context management

RAG systems frequently suffer from hallucinations and factual inaccuracies due to retrieval of irrelevant or weakly relevant passages; existing approaches are limited to coarse-grained, document-level filtering, resulting in low precision. This paper proposes the first LLM-driven, fine-grained block-level filtering framework, integrating semantic chunking, query-block alignment scoring, and LLM-based re-ranking to enable precise noise suppression prior to generation. Its core innovations are: (1) dynamic relevance assessment and filtering at the semantic block level—departing from conventional document-level paradigms; and (2) a lightweight LLM agent for efficient, interpretable block re-ranking and adaptive threshold-based pruning. Experiments demonstrate substantial hallucination reduction and consistent superiority over mainstream RAG baselines across multi-hop reasoning and fact-checking tasks, with an average 12.7% improvement in factual accuracy.

This work addresses key limitations of Retrieval-Augmented Generation (RAG) for knowledge-intensive tasks—namely, high retrieval latency, document selection errors, and architectural complexity. We propose Cache-Augmented Generation (CAG), a novel paradigm that eliminates real-time retrieval by preloading a bounded knowledge corpus into the extended context window of large language models (e.g., 128K+ tokens) and leveraging parameter-based caching for zero-latency, zero-retrieval-error inference. To our knowledge, CAG is the first systematic architecture to replace retrieval entirely with caching, demonstrating that, under constrained knowledge settings, long-context LLMs can bypass retrieval modules altogether—yielding significantly simplified, more robust systems. Experiments across multiple benchmarks show CAG matches or exceeds RAG performance while eliminating the need for vector databases and dedicated retrievers, thereby drastically reducing system complexity.

Existing RAG research suffers from the absence of a unified, lightweight, and scalable standardized framework, hindering efficient method reproduction, comparative analysis, and evaluation. To address this, we propose RAGFlow: an open-source, modular, and highly customizable RAG research toolkit. Its key contributions are threefold: (1) a novel fine-grained modular architecture that decouples core components—including retrieval, re-ranking, prompt engineering, and evaluation—enabling flexible composition and independent optimization; (2) integrated support for 16 state-of-the-art RAG methods and 38 standardized benchmarks spanning textual and multimodal scenarios; and (3) a unified interface abstraction for LLMs and multimodal LLMs (MLLMs), coupled with efficient preprocessing and evaluation pipelines. Implemented in Python, RAGFlow significantly lowers the barrier to algorithmic experimentation and has been widely adopted for validating novel RAG methods and supporting pedagogical practice.

RAG system performance critically depends on the retriever-reader configuration, retrieval depth, and context quality; suboptimal settings often degrade performance rather than improve it. Method: We propose RAGGED, a systematic evaluation framework that—through multidimensional controlled experiments, controlled noise injection, and response attribution analysis—quantifies language models’ sensitivity spectra to contextual signals versus noise, and establishes a behavior-driven diagnostic paradigm for RAG configuration. Contribution/Results: We identify two canonical performance patterns—monotonic improvement and inverted-U—as context quality varies. Crucially, we reveal fundamental disparities across models in noise robustness and signal utilization capacity. Based on these insights, we distill reusable, model-aware configuration principles, validated across multiple DBQA benchmarks for both effectiveness and generalizability.

Existing RAG methods predominantly rely on single knowledge sources, rendering them inadequate for real-world scenarios requiring integration of heterogeneous, cross-domain knowledge—spanning structured (e.g., databases, tables) and unstructured (e.g., text, images) modalities—while lacking standardized benchmarks and systematic evaluation. Method: We formally define the multi-source knowledge RAG task and introduce PruningRAG, a novel framework featuring a three-stage, multi-granularity knowledge filtering mechanism: semantic similarity pruning, domain-adaptive filtering, and hierarchical attention distillation—dynamically suppressing noise and redundancy. Contribution/Results: We release the first open-source benchmark encompassing multimodal, cross-domain, structured, and unstructured knowledge. On this benchmark, PruningRAG achieves an average 12.7% improvement in QA accuracy and a 31.4% reduction in hallucination rate over state-of-the-art RAG baselines.

