Self-Retrieval: Building an Information Retrieval System with One Large Language Model

The article discusses the development of Self-Retrieval, a new information retrieval architecture designed to enhance the performance of large language models (LLMs). The authors argue that existing information retrieval systems are inadequate due to their isolated architecture and limited interaction, which hampers their ability to serve LLMs effectively. Self-Retrieval addresses these issues by integrating the capabilities of information retrieval systems into a single LLM, using a natural language indexing architecture to internalize the corpus to be retrieved.

The retrieval process is then reimagined as a procedure of document generation and self-assessment, which can be executed end-to-end using a single large language model. The authors claim that Self-Retrieval significantly outperforms previous retrieval methods and can greatly enhance the performance of LLM-driven applications like retrieval augmented generation.

Key takeaways

The authors propose Self-Retrieval, an end-to-end, large language model (LLM)-driven information retrieval architecture.
This architecture internalizes the corpus to retrieve into an LLM via a natural language indexing architecture.
The entire retrieval process is redefined as a procedure of document generation and self-assessment, which can be executed using a single large language model.
Experimental results show that Self-Retrieval significantly outperforms previous retrieval approaches and can boost the performance of LLM-driven downstream applications.

Self-Retrieval: Building an Information Retrieval System with One Large Language Model

Key takeaways

Discussion (0)