AquiLLM: Evaluating Faithfulness in Open-Weight RAG-LLM Systems for Scientific Research
本文评估了AquiLLM在科学研究中对开放权重RAG-LLM系统的忠实度，通过天文学案例研究了其在检索和分析任务中的表现。
0 citationsRead paper
Selected work
Representative Papers
AquiLLM: An Architecture for Supporting Tacit Knowledge Capture in Research Groups
This work addresses the challenges research teams face when using commercial AI systems—namely, insufficient transparency, poor reproducibility, and privacy risks—which hinder effective capture and utilization of tacit knowledge. To overcome these limitations, the authors propose an open-source, modular RAG-LLM framework built upon open-weight large language models, enabling local deployment. The framework integrates local embedding and re-ranking, multimodal processing, semantic and contextual memory mechanisms, and extensible skill modules, while offering an OpenAI-compatible API and an optimized user interface. Expert evaluations in domains such as astrophysics and environmental science demonstrate that the system substantially enhances the effectiveness of tacit knowledge capture and improves the trustworthiness, adaptability, and usability of AI tools in scientific research contexts.
0 citationsRead paper
GreenCrossingAI: A Camera Trap/Computer Vision Pipeline for Environmental Science Research Groups
Wildlife camera-trap data processing faces challenges including massive data volume, low annotation accuracy, strong environmental interference, difficulty integrating AI tools, and constrained computational resources. To address these, this study proposes a lightweight, localized, end-to-end AI processing pipeline. Methodologically, it integrates lightweight deep learning models (e.g., YOLOv5s), edge computing architecture, and automated image annotation, enabling heterogeneous device deployment and offline operation. A domain-adapted workflow is designed to support efficient image transmission, on-device inference, species identification, and structured data management. The key contribution is the first cloud-free, low-compute, highly robust end-to-end solution tailored for resource-constrained research teams. Experiments demonstrate significant reductions in annotation cost and processing latency, with a 12.3% improvement in species identification accuracy. The system exhibits strong practicality and scalability across diverse field terrains.
0 citationsRead paper
Recent publications
Latest Papers
AquiLLM: Evaluating Faithfulness in Open-Weight RAG-LLM Systems for Scientific Research
本文评估了AquiLLM在科学研究中对开放权重RAG-LLM系统的忠实度，通过天文学案例研究了其在检索和分析任务中的表现。
0 citationsRead paper
AquiLLM: An Architecture for Supporting Tacit Knowledge Capture in Research Groups
This work addresses the challenges research teams face when using commercial AI systems—namely, insufficient transparency, poor reproducibility, and privacy risks—which hinder effective capture and utilization of tacit knowledge. To overcome these limitations, the authors propose an open-source, modular RAG-LLM framework built upon open-weight large language models, enabling local deployment. The framework integrates local embedding and re-ranking, multimodal processing, semantic and contextual memory mechanisms, and extensible skill modules, while offering an OpenAI-compatible API and an optimized user interface. Expert evaluations in domains such as astrophysics and environmental science demonstrate that the system substantially enhances the effectiveness of tacit knowledge capture and improves the trustworthiness, adaptability, and usability of AI tools in scientific research contexts.
0 citationsRead paper
GreenCrossingAI: A Camera Trap/Computer Vision Pipeline for Environmental Science Research Groups
Wildlife camera-trap data processing faces challenges including massive data volume, low annotation accuracy, strong environmental interference, difficulty integrating AI tools, and constrained computational resources. To address these, this study proposes a lightweight, localized, end-to-end AI processing pipeline. Methodologically, it integrates lightweight deep learning models (e.g., YOLOv5s), edge computing architecture, and automated image annotation, enabling heterogeneous device deployment and offline operation. A domain-adapted workflow is designed to support efficient image transmission, on-device inference, species identification, and structured data management. The key contribution is the first cloud-free, low-compute, highly robust end-to-end solution tailored for resource-constrained research teams. Experiments demonstrate significant reductions in annotation cost and processing latency, with a 12.3% improvement in species identification accuracy. The system exhibits strong practicality and scalability across diverse field terrains.
0 citationsRead paper