Retrieval-augmented generation is often sold as “chat with your PDFs.” Field service cannot use that slogan. Some of the PDF is a customer-safe explanation. Some of it is an internal lockout procedure, a wholesale cost, or a troubleshooting tree you would never print on a fridge.
In ServiceBrief AI, company documents are first-class, but they are not one pile. This post is the visibility problem, which is more important than the embedding model on the box.
Upload is easy. Scope is the work.
Teams upload SOPs, manuals, warranties, quote templates, and general files. Ingest is asynchronous: chunk the file, embed the chunks, store them next to the company id. Status should be visible: processing, ready, failed. Re-ingest exists because people replace a price sheet.
Every chunk still needs a category and a visibility scope. Internal only can guide the technician-facing brief. Customer safe can inform a summary. Quote only should not wander into a diagnosis. Technician only is for the people in the van. If you skip scope, you will eventually watch a customer PDF recite an internal checklist.
Retrieval happens at extraction time
When a visit is processed, the pipeline is not “stuff the whole knowledge base into the prompt.” It classifies the kind of work, retrieves a small set of nearby chunks, and asks extraction to respect those rules. Required SOP fields that never appeared in the transcript can be flagged instead of filled with a guess.
The UI should show sources. A citation chip is not decoration. It is how a manager decides whether the model used the right manual. If you cannot show why a field was flagged, the flag will be ignored.
Customer language is a different output
The internal brief and the customer summary are not the same paragraph with the jargon deleted by hope. Customer output is generated from customer-safe material and the facts of the visit. Internal notes stay internal. That is cheaper than a regex trying to scrub the damage later.
Memory is not policy
Approved jobs can become examples. That is useful: this company always asks for a model number on a boiler. It is dangerous if examples outrank the written SOP. ServiceBrief keeps a hierarchy: official documents and the price book first, approved memory second, drafts never. You can turn memory off. You should never let a draft teach the next draft.
Tenancy is not optional
Vectors without an organization filter are a data breach waiting for a curious prompt. Every search, ingest, and generation path has to be company-scoped on the server, not hidden in the UI. Role-based access still applies on top.
If you are evaluating “we added RAG” on a roadmap, ask three questions. Can a customer summary cite an internal doc? Can company A retrieve company B? Can a quote price appear if it was not in the book? If any answer is unclear, the demo is not a product yet.