Adding another language to a RAG application is not simply a translation task.
In a multilingual ecommerce environment, the same product can have several language-specific names, synonyms and descriptions.
The challenge is keeping the underlying entity consistent while allowing customers to ask questions naturally in their own language.
Our approach separates several concerns:
Canonical product identity: Different names should resolve to the correct underlying product.
Language-specific knowledge: Advice and product descriptions must remain appropriate for the requested language and store view.
Structured constraints: Product attributes and availability rules must remain consistent regardless of query language.
Evaluation: Equivalent questions in different languages should be tested against the same intended result.
Answer quality: Retrieval accuracy and linguistic quality should be evaluated independently.
One important lesson is that fluent answers do not necessarily indicate correct retrieval.
A German answer can sound perfect while referring to an unsuitable product or the wrong catalogue entity.
For multilingual RAG systems, consistency matters just as much as translation quality.