-
Research Article
Evaluating Precision and Recall at Retrieval Time in Retrieval-Augmented Generation (RAG) Systems
Gopichand Agnihotram*
,
Joydeep Sarkar
Issue:
Volume 8, Issue 4, December 2025
Pages:
174-180
Received:
12 September 2025
Accepted:
23 September 2025
Published:
18 October 2025
Abstract: Retrieval-Augmented Generation (RAG) systems signify a pivotal advancement in natural language processing, merging information retrieval with large language models (LLMs) to ground responses in external knowledge. This hybrid approach enhances the factual accuracy and currency of generated content, mitigating common issues like hallucination. The efficacy of a RAG system, however, is fundamentally dependent on the performance of its retrieval component. This paper provides a detailed analysis of precision and recall as critical metrics for evaluating and optimizing this retrieval step. We explore the distinct roles and inherent trade-offs of these metrics within a RAG pipeline, demonstrating their direct influence on the quality of the final output. Through a series of experiments comparing sparse (BM25), dense (DPR), and hybrid retrieval methods, we quantify their performance characteristics. The analysis is further enriched with real-world examples from finance, law, and healthcare, illustrating the practical implications of retrieval quality. Additionally, we outline advanced strategies for improving retrieval effectiveness, such as multi-stage architecture involving rerankers and the use of query transformations. The paper concludes with a set of best practices for deploying robust, enterprise-grade RAG systems, emphasizing the need for continuous evaluation and sophisticated retrieval strategies. By focusing on the systematic optimization of precision and recall, organizations can build more reliable and trustworthy AI applications.
Abstract: Retrieval-Augmented Generation (RAG) systems signify a pivotal advancement in natural language processing, merging information retrieval with large language models (LLMs) to ground responses in external knowledge. This hybrid approach enhances the factual accuracy and currency of generated content, mitigating common issues like hallucination. The e...
Show More
-
Research Article
Beyond Static Retrieval: A Reinforcement Learning Framework for Dynamic and Adaptive RAG
Gopichand Agnihotram,
Joydeep Sarkar,
Magesh Kasthuri*
Issue:
Volume 8, Issue 4, December 2025
Pages:
181-188
Received:
12 September 2025
Accepted:
23 September 2025
Published:
18 October 2025
Abstract: Retrieval-Augmented Generation (RAG) is a widely adopted technique that enhances large language models (LLMs) by grounding their outputs in external knowledge sources. This approach reduces hallucinations, increases factual accuracy, and adapts well to rapidly evolving domains. Despite these strengths, traditional RAG implementations rely on static, heuristic-based retrieval strategies that operate independently of feedback or contextual learning. In today’s fast-changing information landscape, it’s crucial for language models to go beyond static retrieval when grounding their responses. That’s where a RL framework comes into play for RAG. Rather than sticking to fixed, rule-based selection methods, RL allows the retrieval component to learn and adapt over time—much like how a person refines their search strategies with experience and feedback. By framing the process of document selection as a Markov Decision Process (MDP), the system can make context-aware choices that consider both immediate and future gains. This white paper explores how Retrieval-Augmented Generation can be significantly enhanced by integrating Markov Decision Processes (MDPs) and Reinforcement Learning (RL). We present a conceptual framework that models retrieval as a sequential decision-making problem. By treating document selection as an MDP and employing RL algorithms to optimize retrieval strategies, we introduce adaptivity, context sensitivity, and long-term reasoning into the RAG pipeline, leading to demonstrably more accurate and relevant generated content. The paper also outlines applications, implementation strategies, and future research directions that combine symbolic and neural methods for improved decision-making and document relevance.
Abstract: Retrieval-Augmented Generation (RAG) is a widely adopted technique that enhances large language models (LLMs) by grounding their outputs in external knowledge sources. This approach reduces hallucinations, increases factual accuracy, and adapts well to rapidly evolving domains. Despite these strengths, traditional RAG implementations rely on static...
