Preparing company documents for an AI knowledge assistant

Imagine a support team being asked where a customer should request an exchange for an item bought through a distributor. The answer might depend on the service policy, partner instructions and a recent process update. Having all three files does not tell the system which sales channel applies or which passage the update replaces.

Write down representative questions with the context needed to answer them: who is asking, which product or service is involved, and when the relevant event occurred. Ask the process owner to identify an acceptable answer and the evidence supporting it. If the team cannot agree, resolve that uncertainty before treating the documents as a reliable knowledge base.

Include ambiguous questions, questions the documents cannot answer, and requests for information the user is not allowed to see. These reveal whether the system can ask for clarification or decline to answer appropriately.

A simple spreadsheet is enough to begin. Its purpose is to identify the authoritative source for each subject and the person who can confirm it, not to move every file into a new platform immediately.

When a revised manual arrives, do not simply add it beside the old one as an equally valid source. Record whether the earlier version has been fully replaced, withdrawn, or retained for a defined category of older cases.

A PDF that is readable on screen may contain only scanned images. A clear visual table can lose the relationship between column headings and values during extraction. Process a representative sample and inspect the result before loading the whole collection.

For long manuals, keep section context and exceptions connected to the relevant instruction. Separating “a team leader may approve” from the condition that limits that authority can produce an answer that sounds confident but applies the rule incorrectly.

Treat translated versions carefully too. If the Thai manual has been updated while its English equivalent is out of date, record that difference or restrict answers to a version the owner has confirmed. A shared title does not make two files interchangeable.

Retrieval-augmented generation, or RAG, retrieves relevant material and provides it to a model as context. It still depends on well-prepared sources and appropriate access controls; it does not resolve contradictory policies by itself. Microsoft Learn explains the RAG approach.

“Answer like a person” is a difficult acceptance criterion. Instead, specify the useful parts of an answer: a conclusion, important conditions, supporting references and any missing information. A reference should point to material that actually supports the statement, rather than a long document containing similar words.

In the illustrative distributor example, a suitable answer might explain that the contact route depends on the distributor category, ask where the purchase was made, and cite the manual section used to decide. If that distributor is not covered, the system should hand over the question and the evidence it found instead of inventing a policy.

The handover should explain the unresolved point. “Two active sources disagree” or “this product version is not covered” gives the employee something to fix in the knowledge base. A generic failure message leaves the same question waiting for another manual answer next time.

An incorrect answer can result from failing to retrieve the right document or from misinterpreting a document that was retrieved correctly. The remedies differ. The first may require better source preparation or retrieval; the second may require clearer conditions and response instructions.

Have a subject expert assess the prepared question set. Record whether the reference supports the answer, whether any statement lacks evidence, and whether clarification or a handover was needed. Test each supported language independently. Passing an English test does not establish that Thai answers are reliable, or vice versa.

Use accounts with different permissions to check access. Material a person cannot read should be excluded before it becomes answer context. Hiding the citation afterwards does not prevent confidential information from appearing in the response itself.

Decide who approves a replacement manual, when it enters the system, and how the team checks that the obsolete version is no longer being used. There should also be a way to withdraw an incorrect source without waiting for a full application change.

Capture unanswered questions and staff corrections so they can improve the sources and future tests. Keep only the information needed to investigate the problem. Measure supported answer quality and the review effort required, rather than simply counting how many messages the system answers.

The preparation should leave four usable deliverables: real questions, a source register, rules for answering and handing over, and named content owners. Together, they make the difference between a demonstration that can read files and a knowledge service the team can maintain.

If your sources include scans or complex tables, read about AI document extraction and human review to protect the meaning of information before it enters the knowledge base.