A Hybrid RAG-LLaMA Framework for Scalable and Accurate Interpretation of Legal Texts

ABSTRACT Interpreting Indian tax law is challenging due to frequent legislative amendments, complex statutory structures, and the requirement for precise citation. This paper presents TaxFlow, a domain-specific legal assistant that integrates a hybrid Retrieval-Augmented Generation (RAG) framework with the LLaMA-3 language model for statutory question answering. The system is trained on a curated corpus of Indian tax statutes, government gazettes, and case law, processed through structured extraction, segmentation, and dense–sparse indexing. TaxFlow incorporates temporal validity filtering to ensure that retrieved provisions reflect the legally effective version at query time. The architecture combines FAISS-based retrieval, legal-domain adapters, and critique-driven generation to reduce hallucinations and improve citation fidelity. Evaluation is conducted on a unified Indian tax law benchmark using automatic metrics. Experimental results show that TaxFlow achieves 96.5% accuracy and a 95.0% F1-score, demonstrating consistent and substantial improvements over representative baseline systems, while BLEU and ROUGE-L scores demonstrate significant improvements in linguistic quality. The findings confirm the effectiveness of domain-adapted hybrid RAG systems for reliable and scalable legal assistance. Introduction Interpreting and applying tax law in India remains a complex undertaking due to frequent legislative amendments, intricate drafting, and the layered nature of statutory provisions. These characteristics create barriers not only for legal practitioners but also for auditors, tax professionals, and policymakers who require reliable interpretations for decision-making. Traditional retrieval systems – such as static databases and rule-based expert systems – offer structured access to documents but lack adaptability to evolving contexts, often resulting in incomplete or ambiguous outcomes. Consequently, there is a pressing need for advanced computational approaches that can deliver precise and context-aware interpretations at scale. Recent advances in artificial intelligence (AI) and natural language processing (NLP) provide a viable pathway toward addressing these challenges. Large Language Models (LLMs), when coupled with retrieval-augmented generation (RAG), enable dynamic consultation of authoritative corpora before generating responses, thereby enhancing factual reliability and reducing hallucinations (Lewis et al. Citation2021). Foundational contributions, including RAG (Lewis et al. Citation2021), ATLAS (Izacard et al. Citation2022), and RETRO (Borgeaud et al. Citation2022), established the benefits of retrieval–generation integration. Subsequent innovations, such as Self-RAG (Shi et al. Citation2023) and IRCoT (Trivedi et al. Citation2023), demonstrated improved robustness in multi-hop reasoning. Benchmarking efforts (Muennighoff et al. Citation2023) have reinforced the transformative potential of these models, though challenges concerning latency, retrieval fidelity, and scalability remain unresolved. In parallel, the legal domain has increasingly adopted NLP methods, supported by benchmark datasets such as LexGLUE (Chalkidis et al. Citation2021), Pile of Law (Henderson et al. Citation2022), and LegalBench (Holzenberger et al. Citation2022). Surveys highlight progress in classification, statutory interpretation, and case reasoning (Ariai, Mackenzie, and Demartini Citation2025; Quevedo et al. Citation2022). Nonetheless, jurisdictional imbalances persist, with most benchmarks focused on Western legal systems and limited resources available for Indian contexts (Sharma, Iyer, and Srinivasan Citation2024; Xu et al. Citation2023). Furthermore, explainability and citation fidelity remain open challenges in legal AI (Richmond Citation2024). Applied works, such as CaseLaw-RAG (Kshitij, Gupta, and Singh Citation2024) and RAG-Fusion (Rackauckas Citation2024), have shown progress in citation reliability; however, their scope remains narrow, and evaluations lack consistency. In response to these limitations, this study introduces TaxFlow, an AI-driven legal assistant explicitly designed for Indian tax law. The system integrates LLaMA-3 (Meta Citation2024) with a hybrid RAG pipeline, orchestrated using LangChain, Hugging Face inference tools, and FAISS-based vector indexing. Legal sources, including statutory gazettes, government notifications, and case law, are processed through a rigorous pipeline of extraction, cleaning, segmentation, and embedding, enabling scalable and context-sensitive responses. This modular design ensures adaptability to evolving statutory frameworks while delivering outputs that are legally coherent and practitioner-ready. The novelty