An encoder-decoder architecture combines two neural-network roles: the encoder transforms an input sequence into embeddings, and the decoder uses those representations to generate an output sequence. This design is especially useful for sequence-to-sequence tasks. section 2.3.3
The encoder processes semantic content into dense numerical vectors that capture contextual relationships. The decoder then converts those encoded representations into the requested output, extending the Transformer approach beyond using either component alone. section 2.3.3 section 2.0
Choose the architecture according to the amount and form of information being processed:
- Use an encoder to embed large document collections or queries when building retrieval systems. section 2.3.3
- A decoder alone may summarize one document when the article fits within its context length. section 2.3.3
Semantic search for enhanced retrieval
Given the query “Impact of climate change on marine ecosystems Jane Scientist,” the encoder represents the query and repository documents as vectors. The retrieval system can then identify articles matching the query’s meaning, context, author, and related concepts rather than relying only on keywords. section 2.3.3













