Generative retrieval models index and retrieve documents in a single end-to-end pass, avoiding traditional separate indexing steps. A key design choice is how documents are identified. According to a new arXiv paper, numeric Document Identifiers (DocIDs) are generally more efficient than alternatives because they are shorter and typically fixed-length sequences.

The paper, titled "Computed, not Memorized: Bit-Factorized Identifiers for Efficient Generative Retrieval," proposes a method that computes these identifiers using bit-factorization rather than relying on memorized representations. This shift aims to further reduce the cost and complexity of generative retrieval.

The abstract does not provide experimental results or comparisons, so the claimed efficiency gains are presented as the core contribution. The work appears to target the bottleneck of identifier length and sequence structure in generative retrieval systems.