Active Retrieval

Active retrieval focuses on intelligently deciding when to incorporate external knowledge into large language models (LLMs) during tasks like text generation or code creation, improving efficiency and performance. Current research emphasizes developing unified and efficient retrieval strategies, often employing multiple criteria to determine retrieval necessity and leveraging techniques like knowledge distillation and contrastive learning to optimize embedding-based retrieval. This area is crucial for enhancing the capabilities of LLMs across various applications, particularly in scenarios requiring access to external knowledge bases, as demonstrated by its successful deployment in commercial search engines.