Build an AI Assistant Knowledge Base From Trusted Sources

Inventory sources before adding them

A useful assistant knowledge base is a maintained set of authoritative material, not a pile of every document the company can find. Begin with the questions in the assistant’s intended scope. For each question, identify which approved source currently answers it and whether that source is complete enough for a customer-facing reply.

Create a team inventory with the source name, owner, audience, authoritative location, last review date, and a trigger for review. Triggers might include a price update, a policy revision, a service change, or a new location. Keep this register in the team's shared documentation and assign an owner to review it when one of those triggers occurs. It helps a reviewer decide whether a document belongs in the assistant’s material and who can approve changes.

Rank sources by authority and currentness. A signed policy or maintained help guide should take precedence over an old campaign file or a colleague’s notes. Read the document itself rather than relying on its filename. Look for date ranges, exceptions, missing definitions, and statements that apply only to one audience. If two credible sources disagree, ask the policy owner to choose the canonical version before including either as if it were settled.

Curate for one assistant responsibility

Choose material that directly supports the role. A support assistant answering published service questions may need a current service guide and relevant policy, not every internal process document. Exclude drafts, expired copies, irrelevant customer data, and content the organization is not authorized to expose. If a source contains sensitive or internal-only sections, prepare an approved customer-facing version rather than assuming the assistant will ignore them correctly.

Resonoon’s knowledge editor supports PDF and Markdown uploads. Google Docs knowledge is handled through Google Drive Picker, where connecting the account alone imports nothing; a non-empty selection must be confirmed to start import and save it. Choose among the source types available in the editor. If the original material is stored elsewhere, ask its owner to prepare a current, approved PDF or Markdown copy for upload.

Use clear filenames and keep one authoritative version easy to identify. Avoid uploading near-duplicate files with uncertain status. A duplicate can preserve conflicting wording even after the canonical document is corrected. Before removing a candidate, check whether any important customer question depends on it and whether its authoritative replacement is ready.

Keep source ownership explicit

Assign an owner who can confirm the facts and a reviewer who can check how they affect the assistant’s answers. The source owner should define the audience and approve customer-facing language. Record the review date and the event that requires another review. For frequently changing policies, a calendar date alone may not be enough; a price change should trigger review immediately.

When a source changes, compare the previous and current wording around test questions. Recheck the answers that rely on the changed section and any boundary cases. Do not promise instant propagation after editing a connected document. Verify the visible source status and relevant answers before assuming customers will receive the new wording.

Make the inventory useful in practice

Keep the initial inventory modest. A spreadsheet or team register can show what is approved, what is pending, and what should be excluded until a conflict is resolved. Share the inventory with the source owners so unresolved items can be assigned and checked before their next review. Review it during policy changes and after conversations reveal a missing or obsolete fact.

For each source, ask whether a customer can understand it without internal context. Replace unexplained acronyms, resolve contradictory dates, and state conditions near the rule they qualify. Do not rewrite legal or policy content without the owner’s approval. The aim is not to make every document sound alike; it is to give the assistant a coherent, approved basis for the questions in scope.

Where a source is lengthy, identify the sections that support the intended questions and check whether nearby material changes their meaning. Do not remove qualifiers simply to shorten a document. If preparing a shorter customer-facing extract, have the policy owner compare it with the original and approve the final wording. Keep the canonical original easy to locate so a future editor can trace the extract back to its authority.

When several teams own different sections, confirm who approves the complete customer-facing version. A section-level review can miss a conflict introduced by combining two individually accurate passages.

Resonoon’s AI chatbot widget provides context for the customer-facing experience. A well-governed knowledge base depends on named owners, controlled copies, and review triggers. Those practices make it easier to spot when the assistant lacks evidence and to correct the source before a confident answer becomes a customer problem.