Dynamic LLM selection based on input language

[mamadaevv]

Context

Currently, all user input is processed by a single LLM regardless of language. This limits flexibility in terms of performance, cost, and language-specific model strengths.

Proposed solution

Route user input to different LLM models based on the currently active input language (the same language the system already knows from keyboard layout / IME / user settings). No separate language detection is needed.

How it should work

1. Read the active input language from the system (e.g., keyboard layout, IME, or user-selected language context).
2. Map that language to a specific LLM via configuration.
3. If no mapping found for the active language → fallback to a default model.

Config example (YAML)

llm_routing:
  en: "gpt-4o"
  ru: "yandex-gpt"
  zh: "qwen-2.5"
  default: "gpt-4o-mini"

Requirements

• Read active input language from system context
• Configurable language → LLM mapping (file or env vars)
• Fallback to default

Edge cases

• Language not configured → fallback to default
• Model unavailable → fallback to default