Collect answers, not every available file
Connecting an assistant to a large folder can feel like progress, but volume is not the same as usefulness. Old price lists, draft procedures and duplicate presentations create conflicting evidence. Begin with the questions your customers or staff actually ask. Find the approved material that answers those questions and identify the gaps before importing everything.
For a Singapore service team, the first collection might include current service descriptions, booking procedures and escalation contacts. Keep the scope small enough that someone can review it. A carefully chosen collection of reliable documents is a stronger starting point than an archive that mixes current policy with forgotten experiments and informal notes.
Give every document an owner and a status
A document should have a person or team responsible for its accuracy. Record whether it is approved, when it was last reviewed and which audience may access it. These details help the retrieval system distinguish authoritative information from a draft. They also give staff a clear route for correcting the source when an answer is wrong.
Agree what happens when a document expires or is replaced. Removing an old file from a folder may not immediately remove its indexed content from a search service. The update process should cover the stored document, extracted text and search index. Where information has been withdrawn, make sure cached answers do not continue presenting it as current guidance.
Write documents so the answer survives extraction
Use descriptive headings and keep related instructions together. Explain abbreviations and avoid referring to an unnamed attachment as the only source of an important rule. Tables, scanned pages and images may require special handling. Review the extracted text rather than assuming that a visually tidy PDF will become a clean and complete search document.
Keep conditions alongside the rule they qualify. A cancellation policy without its exceptions can produce misleading answers. When information is divided into smaller sections for retrieval, preserve enough context to identify the document, its scope and the relevant qualification. The right unit of information is the smallest one that still communicates the complete meaning.
Treat citations as something to evaluate
A citation is useful only when it supports the answer. Ask the assistant to point to the relevant source and let users open it. Then review whether the cited passage actually establishes the claim. A model can attach a real document to a statement that the document never makes, so a clickable reference is not a guarantee of correctness.
Build a small set of realistic questions, including requests that the collection cannot answer. Include cases where two documents appear similar but apply to different services or teams. Evaluate both retrieval and the final response. If the right document is found but the explanation changes its meaning, the problem is not solved by adding more files.
Turn a vague policy into a usable reference
Consider a fictional service policy that says customers can reschedule with sufficient notice. That sentence leaves the assistant to guess what sufficient means, which services it applies to and whether exceptions exist. Before indexing the document, ask the policy owner to supply the missing rules. A knowledge base cannot make an ambiguous business decision more authoritative by repeating it in polished language.
A clearer reference names the service, the notice period approved by the business, the channel customers should use and the team responsible for exceptions. It also identifies its effective date and the policy it replaces. Keep any qualification next to the rule it changes. The goal is not to make every document longer; it is to make each answer complete enough that a reader can apply it correctly.
Use a consistent document layout where it helps: purpose, applicable audience, rules, exceptions and escalation contact. Avoid filling the collection with repeated copies of the same rule. If several pages describe the same process, establish which one is authoritative and have the others refer to it. Contradictory duplication creates maintenance work and makes retrieval harder to evaluate.
Build an answer key before evaluating the assistant
Choose a manageable starter set of questions taken from actual work, with personal details removed where needed. For each question, record the expected source, the points a correct answer must include and the conditions that require a handoff. Include several questions the collection cannot answer. Those examples test whether the assistant recognises a gap rather than presenting an unrelated document as evidence.
For the rescheduling policy, test a straightforward change request, a request inside an exception window and a question about a different service. Add a question that mentions an old policy name. The expected behaviour may differ across these examples: answer directly, explain the relevant qualification, or route the request to the policy owner. Write that expectation before looking at the generated response.
Compare the retrieved passage with the answer key, then compare the final response with the passage. This separates two common problems. The search may find the wrong document, or the model may find the right document and omit a condition. Keep a few examples for checking after changes so that improving one answer does not quietly break another.
Diagnose missing answers before adding more content
When an answer fails, ask where the evidence was lost. Was the information absent from the source? Did extraction miss a table or scanned page? Did retrieval select the wrong service? Did the response change the meaning of the retrieved text? Each cause needs a different repair. Uploading another folder is unlikely to fix a poor scan or an explanation that ignores an exception.
Use a small failure log with the question, the approved expected answer, the observed problem and the owner of the fix. Keep only the information needed for investigation. Group similar failures so that a single source improvement can solve several recurring questions. Check the repaired examples again before declaring the issue closed.
When a policy changes, follow its path through publication, extraction and indexing. Confirm that the updated answer is retrievable and that a withdrawn rule no longer appears in a routine response. Include access checks in this verification. An accurate answer still creates a problem if it reveals a document to someone who should not be able to read it.
Make upkeep an ordinary team habit
Set a review rhythm that fits the rate at which the information changes. A stable operating procedure may need less frequent attention than delivery guidance or a changing product catalogue. Let staff report missing answers and confusing wording. Use those reports to improve the approved material instead of building an endless collection of one-off prompt exceptions.
Measure whether people can find a useful answer and complete their task. Search volume alone does not tell you whether the system helped. Keep permission boundaries in place as the collection grows, and document how to pause retrieval from a problematic source. A knowledge assistant stays useful through disciplined information ownership, not through a one-time import.