of this work lies in its hybrid retrieval framework, tailored to Indian tax law, which combines dense and sparse retrieval with domain-specific legal adapters. Unlike prior RAG systems, which were primarily tested on generic datasets (Borgeaud et al. Citation2022; Izacard et al. Citation2022; Lewis et al. Citation2021), TaxFlow demonstrates enhanced interpretability and accuracy for jurisdiction-specific statutory QA. Additionally, the study situates the proposed framework within a comprehensive qualitative review of recent retrieval-augmented and legal NLP systems. It conducts an experimental evaluation using a unified benchmark (2021–2025) across multiple performance metrics, in which TaxFlow achieved 96.5% accuracy and 95.0% F1-score, surpassing state-of-the-art methods by a significant margin. By bridging methodological innovation with comprehensive benchmarking (Ajay Mukund and Easwarakumar Citation2025; Ariai, Mackenzie, and Demartini Citation2025; Hou et al. Citation2025; Quevedo et al. Citation2022; Richmond Citation2024), this work contributes to both the practical design of domain-specific legal assistants and the scholarly understanding of retrieval-augmented frameworks in law. Unlike prior legal question-answering systems that are evaluated on heterogeneous datasets or Western-centric benchmarks, this work focuses exclusively on Indian tax law and adopts a unified evaluation protocol. The proposed framework emphasizes jurisdiction-specific corpus curation, hybrid dense–sparse retrieval, and temporal awareness to ensure statutory validity. These design choices directly address limitations identified in earlier retrieval-augmented legal systems, particularly with respect to reproducibility, citation accuracy, and the evolution of legislation. Literature Survey The development of retrieval-augmented generation (RAG) marked a turning point in knowledge-intensive NLP by integrating retrieval mechanisms into generative transformers, thereby strengthening factual grounding (Lewis et al. Citation2021). Building on this, ATLAS (Izacard et al. Citation2022) advanced few-shot learning, while RETRO (Borgeaud et al. Citation2022) demonstrated scalable retrieval for large language models. Recent efforts, such as Self-RAG (Shi et al. Citation2023) and IRCoT (Trivedi et al. Citation2023), have further improved robustness in multi-hop reasoning tasks. Benchmarking initiatives, including MTEB (Muennighoff et al. Citation2023) and BGE embeddings (B. A. A. I. Research Citation2023), have provided standardized evaluation frameworks; however, retrieval accuracy and latency remain bottlenecks (Tang et al. Citation2022). Within the legal domain, systematic reviews (Ariai, Mackenzie, and Demartini Citation2025; Quevedo et al. Citation2022) document significant progress across tasks such as classification, summarization, and judgment prediction. Benchmarks such as LexGLUE (Chalkidis et al. Citation2021), Pile of Law (Henderson et al. Citation2022), and LegalBench (Holzenberger et al. Citation2022) provide standardized evaluation corpora, while LAW-MT5 (Xu et al. Citation2023) extends coverage to multilingual legal systems. Applied works such as CaseLaw-RAG (Kshitij, Gupta, and Singh Citation2024), Contract-RAG (Zhang and Xiao Citation2023), and RAG-Fusion (Rackauckas Citation2024) demonstrate the practical potential of retrieval pipelines for improving citation accuracy. Yet persistent limitations remain, including inconsistent evaluation metrics, limited jurisdictional coverage, and recurring citation errors (Sharma, Iyer, and Srinivasan Citation2024). The advent of advanced LLMs, such as LLaMA-2 (Touvron et al. Citation2023), LLaMA-3 (Meta Citation2024), and GPT-4 (OpenAI Citation2023), has improved reasoning and multilingual performance, although their factual reliability depends on retrieval integration. Efficiency techniques – such as LLM.int8 (Dettmers et al. Citation2022) and QLoRA (Frantar, Ashkboos, and Alistarh Citation2023) – have made fine-tuning feasible on constrained hardware, while evaluation frameworks, including RARR (Shuster et al. Citation0000), attribution metrics (Rashkin et al. Citation2023), and RAGAS (Es, Sharma, and Chen Citation2023), prioritize faithfulness. Nonetheless, surveys indicate that domain-sensitive evaluation for legal QA, particularly for citation correctness and interpretability, remains underdeveloped (Richmond Citation2024). Infrastructure-level advances have also played a role. Vector search engines such as Weaviate (Formal et al. Citation2021) and Milvus (Z. Research Citation2022) support large-scale retrieval pipelines, while hybrid sparse–dense approaches (Boytsov, Lin, and Ma Citation2022; Lin and Ma Citation2021) have improved robustness in heterogeneous corpora. Legal applications of RAG include compliance