Definition
The mechanism by which an LLM or embedding model applies knowledge or semantic patterns learned in a source language to perform tasks in a target language without explicit training. In RAG architectures, this enables cross-lingual retrieval where queries in one language can successfully find semantically relevant documents in another language within a shared vector space.
Distinguish from Machine Translation; this is about shared conceptual representations, not converting text strings.
"A Universal Library Index: Books are written in different languages, but they are all assigned the exact same Dewey Decimal number based on their topic."
Conceptual Overview
The mechanism by which an LLM or embedding model applies knowledge or semantic patterns learned in a source language to perform tasks in a target language without explicit training. In RAG architectures, this enables cross-lingual retrieval where queries in one language can successfully find semantically relevant documents in another language within a shared vector space.
Disambiguation
Distinguish from Machine Translation; this is about shared conceptual representations, not converting text strings.
Visual Analog
A Universal Library Index: Books are written in different languages, but they are all assigned the exact same Dewey Decimal number based on their topic.