How search works during a call

Understanding (at least at a surface level) how semantic search works helps you write documents the assistant can use well.

Semantic search, not textual

The search isn't for exact words but for meanings. A question about 'rates for families with children' can find a paragraph about a 'family unit discount', even if the words don't match. This makes the base resilient to how customers phrase their questions.

Why small blocks work better

Documents are split into small blocks of a few hundred words. Each block is retrieved independently. Because of this, short self-contained paragraphs produce better results than walls of continuous text: a block with a single clear idea is easier to match to a question.

Writing documents the AI uses well

• Explicit headings. Phrases like 'Cancellation policy' or 'Indoor pool hours' make every section easy to identify.
• Self-contained sentences. Avoid references like 'as seen above'. Each paragraph must be able to be read on its own and make sense.
• Explicit numbers and details. Write 'free cancellation is possible up to 48 hours before arrival' instead of 'it's possible to cancel before the stay'. Specific data is more useful than generalities.
• No acronyms without explanation. If your business uses internal terms, define them at least once in the document: it helps both the AI and whoever reads the file tomorrow.