monitoring (Ramakrishnan, Sen, and Rao Citation2022), contract interpretation (Zhang and Xiao Citation2023), judgment prediction (Zhong et al. Citation2023), and statutory QA (Sharma, Iyer, and Srinivasan Citation2024). Human-in-the-loop evaluation frameworks (Aggarwal, Banerjee, and Singh Citation2025) have emerged to address the limitations of automated metrics. From this body of work, four apparent gaps emerge. First, jurisdiction-specific corpora – especially in Indian law – remain underdeveloped (Sharma, Iyer, and Srinivasan Citation2024). Second, retrieval fidelity and citation correctness remain challenges (Kshitij, Gupta, and Singh Citation2024; Rackauckas Citation2024). Third, comprehensive evaluations across accuracy, precision, recall, and F1-score are rare, limiting comparability across systems. Finally, scalable hybrid retrieval pipelines that integrate dense and sparse methods with domain-adapted generators remain underexplored (Richmond Citation2024; Zhang and Xiao Citation2023). This study addresses these gaps by presenting TaxFlow, a system tailored for Indian statutory question answering, benchmarked against various peer systems, and demonstrating superior performance across all primary evaluation metrics. provides a qualitative overview of representative work in retrieval-augmented generation and legal NLP, highlighting datasets, methodological approaches, key contributions, and remaining limitations. This comparison is intended to contextualize the proposed work rather than to provide a quantitative performance benchmark. Since these studies were evaluated under heterogeneous datasets and experimental protocols, direct quantitative comparison across them is not methodologically valid; therefore, performance benchmarking is conducted separately under a unified evaluation setting in Section 6. Methodology The TaxFlow system uses advanced algorithms to interpret complex legal texts and generate accurate, context-aware responses. The choice of methods is motivated by their demonstrated performance in natural language processing (NLP) and information retrieval, as well as their suitability for handling the intricate nuances of legal language. At its core lies the fine-tuned LLaMA-3 model, a transformer-based architecture that has become the benchmark in NLP. The model’s self-attention mechanism captures long-range dependencies and semantic relationships, which are essential for understanding the technical vocabulary and layered structures of Indian tax law. To further strengthen generation, TaxFlow incorporates a Retrieval-Augmented Generation (RAG) framework, which couples generative reasoning with retrieval from external sources. During inference, the model dynamically fetches supporting evidence from a vector database, ensuring responses remain grounded in authoritative legal texts. The FAISS library underpins this retrieval process by implementing state-of-the-art approximate nearest neighbor (ANN) algorithms, enabling efficient searches across large-scale legal corpora. This design ensures rapid access to relevant statutes, case law, and notifications, even under high query volumes. The LLaMA (Large Language Model Meta AI) architecture shown in is based on the transformer framework, but introduces several optimizations to improve scalability and efficiency: Input and Embeddings Raw tokens from user input are first converted into dense vector embeddings. Rotary positional encodings (RoPE) are applied to inject sequence order information into embeddings, ensuring the model can capture word order dependencies. Transformer Layers (Nx stacked blocks) Each transformer block repeats N times to increase model depth. Self-Attention (Grouped Multi-Query Attention with KV Cache): Instead of traditional multi-head attention, LLaMA adopts a grouped multi-query attention mechanism. This reduces memory usage and improves inference speed, while the KV cache accelerates decoding by storing previously computed key-value pairs. o Feed Forward Network (SwiGLU): The standard feed-forward layers are replaced with SwiGLU activation, which improves expressiveness and training stability. o Residual Connections & Normalization (RMS Norm): Residual shortcuts ensure stable gradient flow. LLaMA uses RMS Normalization instead of LayerNorm, providing better performance for very large models. 