How the assistant picks what to use

The assistant does not read your whole knowledge base before it replies. For each incoming message it runs one search over your organization’s indexed passages, takes the best few, and writes its answer with those passages in front of it. Everything else stays where it is.

What gets searched

Only sources in one of three states are searchable, and only the passages belonging to a source’s current version can match. The status badge on each row in Knowledge Base tells you which group a source is in.

Every search carries your organization as a required filter — in the application code and again in the database query itself. Only your own material is searched: what you upload is never searched on behalf of another business, and another business’s material is never searched on behalf of yours. The same filter excludes the passages of a replaced version, so an old price cannot answer in the seconds between a rebuild finishing and the old passages being cleared away.

How a passage is matched

When your knowledge is indexed, each passage is converted into a numeric representation of its meaning. When a message arrives, the same conversion is applied to the customer’s text, and the store returns the passages whose representation is closest to it.

• Up to 5 passages are taken per message. That is the whole of what the assistant is given from your knowledge base for that reply.
• The comparison is on meaning, not on words, so “how much is a men’s cut” can reach a passage that reads Service: Men’s haircut / Price: 90.00 ILS even though none of those words match.
• Nothing filters the search by language. A question is compared against every eligible passage, in whichever language it was written.
• The passages that come back are placed in the prompt as data, together with the last few turns of the conversation. Text inside them is never treated as an instruction — a document that says “ignore your instructions” is stored and read as content only.

Very short replies are handled specially. When a customer sends 80 characters or fewer — “yes”, “and the price?” — the last 240 characters of the assistant’s previous message are added to the search text, so a follow-up searches for the thing that was actually being discussed rather than for the word “yes”.

Why your material is split into passages

A document is not stored as one block. It is cut into passages at indexing time, and how it is cut decides what can be retrieved later.

• A passage targets about 500 tokens and never exceeds 1,200.
• A table row, an FAQ question-and-answer, and a list item are treated as atomic: they are never cut in half. That is what keeps a price with the service it belongs to.
• A heading always starts a new passage and is repeated inside it, so a passage carries the section it came from.
• Consecutive passages overlap by about 50 tokens, carried as whole blocks rather than a cut-off sentence.
• A fragment shorter than 24 tokens is merged into the passage before it, but only when both come from the same section.

Why one long document can be less useful than a short entry

The limit that matters is the 5 passages per message. A 60-page staff handbook becomes hundreds of passages that compete with each other, and the five that come back are the five that looked closest to the customer’s wording — not necessarily the five that answer the question. A cancellation policy written as one FAQ entry is a single self-contained passage that carries its own question, its own answer and its own heading.

• Split a long document into several documents, one topic each, and upload them separately.
• Give documents real headings. A heading is repeated into every passage under it and gives the search something to match.
• Keep prices and opening hours in Business settings rather than buried inside a PDF — they are indexed as labelled rows that stay intact.
• Do not upload the same information twice. Two copies compete for the same five places and one of them will eventually be the outdated one.

What “nothing was found” looks like to the customer

The customer never sees an error, a code, or a message about a knowledge base. When the search comes back empty, the assistant is still asked to reply, with one extra instruction: no specific information was found for this question, so do not invent one — ask a short clarifying question or offer to connect the customer with a team member. What the customer gets is a short, ordinary reply asking what exactly they need, or an offer to pass them to a person.

Finding nothing does not hand the conversation over on its own. Handover happens when a customer asks for a person, sounds frustrated, or raises a sensitive or high-risk topic — see when the assistant hands over. It also means an assistant with an empty knowledge base still greets people and makes small talk; it simply has no business facts to state.

Making the right passage come back

• Ask the real customer question on the knowledge test page and look at which source the passages came from.
• Write entries in the words your customers use, not in internal product names.
• Split any FAQ entry that answers two questions into two entries.
• Remove or disable anything the new material replaces — see update, replace and delete knowledge.
• If nothing comes back at all, check the status badge first: a source that is not Ready, Needs a check or Updating is not being searched.