Show More
-
Research Article
An Integrated Jacket–Helmet Assistive System for Visually Impaired Individuals Using YOLO-Based Object Detection, Depth Estimation, and OCR
Kashvi Ruparelia,
Priyam Parikh*,
Parth Atulkumar Shah
Issue:
Volume 8, Issue 4, December 2025
Pages:
189-205
Received:
12 September 2025
Accepted:
23 September 2025
Published:
30 October 2025
Abstract: This paper presents the design and evaluation of a jacket–helmet assistive system for visually impaired individuals in India. The system integrates a Raspberry Pi 4B with a USB web camera, USB microphone, vibration motor cluster, earphone, pushbuttons, and a rechargeable 7.4 V, 10,000 mAh battery. Two primary functions are implemented: (i) object detection and distance estimation using YOLO algorithms with 2D depth estimation, and (ii) text recognition on posters and hoardings using optical character recognition (OCR). Comparative analysis of YOLOv5, YOLOv7, and YOLOv8 models demonstrated that YOLOv8 achieved the highest mean Average Precision (mAP) of 92.4%, outperforming YOLOv7 (89.6%) and YOLOv5 (87.3%). For monocular 2D depth estimation, MiDaS achieved the lowest mean absolute relative error (0.124) compared to Monodepth2 (0.156) and DPT (0.139). Speech-to-text efficiency was tested across Google Speech Recognition, Vosk, and CMU Sphinx, with Google achieving 94.1% accuracy, followed by Vosk (88.3%) and CMU Sphinx (81.6%). User trials were conducted with ten visually impaired individuals across diverse environments (bus stand, garden, bungalow, and home settings). System usability was measured using the System Usability Scale (SUS), yielding an overall average score of 84.6, indicating “excellent” usability. The proposed system demonstrates high accuracy, robustness, and practicality for real-world navigation and reading assistance, thus contributing to improved autonomy and quality of life for visually impaired users.
Abstract: This paper presents the design and evaluation of a jacket–helmet assistive system for visually impaired individuals in India. The system integrates a Raspberry Pi 4B with a USB web camera, USB microphone, vibration motor cluster, earphone, pushbuttons, and a rechargeable 7.4 V, 10,000 mAh battery. Two primary functions are implemented: (i) object d...
Show More
-
Research Article
Image Reconstruction in Compressive Sensing Using The Level 4 Daubechies 4 (db4) Discrete Wavelet Transform And SP, CoSaMP and ALISTA Algorithm
Issue:
Volume 8, Issue 4, December 2025
Pages:
206-213
Received:
1 October 2025
Accepted:
14 October 2025
Published:
31 October 2025
Abstract: This work proposes an efficient image reconstruction method based on compressive sensing (CS), combining the level-4 Daubechies 4 (db4) discrete wavelet transform with three iterative reconstruction algorithms: Subspace Pursuit (SP), Compressive Sampling Matched Pursuit (CoSaMP), and Analytic Learned Iterative Shrinkage Thresholding Algorithm (ALISTA). The approach follows four key steps: (1) decomposing the original image via the db4 wavelet transform to obtain a sparse representation, (2) performing compressed sampling using a random measurement matrix, (3) reconstructing the sparse signal from the reduced measurements using one of the three optimization algorithms, and (4) recovering the final image through the inverse wavelet transform. Experimental evaluation uses the standard Lena image (200 × 200 pixels) and compares the performance of the three algorithms according to two criteria: reconstruction quality (measured by SSIM) and computational cost (reconstruction time in minutes), across sampling rates ranging from 10% to 60%. Results show that all three methods achieve very similar SSIM scores (up to >0.96 at 60%), indicating high structural fidelity regardless of the algorithm chosen. However, ALISTA stands out significantly for its temporal efficiency, particularly at low sampling rates (<0.1 minute at 10%), while CoSaMP exhibits high and unstable computation times (peaking at ~35 minutes at 40%). SP offers a stable, nearly linear increase in runtime but remains consistently slower than ALISTA. These results demonstrate that ALISTA provides the best trade-off between quality and speed. Thus, this study validates the value of coupling the db4 wavelet basis with modern, learned optimization algorithms for practical CS applications in image processing, where computational efficiency is as critical as reconstruction accuracy.
