Building a knowledge base your voice agent can actually answer from

Teams usually arrive with a folder of policy PDFs and an expectation that pointing an agent at them is enough. It is not, and the reason is mundane: documents are written to be read start to finish by someone who already knows what they are looking for. A voice agent needs the opposite.

Write answers, not documents

Retrieval pulls passages, so the unit of authoring should be a passage that stands on its own. If a paragraph only makes sense after reading the two above it, it will be retrieved without them and answered badly.

• Lead with the answer, then the conditions. Not the reverse.
• Repeat the subject in each entry. The account must be active for 30 days beats It must be active for 30 days.
• Keep one entry to one question. Split anything that answers two.
• Spell out numbers, hours and fees explicitly. Do not refer to a table elsewhere.

One fact, one home

The most common failure we debug is not a hallucination, it is a conflict. The same fee appears in three places, two of them out of date, and retrieval picks whichever one scored highest. The model is behaving correctly on bad input.

Pick a home for each fact and make every other mention a reference to it. When the fee changes, one entry changes.

Signs you have duplicates

• The agent answers the same question differently on repeated calls.
• A correction you made last month has reappeared.
• Two entries score close together in retrieval logs for the same query.

Write in the language callers ask in

If your callers ask in Amharic, entries written only in English force a translation step at retrieval time, which is where meaning slips. Author the entries that get asked in Amharic in Amharic. Lucy handles the rest, but it cannot recover nuance that was never written down.

The same applies to terminology. Callers use the words on the street, not the words in the policy manual. Add both to the entry so retrieval matches either.

Close the loop with transcripts

Every wrong answer is a fixable content gap, and the transcript tells you exactly which one. Read the failed calls weekly for the first month. The list of missing entries gets short fast.