Languages

The assistant replies in English (en), Arabic (ar), or European Portuguese (pt, pt-PT).

Audience: developers serving a multilingual audience. What you will accomplish: control reply language and understand auto-detection and RTL.

Set the language

Pass lang in the request body (see Chatting):

Value	Behavior
`auto` (default)	Detects the language of the input.
`en`	Forces English.
`ar`	Forces Arabic.
`pt`	Forces European Portuguese (pt-PT spelling and vocabulary).

Setting lang explicitly bypasses detection entirely — use it when you want certainty.

How auto-detection works

When lang is auto, a three-tier hybrid resolves the language:

Arabic — detected by Unicode script (instant and definitive).
Portuguese heuristic — distinctive diacritics (ã õ á é í ó ú â ê ô à ç) are decisive; otherwise a Portuguese stopword-frequency ratio decides for sentences of at least four words. A single shared token (e.g. English “no”, also a Portuguese stopword) cannot flip the verdict.
Statistical fallback — only short, unaccented, genuinely ambiguous fragments fall through to an EN/PT statistical model. On any error it defaults to English, so a request never fails on detection.

The resolved language is returned in meta.lang.

Right-to-left (Arabic)

Arabic replies are right-to-left. In your UI, set the text direction for Arabic answers so they render correctly:

<div dir="rtl" lang="ar">…answer…</div>

The reference web client in web/ renders RTL for Arabic automatically; mirror that in your own UI by switching dir based on meta.lang.

Multi-turn follow-ups

Keep a stable X-User-Id and ask naturally. Before retrieval, the bot condenses elliptical follow-ups into a standalone search query, so “and what about damaged items?” resolves its pronouns against the prior turns and retrieves the right documents. This is automatic on multi-turn chats and skipped on the first turn.

Verify your result

Verify: An Arabic question returns an Arabic answer with meta.lang = ar (when auto).
Verify: Setting lang: "pt" returns European Portuguese regardless of the input language.
Verify: A follow-up like “and for damaged items?” retrieves the right documents when the X-User-Id is stable across turns.

Common mistakes and fixes

Mixed-language input returns the wrong language → set lang explicitly instead of relying on auto for short or code-switched inputs.
Follow-up retrieves nothing relevant → confirm you reused the same X-User-Id; memory (and therefore follow-up condensing) is scoped per id and expires after the TTL. See Conversation memory.

Control language per request in Chatting.
Understand the memory TTL behind follow-ups in Conversation memory.

Streaming (SSE)Trust & citations