Parameterized Indices for Generative Document Retrieval

This direction is merged into another direction: generative retrieval. Its last version stays readable and citable here.

Version 1 of 116 papers20 conceptsfrozen 2026-10-04

Traditional information retrieval (IR) operates like a library with a massive card catalog. To find a book, you first convert your request into a code (an embedding), then search through millions of other codes in a physical cabinet (a vector database) to find a match.

Generative Retrieval (GR) proposes a radical alternative: what if the librarian simply memorized the entire catalog? Instead of searching an external index, a language model is trained to map queries directly to document identifiers (DocIDs) using its own internal parameters.

From Vector Spaces to Generative Mappings

In the standard "Dual Encoder" architecture, retrieval is a geometric problem. Models map queries and documents into a high-dimensional space, and search involves finding the nearest neighbors. While effective, this requires storing millions of dense vectors in memory, leading to massive infrastructure overhead.

Generative Retrieval, as explored by Jiangui Chen, Jiafeng Guo, and Ruqing Zhang in their work on UGR, treats retrieval as a sequence-to-sequence task. The model receives a query and "decodes" the unique identifier of the relevant document. By encoding the relationship between text and DocIDs directly into the model's weights, the system functions as a parameterized index.

The Efficiency of Model-Based Compression

The primary breakthrough of this approach is extreme memory efficiency. Traditional dense retrievers require external indices that scale linearly with the number of documents. In contrast, GR models compress this information.

Research on AdMIRaL by Rami Aly and Andreas Vlachos 0001 demonstrates that autoregressive retrieval can achieve a 5x to 10x reduction in memory usage compared to dense methods. Even more strikingly, Xiaoxi Li and Zhicheng Dou show in CorpusLM that by treating DocIDs as natural language tokens, memory overhead can be reduced by over 100-fold. The model weights effectively act as a highly lossy but functional compressor for the entire corpus.

Defining the Document: Substrings and N-grams

A central challenge in GR is deciding what a "Document Identifier" should be. If you use arbitrary integers, the model struggles to learn the relationship between the document's content and its ID.

To solve this, Michele Bevilacqua and Patrick Lewis introduced SEAL, which uses actual substrings from the document as identifiers. By constraining the model to only generate strings that exist in the corpus (using a compressed FM-Index), the model leverages the semantic meaning of the text itself to navigate the index. Similarly, UGR utilizes n-gram-based identifiers to allow retrieval at multiple granularities, from specific entities to full documents.

Unifying Retrieval and Reasoning

Perhaps the most compelling aspect of this direction is the unification of the "Retrieve-then-Generate" pipeline. In traditional Retrieval-Augmented Generation (RAG), the retriever and the generator are two separate systems that communicate via a shared context window.

CorpusLM collapses this distinction. Because the model is the index, it can perform retrieval, reasoning, and answer generation in a single continuous decoding pass. This eliminates the "lost in the middle" problems of long-context windows and ensures that the model's internal knowledge and the external corpus are tightly integrated.

The Challenge of Dynamic Knowledge

If the index is stored in the model's weights, what happens when the world changes? Traditional indices are updated by adding a new vector to a database. GR models, however, require weight updates.

Soyoung Yoon and Minjoon Seo have shown that GR is surprisingly robust to temporal biases in dynamic corpora. To make this practical, CorpusBrain++ introduces a continual learning framework. By using backbone-adapter architectures and "semantic experience replay," the model can learn new documents without forgetting the old ones, maintaining its status as a living, parameterized archive of information.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_parameterized_indices_for_generative_doc_6371f433,
  title = {Parameterized Indices for Generative Document Retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/parameterized-indices-for-generative-document-retrieval/rdr_6371f433660348108bb454a03bcc04ba},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_parameterized_indices_for_generative_doc_6371f433_v1,
  title = {Parameterized Indices for Generative Document Retrieval},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/parameterized-indices-for-generative-document-retrieval/rdr_6371f433660348108bb454a03bcc04ba/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}