Skip to content
Neutrome
Ukrainian-language AI assistant

Semantyka

A family of Ukrainian assistants behind one model catalog

Visit semantyka.com.uaNamespaced catalogsModel compositionAttached toolsFallback and retry

What they do

Semantyka runs Enei, a Ukrainian-language assistant family, as a namespaced catalog with chat, legal, research, and creative variants built from shared base models. You can see it live at semantyka.com.ua.

The problem

  • Ukrainian-language answers needed a curated voice and grounding in Ukrainian legal sources, which no single upstream model provides out of the box.
  • Each assistant variant needed different tools and prompts while still feeling like one product to the client application.
  • The client needed to keep its existing integration while gaining the new product capabilities.

What they built

  • A core/* namespace for shared base models, with voice and turn handling written once.
  • A semantyka/* namespace for the customer-facing chat, legal, research, and poet variants.
  • Three attached tools for Ukrainian legal material, fresh web results, and an internal fan-out reasoning pass.
  • A parallel semantyka-next/* namespace for testing the next chat-model revision against real traffic.

The model catalog

core/voice-1

Shared voice model. It uses a fallback across several retry-wrapped upstreams, so another candidate can answer if one provider fails.

core/turn-1

Shared turn-handling base the assistant variants build on.

semantyka/enei-1-chat

The default assistant. Its executor delegates to the shared base model, so the catalog entry and its behavior can change independently.

semantyka/enei-1-legal

Ukrainian legal variant, grounded by the attached legal-knowledge tool.

semantyka/enei-1-deep-research

Research variant. It uses the chat model with a system prompt for repeated internal reasoning passes.

semantyka/enei-1-poet

Creative-writing variant of the same base.

Attached tools

search_legal_knowledge

Retrieval over a Ukrainian legal corpus for legal answers.

internal_web_search

Fresh web results, for questions where a training cutoff is the wrong answer.

deep_think

An internal fan-out pass that spends extra reasoning on a sub-question and returns the conclusion.

What changed

  • One runtime URL serves the assistant family, and clients select a variant by model id.
  • Changes to the base model apply to every variant that delegates to it.
  • A fallback chain handles failures from an upstream provider.
  • New variants are added as models in the namespace without a client release.
Visit semantyka.com.ua

Build a workspace for your product

Create a workspace, choose a namespace, and deploy the first model.

Open the platformBack to showcases