What to decide before you ground answers in company docs

The job is an answer that can point at a file you already own. A policy, a price list, a handover note — retrieved, cited, refused when the pile is silent.

People ask to “chat with our Drive.” That is a demo. The job is which corpus is in scope, who updates it, and what the desk does when the model is unsure.

The questions that decide the build

Which folders are allowed this month. Whether a page in Arabic and a page in English are the same fact. Who may add a PDF. Where the citation has to appear — Slack, an internal search box, a CRM sidebar.

If the latest price lives in someone’s head, retrieval will only recite last quarter. We will say that.

If the files are real and the stall is search time, AI text generation can sit on a narrow RAG slice: retrieve, answer, show the passage, stop when nothing matches. We will not call that a colleague.

This is not policy form annotation. That job names fields and checkboxes on a government PDF so another system can fill them. Grounded Q&A does not emit a form schema. It is also not a public support bot. A customer channel has a different tone, a different escalation, and a different risk.

When we tell you not to index yet

If two departments publish conflicting PDFs, software should not pick a winner. If the pile is scans with no text layer, the first slice is OCR and a reviewer, not a chat box.

Kuwait operators often mix Arabic circulars with English vendor notes. That changes chunking. It does not change the sequence: name the corpus, name the refuse rule, then decide if consultancy becomes a small integration.

Bring the three files people actually open on a Monday. Book thirty minutes or write hello@jamilglobal.com.