Automatic Language Detection in AI Assistants

The problem: customers who speak different languages

If you sell in Latin America, Spain, Brazil, and the US, your customers speak Spanish, Portuguese, and English. The naive solution — three different chatbots, three duplicate knowledge bases — is expensive and hard to maintain.

The right solution: one assistant that detects language and responds automatically.

How detection works

Citai uses a cascade priority system to determine the response language:

Heuristic detection: fast and accurate

Instead of using an ML model (which adds latency and cost), Citai uses expanded frequent word lists per language. Words like “the”, “and”, “how” identify English; “como”, “para”, “puede” identify Spanish; “como”, “para”, “pode” identify Portuguese.

System prompt injection

Once detected, Citai injects the language into the LLM system prompt, ensuring the entire response is in the correct language.

Multilingual embeddings: the semantic magic

The embedding model (paraphrase-multilingual-MiniLM-L12-v2) understands 50+ languages in the same vector space:

• A Spanish document can answer an English question
• A Portuguese FAQ matches a Spanish query
• No need to duplicate content per language

Content rules: granular content control

Beyond language, you can configure rules controlling what the assistant responds:

BLOCK — Total block

Block questions on specific topics. The user gets a predefined message.

REDIRECT — Redirect to another source

Redirect to a URL or custom message. Useful for topics requiring mandatory human attention.

FILTER — Fragment filtering

Filter retrieved fragments containing blocked terms before sending to the LLM. The LLM never sees that information.

Visual configuration

Language and content rules are configured from the visual interface in each KB’s RAG config. Toggle, select, and inline CRUD — no code needed.

One assistant, all languages. Try Citai →