Generative Retrieval
Generative Retrieval
Generative retrieval replaces the traditional index-then-rank pipeline with a single model that directly generates document identifiers given a query. The model memorizes the corpus during training and retrieves by generating relevant doc IDs autoregressively.
Key Idea
Traditional: Query → [Index Search] → [Rerank] → Doc IDs
Generative: Query → [Seq2Seq Model] → Doc IDs directly
The model parameters serve as the “index” — no separate data structure needed.
Notable Models
| Model | Approach |
|---|---|
| DSI (Differentiable Search Index) | Encode docs into model params; generate hierarchical doc IDs |
| GENRE | Autoregressive entity retrieval; generate entity names directly |
| SEAL | Generate n-grams, then map to documents |
Challenges
- Scalability: Hard to scale to millions of documents
- Corpus updates: Adding new documents requires retraining
- Doc ID design: Choice of ID scheme affects performance significantly (atomic, string, semantic, hierarchical)