Definition
Vector representations where the embedding model ingests a natural language instruction (e.g., 'Represent this document for retrieval') alongside the input text to dynamically adjust the semantic mapping. This improves retrieval precision by aligning the embedding space with the specific intent of the RAG task, though it introduces architectural trade-offs like increased token consumption and higher latency compared to instruction-free models.
Not just task-specific fine-tuning; it is an on-the-fly contextual adjustment using a text prefix at inference time.
"A Multi-Tool Lens: Looking at a map through a 'topography' filter highlights mountains, while a 'transit' filter highlights roads, even though the underlying terrain is the same."
Conceptual Overview
Vector representations where the embedding model ingests a natural language instruction (e.g., 'Represent this document for retrieval') alongside the input text to dynamically adjust the semantic mapping. This improves retrieval precision by aligning the embedding space with the specific intent of the RAG task, though it introduces architectural trade-offs like increased token consumption and higher latency compared to instruction-free models.
Disambiguation
Not just task-specific fine-tuning; it is an on-the-fly contextual adjustment using a text prefix at inference time.
Visual Analog
A Multi-Tool Lens: Looking at a map through a 'topography' filter highlights mountains, while a 'transit' filter highlights roads, even though the underlying terrain is the same.