Multilingual RAG hits language drift

This is less a product launch than a useful field report: multilingual RAG breaks in boring, production-shaped ways before it breaks in benchmark-friendly ways.

• –Retrieved context can overpower user intent when the prompt leaves response language implicit
• –LLM-based language detection is brittle when queries mention foreign names, citations, or legal terms
• –Simple deterministic routing can beat “smart” detection when the target language set is narrow
• –Explicit negative constraints like “never French” can matter when source documents contain strong language cues
• –Teams building RAG for legal, policy, or enterprise corpora should treat output language as a controlled system parameter, not a style preference