Abstract: This work proposes an efficient image reconstruction method based on compressive sensing (CS), combining the level-4 Daubechies 4 (db4) discrete wavelet transform with three iterative reconstruction algorithms: Subspace Pursuit (SP), Compressive Sampling Matched Pursuit (CoSaMP), and Analytic Learned Iterative Shrinkage Thresholding Algorithm (ALIS...
Show More
-
Research Article
Hybrid and Unitary PEFT for Resource-Efficient Large Language Models
Haomin Qi*,
Zihan Dai,
Chengbo Huang
Issue:
Volume 8, Issue 4, December 2025
Pages:
242-255
Received:
1 October 2025
Accepted:
20 October 2025
Published:
19 December 2025
DOI:
10.11648/j.ajcst.20250804.17
Downloads:
Views:
Abstract: Fine-tuning large language models (LLMs) remains a computational bottleneck due to their scale and memory demands. This paper presents a comprehensive evaluation of parameter-efficient fine-tuning (PEFT) techniques, including LoRA, BOFT, LoRA-GA, and uRNN, and introduces a novel hybrid strategy that dynamically integrates BOFT’s orthogonal stability with LoRA-GA’s gradient-aligned rapid convergence. By computing per-layer adaptive updates guided by gradient norms, the hybrid method achieves superior convergence efficiency and generalization across diverse tasks. We also explore, for the first time, the adaptation of unitary RNN (uRNN) principles to Transformer-based LLMs, enhancing gradient stability through structured unitary constraints. Across GLUE, GSM8K, MT-Bench, and HumanEval with models from 7B to 405B, the hybrid approach yields consistent gains across three independent runs per task and model, approaching the quality of full fine-tuning while reducing training time by about 2.1 × and peak memory by nearly 50%, indicating practical significance under resource constraints. A compact multilingual and low-resource study on XNLI and FLORES with 32 examples per language shows consistent gains under the same budget with a small, stable footprint. These results indicate a practical and scalable path to accessible LLM fine-tuning under resource constraints.
Abstract: Fine-tuning large language models (LLMs) remains a computational bottleneck due to their scale and memory demands. This paper presents a comprehensive evaluation of parameter-efficient fine-tuning (PEFT) techniques, including LoRA, BOFT, LoRA-GA, and uRNN, and introduces a novel hybrid strategy that dynamically integrates BOFT’s orthogonal stabilit...
Show More
-
Research Article
Image Reconstruction in Compressive Sensing Using Daubechies 7 (db7) and Lifting Wavelet Transforms with SP, CoSaMP, and ALISTA Algorithms
Issue:
Volume 8, Issue 4, December 2025
Pages:
214-227
Received:
20 October 2025
Accepted:
3 November 2025
Published:
9 December 2025
Abstract: This paper proposes an efficient image reconstruction method in compressive sensing (CS) that combines the Lifting Wavelet Transform (LWT) with the Daubechies 7 (db7) wavelet and three reconstruction algorithms: Subspace Pursuit (SP), Compressive Sampling Matching Pursuit (CoSaMP), and the Analytic Learned Iterative Shrinkage Thresholding Algorithm (ALISTA). Unlike the conventional Discrete Wavelet Transform (DWT), whose implementation relies on computationally expensive convolution operations, the LWT enables a faster sparse representation while preserving the sparsity properties essential for reconstruction. The proposed methodology is based on a key observation: among the four subbands generated by the LWT: approximation (CA) and detail coefficients (LH, HL, HH) only the latter three are inherently sparse. Consequently, compression is applied exclusively to these detail components, while the approximation subband is kept intact, thereby preserving critical low-frequency information. Experiments were conducted on two types of images a natural image (Lena) and a medical image (MRI) across various resolutions (from 200×200 to 512×512 pixels) and sampling rates (from 10% to 80%). Performance was evaluated using the Structural Similarity Index (SSIM) and reconstruction time. Results consistently show that ALISTA significantly outperforms SP and CoSaMP in both reconstruction quality and speed. At 80% sampling, ALISTA achieves an SSIM of 0.9936 for Lena and 0.9764 for MRI, compared to approximately 0.975 and 0.934 for the other methods, respectively. Moreover, ALISTA maintains extremely low reconstruction times under 4 seconds even for 512×512-pixel images. This research confirm that the ALISTA + LWT/db7 combination achieves the best quality–speed trade-off and exhibits robustness regardless of image type or size.
Abstract: This paper proposes an efficient image reconstruction method in compressive sensing (CS) that combines the Lifting Wavelet Transform (LWT) with the Daubechies 7 (db7) wavelet and three reconstruction algorithms: Subspace Pursuit (SP), Compressive Sampling Matching Pursuit (CoSaMP), and the Analytic Learned Iterative Shrinkage Thresholding Algorithm...
Show More
-
Review Article
Systematic Review of Models Examining Factors Influencing SaaS Adoption in Higher Education Institutions
Issue:
Volume 8, Issue 4, December 2025
Pages:
228-241
Received:
13 October 2025
Accepted:
28 October 2025
Published:
19 December 2025
DOI:
10.11648/j.ajcst.20250804.16
Downloads:
Views:
Abstract: Since the early 2000s, Higher Education Institutions (HEIs) have increasingly adopted Software as a Service (SaaS). Despite benefits such as improved collaboration, efficiency, and reduced costs, adoption remains low in many developing regions, particularly Sub-Saharan Africa. Most research emphasizes organizational-level adoption, especially in East Africa, with limited focus on individual user factors. Using the PRISMA protocol, this review analyzed 34 studies from 2015–2025 sourced from Google Scholar, ScienceDirect, SpringerLink, and IEEE. The review examines factors influencing individual SaaS adoption in HEIs and evaluates three key models: TAM, UTAUT, and TRI, highlighting their strengths and limitations. Results show that UTAUT explains up to 70% of behavioral intention variance, compared to 40–60% for TAM. UTAUT expands TAM with additional constructs, while TRI is most effective when integrated with other frameworks. Key variables influencing user adoption were identified, offering insights to enhance individual uptake of cloud technologies in HEIs. Many studies apply TAM and UTAUT without contextual adaptation. Core constructs such as Perceived Usefulness (PU–39%), Perceived Ease of Use (PEU–94%), Facilitating Conditions (FC–50%), Social Influence (SI–44%), and Effort Expectancy (EE–50%) show higher significance than local/ External factors like skill transfer (6%), transition to e-learning (6%), and institutional policies (6%). Security and price emerge as the most prominent contextual factors that are addressed in roughly 39% and 17% of the studies respectively, this highlights their relative influence on technology adoption. Regionally, 80% of studies originate from Asia, reflecting strong academic interest and rapid digital growth. The Americas (20%) show less focus, possibly due to maturity in cloud adoption. Europe (40%) exhibits moderate engagement, while the Middle East/West Asia and Africa (30% each) show emerging research hindered by infrastructure, trust, and capacity challenges, leading to more emphasis on organizational adoption. The review identifies key variables shaping user adoption, offering insights to strengthen individual uptake of cloud technologies in HEIs across the region.
Abstract: Since the early 2000s, Higher Education Institutions (HEIs) have increasingly adopted Software as a Service (SaaS). Despite benefits such as improved collaboration, efficiency, and reduced costs, adoption remains low in many developing regions, particularly Sub-Saharan Africa. Most research emphasizes organizational-level adoption, especially in Ea...
Show More
