Generation-Augmented Retrieval: A New Approach to Information Retrieval

Document retrieval technology forms the foundation for numerous applications, from web search and open-domain question answering to retrieval-augmented generation. The prevailing approach, the so-called bi-encoder method, calculates the semantic similarity between a search query and a document. However, this method reaches its limits when it comes to fully capturing the complexity of the relationship between query and document.

Challenges of Traditional Retrieval Methods

The reduction of the complex semantic relationship to a scalar similarity score offers only limited insights into the actual relevance of a document. Especially with longer documents, it is difficult to identify the most relevant sections that significantly contribute to the similarity score. Tasks such as selecting relevant sentences, highlighting search results, and precisely locating information require a deep and fine-grained text understanding that goes beyond the capabilities of bi-encoder models.

GeAR: A Generation-Augmented Approach

To address these challenges, a new approach called Generation Augmented Retrieval (GeAR) has been developed. GeAR integrates special fusion and decoding modules that enable the model to generate relevant text passages from documents. This generation is based on a combined representation of the search query and the document, allowing the model to focus on fine-grained information. Notably, GeAR as a retriever does not incur any additional computational overhead compared to bi-encoders.

Data Synthesis and Model Training

To support the training of this new framework, a pipeline has been developed that synthesizes high-quality training data using large language models (LLMs). This approach allows GeAR to be effectively trained and evaluated on diverse scenarios and datasets.

Advantages of GeAR

GeAR offers several advantages over traditional retrieval methods:

Fine-grained Text Understanding: By generating relevant text passages, GeAR enables a deeper understanding of the relationship between search query and document. Improved Localization: GeAR facilitates the identification of the most relevant sections within a document. Interpretability: The generated text passages provide additional insights into the retrieval results and improve the transparency of the process. Efficiency: GeAR does not cause any additional computational overhead compared to bi-encoders.

Outlook

GeAR represents a promising approach for the future of information retrieval. The combination of retrieval and generation allows for improved text understanding and more precise localization of relevant information. The release of the code, data, and models after the completion of the technical review will promote further research in this area and enable the development of innovative applications. Especially for companies like Mindverse, which specialize in AI-powered content creation and research, GeAR offers the potential to further increase the efficiency and quality of their solutions. The development of chatbots, voicebots, AI search engines, and knowledge systems could be significantly improved by integrating GeAR.

Bibliography Liu, H., et al. "GeAR: Generation Augmented Retrieval." arXiv preprint arXiv:2501.02772 (2025). Liu, H., et al. "GeAR: Generation Augmented Retrieval." arXiv preprint arXiv:2501.02772v1 (2025). Shen, Z., et al. "GeAR: Graph-enhanced Agent for Retrieval-augmented Generation." arXiv preprint (2025). Mao, Y., et al. "Generation-Augmented Retrieval for Open-Domain Question Answering." Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021. Maiworm, B. "Understanding Retrieval Augmented Generation: A New Frontier in AI." AmberSearch Blog (2023). Maulini, G. "Retrieval-Augmented Generation (RAG) Basics." 7Rivers Blog (2024).