A trilingual health glossary with Kazakh lands on Hugging Face, plus an MCP server without an LLM

If you build search systems or chatbots for Central Asia, you already know what a real query looks like. Someone types “qalqansha bez,” drops in “shchitovidka” a word later, and adds “ttg” at the end. Three languages and two alphabets in one line. Standard models stumble over this because Kazakh is considered a low-resource language in NLP: publicly available texts are scarce, and annotated parallel health data is almost nonexistent.

One author tackled this problem by hand. She runs a website with more than a hundred pages on women's health and psychosomatics. The texts were written in Russian and translated into Kazakh and English. She used them to compile a small open glossary covering the body and well-being and published it on Hugging Face. The second part of the story is an MCP server for ChatGPT and Claude, drawing on the same website, with no language model inside.

Why a small glossary matters

A large corpus and a small glossary solve different problems. A corpus teaches a model a language. A glossary helps a system understand that “qalqansha bez,” “shchitovidnaya zheleza,” and “thyroid” refer to the same concept. For website search, query normalization, data annotation, and translation quality checks, a reference like this is often more useful than another fine-tuned checkpoint.

A practical example: you have a medical or wellness product, and a user types “ttg qansha bolu kerek” (“what should the TSH level be?”). Before sending the query to an LLM, you can map it to shared concepts (the hormone TSH and the thyroid gland). Then retrieve the right pages or answers in the required language. This eliminates errors in areas where the model simply has not seen enough Kazakh examples.

An important caveat: the source is an announcement, which does not specify the glossary's size, license, or format. Before using it, open the dataset card on Hugging Face and check those details yourself. The health texts come from the author's own website, so this is not a clinical reference and must not be used to make medical decisions.

An MCP server with no model inside

The second part is more interesting for anyone building agents. MCP (Model Context Protocol) is a protocol that lets clients such as Claude or ChatGPT connect to external tools and data. A “smart server” usually implies something with an LLM inside. Here, the logic is different: the model on the client side handles the reasoning, while the server remains a simple access layer for vetted content.

This approach has clear advantages:

• Low cost. You do not need to pay for inference or run a GPU. The server serves existing pages and terms.
• Predictability. The server's response is your content, rather than generated text. For health topics, that is a significant advantage.
• Easy maintenance. Update a page or the glossary, and the response updates too.
• No second source of hallucinations. Only the client model can make a mistake, and at least you know which sources it draws its facts from.

What you can take from this

First: if your language is low-resource, do not wait for someone to release a large dataset. A small glossary compiled from your own content and published openly already provides real value, and other developers can expand it.

Second: before embedding an LLM in your MCP server, ask yourself whether it needs one. If you have your own vetted content, the server can be a thin layer with search and a glossary. The client can handle all the reasoning.

Third: multilingualism in real queries means a mixture of languages, rather than three separate sets. Test search on those mixed-language strings, rather than clean translations.

The announcement does not explain exactly how the server works or how the glossary was built. Look for details in the full article on Habr and in the repository. The overall direction is sound: open data for low-resource languages and a simple server backend without unnecessary models.