ICICLE: In-Context Retrieval for Generative Indexing
Generative retrieval (GR) systems integrate document identifiers within model parameters, leading to high costs for corpus expansion and risks of catastrophic forgetting. A recent arXiv publication introduces ICICLE, an in-context indexing framework that utilizes newly incorporated documents as evidence during inference. ICICLE employs a [COPY]-based routing approach, preference-based calibration, and large context adaptation to differentiate between context-grounded retrieval and parametric retrieval. Tests conducted on MS MARCO and NQ320K demonstrate enhanced retrieval performance for newly added documents.
Key facts
- Generative retrieval maps queries to document identifiers using parametric knowledge.
- Adding new documents to GR requires updating model parameters.
- ICICLE is an in-context indexing framework.
- ICICLE performs source-aware docid generation over parametric memory and context-provided pairs.
- ICICLE uses a [COPY]-based routing mechanism.
- ICICLE uses preference-based calibration.
- ICICLE uses large context adaptation.
- Experiments were conducted on MS MARCO and NQ320K datasets.
Entities
Institutions
- arXiv