---
title: Searching the knowledge
summary: Search finds entries by meaning, so a client's own words or another language still lead to the right entry. When AI is stopped, search falls back to words.
tools: [talemo_find_knowledge]
related: [knowledge-entry, online-admin-and-knowledge, versions-and-removed, suggestions]
search: [semantic search, find answer]
---

# Searching the knowledge

## Where search is used

- The **Find in knowledge** field in the Knowledge section.
- The online admin, before it answers a client, and the assistant, when it looks
  something up.
- A program connected over MCP: `talemo_find_knowledge` with `q`. Without `q` the
  same tool lists entries by topic.

A question is one line of text. In the field and over MCP it is not combined with
the topic, removed or service filters.

## By meaning

The question is compared with pieces of the entries, and the closest live entries
come back, the closest first, each with the piece where it was found. The
screen says how many entries were found and that the search went by meaning. A client may write in
their own words or in another language, and the right entry is still found.

If nothing is close enough, the screen says "Nothing close. No entry is about
this yet. If clients ask it, write one". Over MCP the list is empty.

## By words

Search goes by words only in three cases:

- AI is stopped because the balance came down to the reserve. The screen says
  meaning search returns when the balance is above the reserve.
- The monthly AI spending limit is used up. Meaning search returns next month,
  or at once when the owner raises the limit.
- Meaning search did not answer this time.

A word search can miss an entry that says the same thing in other words, and the
screen says so. Over MCP the answer tells that the search went by words and gives
the reason: the balance is at the reserve, the limit, or no answer from the meaning search.

## New and removed text

An entry is cut into pieces when it is saved. The meaning index of the new text is
added a little later, and until then the entry is found by words. While AI is
stopped, new text is indexed once AI works again. A removed entry is not indexed
and is not found.

## What it costs

Searching by meaning and indexing new text are paid from the business's balance.
On the **Spending this month** tab they appear on the **Knowledge search** line, for the
assistants, the online admin and the search field together. The person or agent
that searches pays for it, and it counts toward the monthly AI limit (see
[Spend of a run](../runs/run-spend.md)).

## Tools and refusals

- `talemo_find_knowledge` with `q` searches: the question in words, up to 300 characters. Ask it the way the client did, in the client's words and language, not translated and not cut to keywords: a question nothing answers reaches the business in these words. It takes no `topic`, `archived`, `service` or `after`.
- The answer lists the closest live entries, the closest first, each with the piece where it was found. Nothing close enough means an empty list.
- `by_meaning` false means the search went by words only, and `words_reason` says why: `credits` (the balance came down to the reserve), `ceiling` (its AI spending limit for the month is used up) or `unavailable` (the search by meaning did not answer this time). A word search can miss an entry that says the same thing in other words, so do not conclude from it that the business has no answer.