Taking RAG Systems to the Next Level: Reasoning That Doesn't Blow Up Your Context

Retrieval-Augmented Generation (RAG) is a widely adopted approach for enabling Large Language Models (LLMs) to answer questions using external documents. The standard pipeline typically retrieves a set of passages (top- k), includes them in the prompt, and generates a response. While effective for single-source queries, this approach often fails in multi-hop scenarios where answers must be synthesized by connecting facts across multiple sources. The common response to such failures is to increase the retrieval depth by appending more passages. However, in practice, this often degrades performance. As the context window fills with partially relevant text and distractors, the model suffers from context dilution, missing the critical 'bridge' facts buried in noise. In this paper, we present SEAL-RAG: a simple, training-free method that strengthens multi-hop RAG without expanding the context window. The core idea is to treat the evidence slots as a fixed budget. Instead of growing the context, we maintain a constant size of exactly k passages and iteratively improve the information density of those slots through a controlled loop. SEAL-RAG utilizes an Entity Ledger to extract facts from initial passages and identify missing information gaps. It transforms these gaps into targeted follow-up queries, retrieves high-utility candidates, and executes a replacement-based policy to evict low-quality distractors from the fixed- k set. We evaluate SEAL-RAG against industry-standard RAG baselines on two multi-hop QA datasets. Our results demonstrate significant improvements in both answer accuracy and evidence precision. We release our code and data in the supplementary material to support reproducibility.