Definition
A method that introduces content retrieved from external materials during generation to supplement knowledge stored in the model's parameters.
The 2020 RAG research combined pretrained generative models with retrievable external memory. This entry describes the broader retrieval-augmented generation approach, without requiring every implementation to reproduce the original paper's joint training method.
Retrieved content can be updated, but retrieval accuracy, source quality, and generation faithfulness still need separate evaluation. Adding reference material does not automatically eliminate errors.
Principles and impact
Problems solved
- Supplements knowledge stored in model parameters that is difficult to update or trace.
Principles
- Retrieves relevant material and uses it to condition generation, allowing the model to use information beyond its parameters.
Improvements
Not recorded
Significance
- Connects information retrieval with language generation.
History and applications
People
Patrick Lewis
Organizations
Not recorded
Events
Not recorded
Applications
Knowledge-based question answering、Source-assisted generation
Limitations
- Missed retrieval results, incorrect source material, and unfaithful generation can all affect the outcome.
Upstream and downstream relations
Upstream
No direct upstream relations recorded
Downstream
Sources: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks