# Designing AI for teams that work in Turkish and English.

> In this field note, veridive explains how to design a bilingual AI assistant for teams that work in Turkish and English: answer in the user’s language, search both languages with multilingual embeddings and glossary-expanded keyword search, quote sources in their original language with a translation, handle mixed-language messages, and evaluate each language with native reviewers.

In many organizations the question arrives in Turkish, the policy is in English and the customer writes in both. Decide which language answers go out in, search across languages, cite the source in its original language and evaluate in both.

## Key takeaways

- Answer in the language the user wrote in, and quote the source in its original language with a translation marked as such.
- Search across languages with multilingual embeddings plus keyword search in both languages, expanded with a company glossary.
- Treat mixed Turkish and English messages as normal input: use the conversation’s language, keep familiar terms and match suffixed English words.
- Split the evaluation set by language pair, have native reviewers score each language and report every slice separately.

A regional sales manager types into the company assistant: “Merhaba, travel policy’de otel limiti ne kadar? Per diem yurt dışında farklı mı?” The group travel policy is in English. The local annex that adjusts it is in Turkish. The answer sits across both documents, and the question is already in both languages.

This is ordinary in organizations that work across Türkiye, Europe and [the Gulf](https://veridive.com/insights/arabic-english-ai-systems/), and a system designed for one language fails here unnoticed. It searches only in the language of the question, answers in whatever language the source happened to use, or translates the policy and loses the wording that matters.

Four decisions fix most of it: answer in the user’s language, search across both languages, quote the source in its original language with a translation beside it, and evaluate each language with native reviewers. Here they are as steps, each with what done looks like.

## What does bilingual work look like in practice?

Map it before designing anything: for one workflow, list the languages questions arrive in, the language of each source and the language each output is read in. Typical patterns:

- **Group documents in English, local ones in Turkish.** Policies, contracts and product documentation from headquarters; procedures, forms and annexes written locally.
- **Customers in Turkish, knowledge in English.** Product and returns information kept by an international team, questions from local customers.
- **Threads that switch.** A Turkish email thread forwarded to a colleague in Europe, who replies in English.
- **Chat full of English terms.** “PO onaylandı mı?”, “Deadline’ı kaçırdık”, “Case’i escalate ettim”.

**Done looks like:** a one-page language map per workflow, including which version is authoritative when a document exists in both languages. **What goes wrong:** assuming questions and sources share a language.

## Which language should the answer be in?

Default to the language the user wrote in, or the one in their profile when a message is too short to tell. A Turkish question gets a Turkish answer, even when every source is English. Write down the exceptions:

- **Records have their own language.** A customer gets a reply in Turkish, while the ticket summary for a regional team may be in English. Decide per output, not per conversation.
- **Keep the user’s terms.** If someone asks about “per diem”, don’t answer only with “harcırah”; show the company’s term once beside theirs.
- **Customers get customer language.** Internal shorthand such as “PO” belongs in internal answers, not in replies to customers.

**Done looks like:** a written rule per channel, tested on examples. **What goes wrong:** the model answers in the language of the retrieved passage, so a Turkish question gets an English answer because the policy was English. It is a common failure, and easy to test.

## How do you search across languages?

Use two methods and combine them. **Multilingual embeddings**, vector models trained on many languages, place a Turkish question and an English passage with the same meaning close together, so semantic search can cross the language line. Their quality varies by model and domain, so test on your own question and document pairs.

**Keyword search in both languages** catches what embeddings blur: policy names, product codes, legal terms. Expand the query with a company glossary (“harcırah” and “per diem”, “masraf” and “expense”, “avans” and “advance”), run each language through its own analyzer, then merge and rerank the results, as in [hybrid search](https://veridive.com/insights/hybrid-search-rag/). Keep the original documents as the source of truth; if you also index translations, cite the original. Building this layer on approved, permission-aware sources is the core of [data and AI foundations](https://veridive.com/services/data-ai-foundations/).

**Done looks like:** on the evaluation set, Turkish questions find the right English passages near the top, and the reverse. **What goes wrong:** only the question’s language is searched, or nobody maintains the glossary.

