# 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/2403.10081v3/

## Key points

- In the realm of text generation processes for Large Language Models (LLMs), the paradigm is crucial for determining when and what information to retrieve during generation.
- The paradigm consists of two essential components: identifying the optimal moment to activate the retrieval module and crafting the appropriate query once retrieval is initiated.
- Existing dynamic RAG methods face limitations in deciding when to retrieve due to reliance on static rules and in determining what to retrieve by focusing only on recent sentences or a few tokens.
- A new framework called DRAGIN (Dynamic Retrieval Augmented Generation based on the Information Needs of Large Language Models) has been introduced to address these shortcomings.
- DRAGIN is designed to make informed decisions on when and what information to retrieve based on real-time information requirements during text generation.
- Comprehensive experiments conducted over four knowledge-intensive generation datasets show that DRAGIN outperforms existing methods across all tasks, demonstrating superior performance in meeting real-time information needs during text generation.
- The authors behind DRAGIN are Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. Their research delves into how this framework revolutionizes dynamic RAG methods by addressing key limitations and enhancing performance outcomes.
- All code, data, and models associated with DRAGIN have been made openly accessible through GitHub at https://github.com/oneal2000/DRAGIN/tree/main for further exploration or replication of findings.

## Summary

In the realm of text generation processes for Large Language Models (LLMs), the  paradigm plays a crucial role in determining when and what information to retrieve during the generation process. This paradigm consists of two essential 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 limitations in both aspects. The strategies for deciding when to retrieve often rely on static rules, which may not adapt well to real-time information needs. Additionally, the strategies for determining what to retrieve typically focus on the LLM's most recent sentence or a few tokens, potentially overlooking critical context spread throughout the entire text. To address these shortcomings, a new framework called  has been introduced. This framework, , is specifically designed to make informed decisions on when and what information to retrieve based on the LLM's real-time information requirements during text generation. To evaluate the effectiveness of , comprehensive experiments were conducted over four knowledge-intensive generation datasets. The results demonstrate that  outperforms existing methods across all tasks, showcasing its superior performance in meeting real-time information needs during text generation. The authors behind this innovative framework are Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. Their research titled "DRAGIN: Dynamic Retrieval Augmented Generation based on the Information Needs of Large Language Models" delves into how this framework revolutionizes dynamic RAG methods by addressing key limitations and enhancing performance outcomes. For those interested in exploring further details or replicating these findings, all code, data, and models associated with  have been made openly accessible through GitHub at https://github.com/oneal2000/DRAGIN/tree/main.