To address low semantic recall accuracy and high latency in large-enterprise knowledge base retrieval, this paper proposes a retrieval-augmented framework leveraging large language models (LLMs) to dynamically generate structured metadata. Methodologically, it integrates semantic, recursive, and naive chunking strategies; enhances vector representations via TF-IDF–weighted embeddings and prefix fusion; and employs a cross-encoder for re-ranking to construct high-quality ground-truth labels. A novel metadata consistency metric is introduced to rigorously evaluate retrieval effectiveness. Experimental results demonstrate that recursive chunking combined with TF-IDF–weighted embeddings achieves 82.5% precision, while naive chunking with prefix fusion attains a Hit Rate@10 of 0.925—both significantly outperforming content-only baselines. The framework delivers superior accuracy, low latency, and strong scalability, making it suitable for enterprise-scale knowledge retrieval.

This work addresses the challenge of generating personalized, factually accurate, and topically relevant reading materials tailored to user-specified queries and target readability levels. To this end, the authors propose a four-module system that integrates retrieval-augmented generation (RAG) with large language models (LLMs), uniquely combining RAG with multiple prompting strategies—including Chain-of-Thought, zero-shot, and few-shot prompting—for personalized reading recommendations. The system further incorporates an LLM-as-a-Judge mechanism to automatically evaluate the factual accuracy, relevance, and readability alignment of generated content. Experimental results demonstrate that RAG consistently enhances the performance of diverse models—including LLaMA 4 Scout, LLaMA 3.1 8B, and Gemma2 9B—across all prompting strategies, yielding improvements of up to 26–35 percentage points in relevance and factual accuracy, thereby enabling high-quality customized reading material generation.

This study addresses the limitations of current retrieval-augmented generation (RAG) systems in handling complex engineering documents from the oil and gas industry—particularly multimodal,图文混排 P&ID diagrams—where suboptimal chunking strategies hinder performance. For the first time in a real-world oil and gas documentation setting, the authors systematically evaluate four chunking approaches: fixed sliding window, recursive, semantic breakpoint, and structure-aware chunking, comparing their retrieval effectiveness and computational overhead within an end-to-end RAG framework. Results demonstrate that structure-aware chunking significantly outperforms other methods by preserving logical document structure, achieving superior top-K retrieval metrics while incurring lower computational costs. Moreover, all text-only chunking strategies exhibit fundamental limitations when applied to multimodal documents like P&IDs, underscoring the necessity of integrating multimodal information for effective retrieval.

This work addresses the challenge of improving retrieval-augmented generation (RAG) accuracy and cost-efficiency without scaling up language model (LM) parameters. We propose the “corpus–generator trade-off” principle, systematically demonstrating that expanding high-quality retrieval corpora significantly reduces reliance on large-parameter LMs. Through rigorous experimentation, we show that a medium-scale LM (e.g., 7B) augmented with a quadrupled corpus achieves performance on multiple open-domain QA benchmarks comparable to that of a much larger LM (e.g., 70B). The primary driver of improvement is increased coverage of answer-relevant passages; however, gains exhibit diminishing returns and saturate as corpus size grows. Crucially, this study provides the first quantitative characterization of corpus expansion as a viable alternative to model scaling—establishing its effectiveness, practical limits, and scalability properties. Our findings establish a new paradigm for lightweight, cost-effective RAG deployment grounded in corpus optimization rather than LM parameter growth.

Large language models (LLMs) suffer from static training data, limiting their adaptability to dynamic information updates and domain-specific tasks; meanwhile, conventional retrieval-augmented generation (RAG) exhibits diminishing marginal utility as LLM capabilities advance. This paper systematically analyzes RAG’s core components, applicability boundaries, and critical failure modes through a comprehensive literature review, multi-scenario case studies, and systematic evaluation of collaborative architectures. Results demonstrate that RAG retains indispensable advantages in complex reasoning, real-time knowledge integration, and expert-domain question answering. We propose a novel conceptualization of RAG in the LLM era—not merely as a capability “patch,” but as a synergistic enhancement paradigm targeting controllability, interpretability, and knowledge recency. The study establishes a theoretical framework and practical design principles for next-generation RAG systems that are lightweight, precise, and verifiable.