Which parts of a reply can change language, and which never will

A visitor writes in one language and gets a reply, a refusal or a form back in another.

In rough order of how often each one turns out to be the answer. Work down rather than across: each carries a way to tell whether it is yours before you change anything.

• 1

  It is a pleasantry, answered from fixed sentences before anything runs

  Why

  Greetings, thanks, goodbyes, apologies, asking what this is and asking for a person are recognised and answered instantly from a small set of fixed sentences that name your brand. No matching runs, nothing is asked, and nothing about those sentences varies. They exist because refusing somebody for saying hello reads as broken, and the price is that they are always in one language.

  How to confirm it is this one

  It arrives instantly with no typing indicator, it carries no source underneath, and it is identical every time. Send the same pleasantry twice and compare the two replies character for character.

  Fix

  Nothing to configure, and nothing is going wrong. Anything that is a real question, including one that opens with a greeting, goes down the normal path and is answered normally.

• 2

  It is your refusal, stored as one string exactly as you wrote it

  Why

  The reply when nothing in your material covers a question is the sentence in the When it has no answer box on the Behaviour tab, served verbatim. It is one string, so it cannot follow the language of anything, and it is the reply most likely to be reached by a visitor whose language your material does not cover.

  How to confirm it is this one

  It matches that box word for word, and the question turns up on the Insights tab under Unanswered questions. Both are true or neither is.

  Fix

  Write that box for every visitor you actually get. Two short sentences in two languages fit inside it comfortably and are better than one sentence half your visitors cannot read.

• 3

  It is panel furniture rather than a reply

  Why

  Some of what a visitor reads comes from your settings and some is built into the panel. The brand name, the greeting and the input placeholder are yours, from the Appearance tab. The status line, the close label, the question asking whether an answer helped, and every label and button on the details form are fixed and reach no setting at all.

  How to confirm it is this one

  Point at the exact words that are wrong and go looking for them on the Appearance tab. If they are in a box there, they are yours and you can change them. If they are in no box anywhere, they are fixed, and no amount of searching will turn one up.

  Fix

  Change what is yours. For the rest there is no fix to offer, and saying so is more use than sending you round the settings a second time.

• 4

  Your own instructions are pinning a language

  Why

  What you write on the Behaviour tab sits ahead of the standing rules, and one of those rules asks for the language of the question. An instruction naming a language, or written in a way that implies one voice, competes directly with it, and the specific instruction tends to win.

  How to confirm it is this one

  Read your own instructions back and look for a language name, and for wording that assumes one. This is also the only cause here that changes if you edit that box, which makes it easy to prove either way.

  Fix

  Take the language out, or put it in deliberately. An explicit instruction to answer in the language the visitor used is honoured, and is worth adding on any site with visitors in more than one.

• 5

  The reply came out of the cache, which matches meaning and ignores language

  Why

  An answer that cited a source is kept for that assistant and returned when a later question is close enough in meaning. Nothing in that comparison looks at what language either question was written in, so a close enough match returns the stored answer exactly as it was written the first time.

  How to confirm it is this one

  In the network tab, the response to the chat request opens with a meta event carrying cached set to true. The Insights tab also carries a percentage of replies served from the cache, which tells you how much of your traffic this could apply to.

  Fix

  Nothing, and it is not worth switching off even if you could. It clears whenever a source indexes new text and expires within a day, and it is the reason repeated questions cost nothing.

One limit sits behind all of this and no setting moves it. The assistant may draw facts only from your material, so the sentences it writes can follow the visitor while everything it quotes cannot: a policy heading, a product name, a price band or a whole quoted line arrives in the language your documents are written in. An assistant expected to serve several languages properly needs its material in those languages, not an instruction asking it to translate material it is forbidden to go beyond.

Is there a language setting?

No, and there is no language detection either. What exists is a standing instruction to answer in the language of the question, which covers the sentences the assistant writes and nothing else on the screen.

Should I run one assistant per language?

It is the cleaner arrangement when you have the material. Each one gets its own documents, its own instructions, its own refusal and its own key, so the language version of your site installs a different key and everything else follows from that.

Why do the fixed pleasantry replies exist at all?

Because without them a greeting goes through matching, finds nothing resembling it, and is refused. An assistant that answers hello with an apology about its documentation feels broken before it has been asked anything real, and that was the behaviour these replaced.