# Large Search Model: Redefining Search Stack in the Era of LLMs

**Authors:** Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, Furu Wei  
**arXiv:** [2310.14587](https://arxiv.org/abs/2310.14587)  
**Date:** 2023-10-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.14587v1/

## Key points

- Various components in search engines are optimized and deployed independently, resulting in a complex search stack.
- The large search model (LSM) is a novel conceptual framework that redefines the search stack by using a large language model (LLM) as a unified solution for all search tasks.
- The LSM formulates all tasks as autoregressive text generation problems and allows customization through natural language prompts.
- LLMs like GPT-4 and LLaMA have demonstrated remarkable zero-shot and few-shot learning capabilities, making them ideal for unified modeling of search tasks.
- In the LSM, all information retrieval tasks except first-stage retrieval are handled by a single large search model.
- Natural language prompts serve as interfaces to customize the behavior of the large search model for different tasks.
- Ongoing research on multi-modal LLMs enables modeling full document contents beyond just textual information.
- Challenges include high inference cost, efficient long context modeling, and ensuring responsible AI principles in content generation.
- A simplified version of the large search model instantiated with LLaMA shows competitive performance in joint listwise ranking and answer generation tasks. Further research is needed to address challenges related to model architecture, training, inference, etc.

## Summary

In the field of search engines, various components such as query understanding, retrieval, ranking, and question answering are optimized and deployed independently. However, this approach results in a complex and difficult-to-maintain search stack. Additionally, the quality of search results for complex information needs is often unsatisfactory. To address these challenges, a novel conceptual framework called large search model (LSM) is proposed. The LSM redefines the conventional search stack by utilizing a large language model (LLM) as a unified solution for all search tasks. By formulating all tasks as autoregressive text generation problems, the LSM allows for customization through natural language prompts. This framework leverages the strong language understanding and reasoning capabilities of LLMs to enhance search result quality while simplifying the existing cumbersome search stack. LLMs like GPT-4 and LLaMA have demonstrated remarkable zero-shot and few-shot learning capabilities, surpassing human performance on professional exams. These features make LLMs an ideal choice for unified modeling of search tasks. In the LSM, all information retrieval (IR) tasks except first-stage retrieval are handled by a single large search model. The model generates various elements of the Search Engine Results Page (SERP), including ranked document lists, document snippets, and direct answers based on user queries and retrieved documents. Natural language prompts serve as interfaces to customize the behavior of the large search model for different tasks. The adoption of LLMs also enables performing new tasks that were not explicitly trained. Ongoing research on multi-modal LLMs allows for modeling full document contents beyond just textual information. However, there are several challenges that need to be addressed before implementing this framework in production systems. The inference cost of LLMs is currently high due to their autoregressive nature, making real-time applications challenging. Efficient long context modeling without compromising quality is another open problem. Moreover, ensuring responsible AI principles in content generation is crucial for deployment. To validate the proposed approach, a simplified version of the large search model is instantiated using the open-source LLaMA model. Preliminary experiments on joint listwise ranking and answer generation tasks show competitive performance compared to strong baselines; however further research is needed to establish benchmarks and develop new methods to address challenges related to model architecture, training, inference, and more.