A technician asks which part belongs to a particular machine. AI finds an answer in a manual. A useful solution must also show the document, its version and the passage behind the answer. A fluent sentence alone is not enough.
Search does not automatically mean training a model
Document search first retrieves relevant passages and uses them to ground an answer. Training a custom model need not be part of the solution. OpenAI File Search supports search in uploaded files and file citations in responses.
The implementation must check that a citation actually supports the answer. Attaching a document name containing a similar word is not enough.
An old manual can produce a convincing wrong answer
With two manuals describing different procedures, the system must know which model and version each applies to. Label each document with its device type, date or version and validity status.
Start with one maintained collection. Check text and table readability in scanned PDFs. A misread part number can make its way into the answer.
Apply permissions before retrieval
A service employee may access manuals but not customer contracts. Search must only use documents the user is authorised to access. A text instruction saying “do not show contracts” cannot replace access control.
OpenAI states that API and business-product data is not used for model training by default. That does not determine who sees a document in your application, how long it is retained or which data you send. Those settings need separate decisions.
Test questions with no available answer
Include questions with a clear source, questions about obsolete manuals and questions outside the documentation. Add a case where the user lacks permission for the required file.
When grounding is missing, the system should say it cannot substantiate an answer. Check citation accuracy and usefulness, not just speed. Documents can contain instructions too; these must not change assistant rules or permissions.
What the first pilot should do
Choose one team and one document collection. Agree typical questions, acceptable answers and a reviewer. The pilot can retrieve and explain; changing orders or operating machines should be a separate step.
For service and safety procedures, users must be able to open the original manual and check the context. AI can shorten the search, but acting on an unverified answer is not the pilot’s goal.