpus. Masking lets them through but strips the structure and identifier relationships that made them useful. What enters the index, and in what form, is a design decision to make up front.