Large language models (LLMs) have revolutionized how we interact with information, but they sometimes struggle with knowledge-intensive tasks. One common solution is retrieval-augmented generation (RAG), where LLMs access external documents to supplement their knowledge. Traditional RAG systems rely heavily on similarity between the user's query and the retrieved documents. However, a new research paper argues that similarity isn't enough. The paper introduces METRAG, a Multi-layEred Thoughts enhanced Retrieval-Augmented Generation framework. METRAG goes beyond simple similarity by incorporating two key innovations: utility and compactness. First, it trains a "utility model" that learns which documents are most helpful in answering a specific query, even if those documents aren't the most semantically similar. This model uses feedback from an LLM to understand the actual usefulness of a document, not just its surface-level relevance. Second, METRAG employs a "task-adaptive summarizer" to condense the retrieved information, focusing on the most pertinent details. This summarizer is trained to align with the specific task, ensuring that the LLM receives the most relevant information without being overwhelmed by extraneous details. By combining these two layers of thinking—utility and compactness—METRAG significantly improves the performance of LLMs on knowledge-intensive tasks. Experiments on various question-answering datasets show that METRAG outperforms traditional similarity-based RAG systems, especially on complex questions requiring deeper reasoning. This research highlights a crucial shift in how we think about information retrieval for AI. It's not just about finding similar documents; it's about finding the most useful and concise information to empower LLMs to answer complex questions accurately and efficiently. Future research could explore applying METRAG to even more complex scenarios, such as legal or medical document analysis, where the ability to extract and synthesize key information is paramount.