## How should citations to the other language work?

Quote the source in its original language, verbatim, with a link to the exact passage and its version. Put a translation next to it, marked as a translation. The answer rests on the original because translation shifts meaning exactly where it matters: English “should” can come out in Turkish as a firm “-meli” or a softer “önerilir”, turning a recommendation into an obligation or the reverse. When Turkish and English versions of a policy differ, show both and say which is authoritative, or flag the conflict for the owner. When the translation itself needs a reviewer, see [AI translation at work](https://veridive.com/insights/ai-translation-for-business/).

Consider an illustrative case. An employee asks: “Yurt dışı iş seyahatinde taksi masrafımı geri alabilir miyim?” The answer lives in the English group travel policy, and retrieval finds it through the glossary and the multilingual embedding. The assistant answers in Turkish: “Evet. İş seyahatindeki taksi ücretleri, fişle belgelendiğinde geri ödenir.” Below that, it quotes the English sentence it relied on (“Taxi fares for business travel are reimbursable when supported by a receipt.”) with the section and a link, followed by a Turkish translation labeled as one. The employee gets a usable answer; a reviewer gets the exact wording.

> Answer in the user’s language. Quote the source in its own.

## How do you handle mixed-language messages?

Code-switching, English terms inside Turkish sentences, is normal at work: “Call’a giremedim, notları mail’den atar mısın?” or “Invoice’u girdim ama PO ile eşleşmiyor.” Three things break on it:

- **Language detection.** Short mixed messages get misclassified. Use the conversation’s language and the user’s profile, not a guess from one line.
- **Search.** “Invoice’u” must match “invoice”. Split at the apostrophe, and handle suffixes attached without one, as in “maili”.
- **Preprocessing.** Cleaning steps that “correct” English words into Turkish, or strip them, remove the most informative words in the message.

Reply in the message’s main language, usually the one carrying the grammar, and keep the English terms the team actually uses. More on the Turkish side in [what makes Turkish hard for AI](https://veridive.com/insights/turkish-language-ai/).

**Done looks like:** mixed messages are a tagged slice of the evaluation set. **What goes wrong:** the test set is “cleaned” into pure Turkish, so the system never meets real input before launch.

## How do you evaluate a bilingual system?

Split the evaluation set by language pair, and report each slice separately:

| Slice | Example | What to check |
|---|---|---|
| Turkish question, English source | A leave question answered from a group policy | Right passage found; answer in Turkish; quote verbatim |
| English question, Turkish source | A question about a local procedure | Right passage; faithful translation of the quote |
| Same language | Turkish to Turkish, English to English | The baseline to compare against |
| Mixed messages | “PO’yu kim onaylıyor?” | Routing, answer language, terms kept |

Native reviewers score each language: Turkish speakers judge Turkish answers for grammar, register and terminology, English speakers judge the English ones, and a bilingual reviewer checks each quote against its translation. An overall average hides a system that is strong in one slice and weak in another.

## One workflow, two languages

Pick one workflow where both languages already meet, such as HR policy questions or customer replies. Draw its language map, collect a few dozen real questions per slice, and test whether search finds the right passage across languages before anyone writes a prompt. For internal documents, this is what [knowledge and document intelligence](https://veridive.com/solutions/knowledge-document-intelligence/) is built for; for customer messages, the same design runs inside [customer operations](https://veridive.com/solutions/customer-operations/).

## Frequently asked questions

### Should an AI assistant answer in the user’s language or the document’s language?

In the user’s language, with the source quoted in its original language. A Turkish question answered from an English policy should get a Turkish answer, followed by the English sentence it relies on and a translation marked as a translation. The answer stays easy to use and the evidence stays exact, because the original wording is what the policy actually says.

### How does multilingual RAG find documents in another language?

It combines two methods. Multilingual embeddings place a question and a passage with the same meaning close together even when they are in different languages, so semantic search can cross languages. Keyword search runs in both languages, with a company glossary translating key terms. The results are merged and reranked, and tests on your own question and document pairs show whether it works.

### How should an AI system handle code-switching between Turkish and English?

Treat mixed Turkish and English messages as normal input, not noise. Detect the conversation’s main language instead of guessing from one short message, keep the English terms your teams actually use, and make sure search matches English words that carry Turkish suffixes, such as “invoice’u”. Add real mixed messages to the evaluation set and score routing and answers on them separately.