Experimental results show that SAC-RAG reduces token consumption by 38%–58% at the cost of only a 1–2 percentage point EM drop, with EM actually improving after compression for reasoning-type questions, achieving the optimal quality–efficiency trade-off in terms of token consumption.
Abstract
Retrieval-Augmented Generation (RAG) effectively mitigates the knowledge cutoff and hallucination issues of large language models by incorporating external knowledge bases. However, standard RAG applies a uniform retrieval strategy to all question types and feeds redundant retrieval results directly into the generator, leading to high token consumption and potentially degraded answer quality. This paper proposes SAC-RAG (Semantic Adaptive Compress RAG), a unified framework that integrates question-type-adaptive retrieval with context compression. SAC-RAG first classifies queries into three types—factual, definitional, and reasoning—via an LLM classifier, applies a differentiated retrieval strategy for each type, then uses a unified LLM context compression module to extract evidence snippets, and finally generates answers from the compressed and refined context. To independently quantify the contribution of each component, we design four progressive ablation experiments on three multi-hop QA datasets—HotpotQA, 2WikiMultiHopQA, and MuSiQue—and evaluate using four metrics: Exact Match, F1, Recall@10, and token consumption. Experimental results show that SAC-RAG reduces token consumption by 38%–58% at the cost of only a 1–2 percentage point EM drop, with EM actually improving after compression for reasoning-type questions, achieving the optimal quality–efficiency trade-off in terms of token consumption, with a moderate increase in per-query latency due to additional LLM calls. SAC-RAG offers a modular and interpretable optimization solution for the efficient deployment of RAG in real-world settings.
The Adaptive Multi-Stage Vector Retrieval (AMSVR) framework is proposed, prioritising weighted, drift-resistant composition over uniform fusion, and offers tailored configurations: AMSVR-Scientific (dense + tuned hybrid) peaks at NDCG@10 = 0.7570 on SciFact, while AMSVR-Full (seven stages) targets broader, noisier corpora where Recall@100 matters most.
Samsudeen Alabi Bankole, Yakub Kayode Saheed· NLP & Big Data· 0 citations
Large language models may suffer from insufficient context use and unsupported generation in question answering tasks that require external knowledge. This study compares the main strategies affecting retrieval and generation performance in retrieval-augmented generation systems within a common experimental setting. In addition to a standard dense retrieval baseline, we evaluate multi-query, hypothetical document, hypothetical question, sparse-dense hybrid retrieval, and chunk compression, and we also propose an integrated method (FUSERAG) that combines these components. Experiments are conducted on a dataset containing 223 documents, 7933 chunks, and 1488 question-answer instances. Results show that the proposed method achieves the highest Mean Reciprocal Rank, nDCG at 5, and recall at 5 at the retrieval level, while the sparse-dense hybrid approach yields the best generation results.
Buğra Şimşek, Korhan Sevinç, Helen Parlar et al.· Signal Processing and Commun...· 0 citations
Retrieval-Augmented Generation (RAG) has established itself as a compelling strategy for grounding large language model outputs in documentary evidence. However, production deployments continue to rely almost exclusively on homogeneous text corpora, even as enterprise repositories grow increasingly heterogeneous blending technical schematics, radiological images, annotated diagrams, and unstructured prose within the same archival system. This mismatch between system design and data reality motivates the present work. We propose a Multimodal RAG framework that unifies text and image retrieval through four tightly coupled components: a dual-stream embedding engine, a learned four-class AI query router, an adaptive confidence threshold, and a session-aware context store. Text is encoded with the allmpnet-base-v2 Sentence Transformer; images are embedded in the same 768-dimensional space via a domain-adapted CLIP ViT-L/14 model fine-tuned on approximately 120,000 technical and clinical text-image pairs. Concatenating 768-d vectors yields a 1,536-d composite query that drives a single approximate nearest-neighbor (ANN) search simultaneously across both modalities. A fine-tuned DistilBERT router assigns each query to one of four retrieval pathways text-only, image-only, hybrid, or conversational at 91.3% accuracy. Experiments were run on three enterprise corpora (engineering manuals, clinical case summaries, and legal paperwork) with the results being a mean F1@5 of 0.90, an increase of 22% over a dense text-only baseline, and a reduction of 49% in the number of hallucinations. The median first-token latency is 1.34s, which meets the interactive-use target deployment-contexts. These results show that it is possible to implement modality-aware retrieval in a simple and realistic setting with real-world organizational constraints and that the results are reproducible.
E.Vijayakumar, Ganesh A· 2026 4th International Confe...· 0 citations
Due to the surge in digital document creation, intelligent systems that can effectively retrieve accurate information from unstructured PDF content have become a necessity. Current keyword-based search techniques lack semantic meaning, while large language models (LLMs), when utilized individually, produce hallucinated results when they are not contextually aware of document content. In this paper, we introduce a novel real-time question answering system for PDF content using a lightweight approach to Retrieval-Augmented Generation (RAG). Our system combines a HuggingFace sentence transformer model for semantic embeddings with a FAISS vector similarity search for efficient retrieval, leveraging a Groq-hosted LLaMA 3.3 70B model for context-grounded answer generation. The system is built using LangChain for orchestration and Streamlit for web-based deployment, and requires no GPU, no fine-tuning, and no pre-indexed corpus. Evaluation across five complementary metrics—Exact Match, Token-level F1, BLEU, ROUGE, and semantic similarity—demonstrates strong alignment between generated and reference answers, with a peak semantic similarity of 0.8709 and a mean Token F1 of 0.4701, confirming the viability of this approach for real-world document intelligence.
Mukesh Lakshmi Sai Medikonda, Kalva Vishnu Teja, K. Greeshma et al.· International Conference Com...· 0 citations
Retrieval-Augmented Generation (RAG) systems face a fundamental trade-off: query expansion can improve retrieval effectiveness for ambiguous or underspecified queries, yet indiscriminate expansion introduces unnecessary latency and retrieval noise. Existing RAG pipelines typically apply expansion uniformly, failing to distinguish between easy and retrieval-challenging queries.To address this issue, we propose DARE-RAG, an adaptive retrieval framework that activates LLM-based query expansion only for retrieval-challenging queries. Our method formulates expansion activation as a lightweight binary classification problem using probe retrieval signals, including score margin, variance, entropy, query length, and lexical specificity. A lightweight MLP predicts whether expansion is likely to improve retrieval quality, and expansion is triggered only when the predicted confidence exceeds a percentile-calibrated threshold.DARE-RAG further integrates a dual-path hybrid retrieval architecture combining BM25 sparse retrieval and BGE dense retrieval, fused via Reciprocal Rank Fusion (RRF), followed by a Cross-Encoder reranker for context refinement. Experiments on NQ-Open and HotpotQA demonstrate that DARE-RAG consistently improves retrieval effectiveness and end-to-end QA accuracy while clearly reducing average end-to-end latency compared with corresponding always-expand variants of BM25, BGE-m3, and their RRF-fused hybrid retriever. Extensive ablation studies and efficiency analyses verify the effectiveness of our utility-guided expansion strategy.
Lixiang Zhu· Poster Volume 0007 The 2026...· 0 citations
Results demonstrate a 157% improvement in retrieval quality over a Naive-RAG baseline, with only 50 ms additional la tency, while Qwen2-VL-2B-Instruct achieved results comparable to cloud-based models in BERTScore, validate that open-source optimized SLMs, paired with advanced retrieval strategies, can provide competitive performance for document understanding without relying on cloud-based models.
