Definition
An architectural framework where an Encoder processes input sequences into a dense latent representation and a Decoder transforms that representation into a target sequence. In RAG pipelines, this structure (e.g., T5 or BART) is often used for abstractive summarization of retrieved context, offering high output quality at the cost of increased computational latency compared to decoder-only models.
Distinguish from simple character encoding (UTF-8) or standalone embedding models; this refers to the dual-stage neural transformation process.
"An hourglass: complex data is squeezed through a narrow neck of 'meaning' before being expanded back out into a new, structured form."
Conceptual Overview
An architectural framework where an Encoder processes input sequences into a dense latent representation and a Decoder transforms that representation into a target sequence. In RAG pipelines, this structure (e.g., T5 or BART) is often used for abstractive summarization of retrieved context, offering high output quality at the cost of increased computational latency compared to decoder-only models.
Disambiguation
Distinguish from simple character encoding (UTF-8) or standalone embedding models; this refers to the dual-stage neural transformation process.
Visual Analog
An hourglass: complex data is squeezed through a narrow neck of 'meaning' before being expanded back out into a new, structured form.