The memory limitation of large language models (LLM) has always been an urgent problem in the AI field. This article explores an alternative solution to vector databases: leveraging improved search engine technology. This solution combines keyword and vector searches and reorders search results through LLM, thereby improving search efficiency and reducing costs. Although this approach has great potential, it also faces challenges such as search engine performance evaluation and deployment.
Researchers believe that building an advanced search engine, combining keyword and vector search technology, and then using LLMs to reorder search results can effectively solve the problem of insufficient LLM memory and eliminate the need to build a specially built ranking model, reducing costs. This provides a new idea for solving the LLM memory bottleneck. However, the article also points out that this solution requires further research and improvement in terms of performance evaluation and actual deployment.
Although this approach holds great promise, there are still practical challenges that need to be overcome. Future research directions should focus on improving search engine performance and solving problems that may arise during its deployment, in order to better meet the needs of LLM applications.