Despite the remarkable progress of Large Language Models (LLMs), their performance in question answering (QA) remains limited by the lack of domain-specific and up-to-date knowledge. Retrieval-Augmented Generation (RAG) addresses this limitation by incorporating external information, often from graph-structured data. However, existing graph-based RAG methods suffer from poor graph quality due to incomplete extraction and insufficient utilization of query information during retrieval. To overcome these limitations, we propose Clue-RAG, a novel approach that introduces (1) a multi-partite graph index that incorporates text Chunk, knowledge unit, and entity to capture semantic content at multiple levels of granularity, coupled with a hybrid extraction strategy that reduces LLM token usage while still producing accurate and disambiguated knowledge units, and (2) Q-Iter, a query-driven iterative retrieval strategy that enhances relevance through semantic search and constrained graph traversal. Experiments on three QA benchmarks show that Clue-RAG significantly outperforms state-of-the-art baselines, achieving 32.55%/12.42% higher average F1 score/Accuracy while delivering $2.2 \times-2.4 \times$ faster offline indexing and $2.7 \times-3.9\times$ lower online query latency compared to strong baseline HippoRAG2. Remarkably, Clue-RAG matches or outperforms baselines even without using an LLM for indexing. These results demonstrate the effectiveness and cost-efficiency of Clue-RAG in advancing graph-based RAG systems. The source code is available in https://github.com/Feesuu/ClueRAG.
Enterprises increasingly rely on Large Language Models (LLMs) for knowledge-driven applications such as clinical question answering, technical support, and legal analysis. However, conventional Retrieval-Augmented Generation (RAG) systems face two major obstacles in enterprise deployment: data privacy concerns when relying on public LLM APIs and high resource requirements for on-premise deployment. This paper presents GalaxyRAG, a privacy-preserving and resourceefficient hybrid GraphRAG framework designed for secure and scalable enterprise reasoning. Built on the distributed graph database Galaxybase, GalaxyRAG unifies knowledge graphs and index graphs within a single retrieval engine, combining symbolic traversal with semantic search. A step-wise reasoning loop enables LLMs to perform multi-hop reasoning without fine-tuning, while a two-stage hybrid evaluation protocol ensures stable and semantically robust performance measurement. Across three multi-hop QA benchmarks and enterprise datasets, GalaxyRAG achieves up to 15-point accuracy gains, while our hybrid evaluation reduces the standard deviation of evaluation scores by over 50%.
Embedding vectors are widely used in various database-related domains. However, efficiently training largescale embeddings directly on DB-resident data remains a challenge, despite the ongoing efforts on supporting machine learning algorithms in database management systems, known as in-DBMS ML. Although current UDAF-based (User-Defined Aggregate Function) approach can support distributed training across parallel DBMS nodes, it enforces strict consistency during model synchronization, leading to high synchronization overhead. Additionally, the incompatibility between DBMS storage architectures and ML workloads further complicates efficient integration. To address these challenges, we propose DistVec, an inDBMS ML framework for efficient distributed training. DistVec introduces multi-query training primitives, that map distributed ML training tasks into concurrent database queries, allowing bounded staleness in model parameters to reduce synchronization costs. Besides, DistVec features an ML-customized model processing pipeline in database, including model materialization, buffer pool eviction, and hotness-aware caching, which minimize disk I/O and remote communication for the ML training workload. Experiments on the Greenplum database show that DistVec achieves an average speedup of 4.53× over UDAF-based training approaches while preserving the model quality.
Deepfakes, visual content created or modified by AI algorithms are increasingly being used to spread misinformation. Consequently, social media platforms are now labelling AI-content. To implement deepfake detectors in a way that permits user acceptance and trust in the decision, it is important to provide explanations of the model decision that align with user preferences. In order to better understand these user preferences, we will compare user preferences along the dimensions of acceptance and trust between (C1) just the output label, (C2) model output + feature importance mask over regions of the image that are salient to model output, (C3) model output + natural language explanation of image decision, and (C4) model output + mask + natural language explanation. Preferences will be measured using Likert scale and statistical significance performed via Friedman test. We find that the presence of a natural language component (C3 and C4) produces clearer explanations, and increases user trust and acceptance over the baseline (C1) and mask (C2).
To deter SNS users from posting or reposting misinformation/disinformation, we develop a method to quantify each user’s degree of responsibility when misinformation/disinformation spreads. We utilize the Banzhaf index, a technique for quantitatively evaluating player contributions in participatory games, for this purpose. Although the computation of the Banzhaf index has exponential time complexity with respect to the number of users, we developed an algorithm computable in realistic time by limiting it to tree structures modeling information propagation in social networks. We also discuss challenges for practical implementation of the proposed method and possibilities for extending the current work.
