Choose Documents That Help a Voice Assistant Answer Clearly

Choosing documents for voice assistant knowledge is an editorial task as much as a technical one. A folder containing every brochure, staff note and historical policy may look comprehensive while making the assistant's answers harder to trust. The useful collection is the smallest set that reliably supports the questions callers are allowed to ask.

Begin with the caller's needs. A sales presentation, internal manual and public FAQ serve different audiences. The assistant should not expose internal instructions or treat promotional language as a binding service promise simply because both appear in an uploaded file.

Make a question inventory

List the common questions the assistant should answer, then identify the authoritative source for each. If a question has no approved answer, mark that gap instead of hoping a large document collection will compensate.

A hypothetical equipment rental business might need service areas, collection rules, opening hours and the process for reporting a fault. It may not need the full internal maintenance handbook. The intended task determines the document set.

Burki's voice-agent building guide provides wider setup context. Knowledge selection should follow the assistant's agreed scope rather than expanding that scope by accident.

Prefer explicit, self-contained facts

Good source material states who a rule applies to, when it is effective and what exceptions exist. A short paragraph saying “returns are accepted within the stated period for unopened items” is more useful when the period and exclusions appear with it, not on an unrelated page.

Avoid relying on unexplained acronyms, image-only tables or references such as “see the policy above” when the relevant context may not be retrieved together. Review the supported ingestion format and inspect what the system actually extracts.

LiveKit's external data and RAG guide explains how external information can supply context. Retrieval does not make an ambiguous sentence unambiguous; source quality remains the business's responsibility.

Remove competing versions

Select one active version of each policy and keep archived material outside the active answer set unless there is a specific supported reason to include it. Label effective dates and owners clearly.

A common hypothetical conflict is a current FAQ saying a service is unavailable while an older brochure still advertises it. The right fix is to resolve the source selection, not merely add a longer prompt telling the assistant to “be accurate.”

Do not remove useful evidence of past decisions from the business archive. Separate archival retention from what the assistant should use for present-day answers.

Separate static knowledge from live state

Documents work well for approved explanations and stable rules. Current account status, availability or a changing price may need a supported live lookup. A static file cannot honestly establish that it reflects the latest external state unless the update process does so.

Write the boundary into the content. For example, the document can explain how a quote is determined without giving the assistant permission to invent a firm quote for every caller.

The assistant should also know what to do when the necessary live information is unavailable. A concise explanation and supported next step are preferable to a plausible answer assembled from old examples.

Test answers in spoken form

Ask a reviewer to listen to responses generated from the selected documents. Check factual completeness, whether essential conditions survive summarisation and whether the answer is short enough to follow without a screen.

Include a question that crosses two policies and one that asks for an unsupported exception. These expose whether the assistant combines facts correctly or turns missing information into permission.

Record the source used for each important answer where your review process allows it. If the answer cannot be traced to an approved source or action result, investigate before expanding the knowledge collection.

Burki's reviewed learning guide supports this maintenance approach. Start with a question inventory and an authoritative source for each answer; add documents because they close a real gap, not because more pages appear more capable.