FLARE (Forward-Looking Active Retrieval Augmented Generation) triggers retrieval during generation when the model’s confidence in its emerging content drops. It extends naive RAG by deciding when evidence is needed instead of retrieving only at the start. section 4.7.2
FLARE responds to weaknesses in naive RAG:
- Naive RAG can produce disjointed context, unclear attribution, and minimum response times of two to three seconds. section 4.7.2
- Unlike HyDE, Self-RAG, and Fusion-in-Decoder, FLARE focuses on active retrieval during sentence generation; it is described but not implemented here. section 4.7.2
