# DRAGIN: Dynamic Retrieval Augmented Generation based on the Information Needs of Large Language Models

**Authors:** Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, Yiqun Liu  
**arXiv:** [2403.10081](https://arxiv.org/abs/2403.10081)  
**Date:** 2024-03-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.10081v2/

## Key points

- The Dynamic Retrieval Augmented Generation (DRAGIN) paradigm is crucial in text generation using Large Language Models (LLMs).
- DRAGIN focuses on identifying the optimal moment to activate the retrieval module and crafting the appropriate query during text generation.
- Existing dynamic RAG methods face challenges in deciding when to retrieve and what to retrieve, often relying on static rules and overlooking relevant information spanning across the entire context.
- A new framework called DRAGIN has been introduced to address these limitations, making informed decisions based on real-time information requirements.
- In evaluations over four knowledge-intensive generation datasets, DRAGIN outperformed existing methods across all tasks, showcasing its effectiveness.
- The authors of DRAGIN include Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu.
- Their research paper delves into how DRAGIN revolutionizes text generation by dynamically adapting retrieval decisions for LLMs' evolving information needs.
- All code, data, and models associated with DRAGIN are openly accessible on GitHub at https://github.com/oneal2000/DRAGIN/tree/main.

## Summary

In the realm of text generation using Large Language Models (LLMs), the Dynamic Retrieval Augmented Generation (DRAGIN) paradigm plays a crucial role in determining when and what information to retrieve during the generation process. This paradigm focuses on two key components: identifying the optimal moment to activate the retrieval module and crafting the appropriate query once retrieval is initiated. However, existing dynamic RAG methods face challenges in both areas. The strategies for deciding when to retrieve often rely on static rules, limiting their adaptability to real-time information needs. Additionally, the strategies for determining what to retrieve typically focus on recent sentences or tokens, overlooking the potential relevance of information spanning across the entire context. To address these limitations, a new framework called DRAGIN has been introduced. DRAGIN stands for Dynamic Retrieval Augmented Generation based on the Information Needs of LLMs and is specifically designed to make informed decisions on when and what to retrieve based on real-time information requirements during text generation. In an extensive evaluation over four knowledge-intensive generation datasets, DRAGIN outperformed existing methods across all tasks, showcasing its effectiveness in enhancing text generation processes. The authors of this innovative framework include Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. Their research paper titled "DRAGIN: Dynamic Retrieval Augmented Generation based on the Information Needs of Large Language Models" delves into how DRAGIN revolutionizes text generation by dynamically adapting retrieval decisions to meet LLMs' evolving information needs. The paper not only presents the theoretical underpinnings of DRAGIN but also provides empirical evidence through comprehensive experiments that validate its superior performance compared to existing methods. For those interested in exploring DRAGIN further or implementing it in their own projects, all code, data, and models associated with this framework have been made openly accessible on GitHub at https://github.com/oneal2000/DRAGIN/tree/main. This transparency underscores the authors' commitment to advancing research in dynamic retrieval augmented generation within the domain of Large Language Models.