3. Output Layer The top of the model consists of a linear projection layer followed by a softmax function to generate probability distributions over the vocabulary. This allows the model to produce the following word/token in the sequence. The orchestration of retrieval and generation is managed through LangChain, which provides a modular workflow to integrate multiple components seamlessly, as shown in , RAG pipeline. For inference and fine-tuning, the system leverages the Hugging Face ecosystem, enabling scalable deployment and alignment with domain-specific requirements. Together, these technologies ensure that generated responses are both factually accurate and legally consistent. By combining transformer-based language models, retrieval-augmented pipelines, and high-performance vector search, TaxFlow delivers a scalable solution for real-time legal question answering. The architecture shown in addresses not only the inherent complexity of statutory interpretation but also enhances precision, transparency, and contextual reliability in responses, thereby providing meaningful support to tax professionals, policymakers, and legal practitioners. The proposed system, TaxFlow, introduces a hybrid retrieval-augmented generation (RAG) framework designed explicitly for statutory question answering in the Indian legal domain, as illustrated in Figure TaxFlow. Unlike generic RAG implementations that rely on large open-domain datasets, our approach integrates domain-adapted retrieval, specialized legal embeddings, and a critique-driven generation module to ensure both factual accuracy and legal faithfulness. The methodological pipeline is structured into the following key components. This methodology is designed to ensure that the system can accurately interpret complex legal texts and provide context-aware responses in real time. Corpus Construction and Preprocessing A dedicated corpus was curated from statutory texts, government gazettes, and case law repositories. Each document underwent a structured preprocessing pipeline consisting of text normalization, removal of extraneous metadata (headers, footers, and OCR noise), and token-level segmentation into overlapping passages of approximately 1000 tokens. This ensures manageable context length while preserving cross-references within statutes. Metadata, such as the Act name, section, and date of publication, were retained to support precise attribution. Each document chunk retains metadata, including Act name, section number, source identifier, publication date, and amendment validity period. This metadata enables temporal filtering during retrieval, ensuring that only legally effective provisions are considered at query time. Hybrid Retrieval Mechanism To overcome the limitations of single-mode retrieval, a hybrid strategy combining dense and sparse indexing was employed. Dense retrieval: Documents were embedded using domain-tuned BGE embeddings, allowing semantic similarity search via FAISS indexing. Sparse retrieval: A BM25/SPLADE-based inverted index complemented the dense retrieval by capturing exact legal terminology, abbreviations, and statutory phrases. A fusion re-ranking layer based on a cross-encoder ensured that retrieved passages reflected both semantic relevance and statutory fidelity. This hybrid design improves recall in cases where legal wording is highly technical or context-dependent. Retrieval results are filtered and re-ranked based on both semantic relevance and statutory validity, ensuring alignment with the applicable legal version at the time of the query. Generation and Legal Adapters For answer generation, a transformer-based large language model (LLaMA-3) was integrated with legal adapters fine-tuned on statutory QA datasets. Retrieved passages were passed into the generator with structured prompts that explicitly requested citations to relevant Acts and sections. To mitigate hallucinations, a self-critique module, inspired by the RARR and RAGAS frameworks, evaluated each draft response against retrieved evidence. If discrepancies were detected, the generator re-invoked with refined context until a factually consistent answer was produced. Attribution and Citation Verification Beyond conventional QA, the system ensures transparent attribution. Extracted answers are post-processed to identify cited sections and cross-validated against the retrieval set. A citation alignment mechanism ensures that quoted text in the response directly corresponds to the verified statutory passages, thereby enhancing the reviewer’s confidence in the legal reliability of the response. Evaluation Strategy Evaluation Methodology The evaluation of TaxFlow is designed to ensure methodological rigor, reproducibility, and domain relevance. Unlike prior studies that compare results across heterogeneous datasets, all experiments are conducted using a single benchmark corpus and uniform evaluation settings. Dataset The evaluation corpus comprises Indian tax statutes, government gazettes, notifications, and selected case law spanning 1995 to 2025. After preprocessing and segmentation, the corpus contains approximately 85,000 documents. The dataset is divided into training (70%), validation (10%), and test (20%) splits, with no document overlap across partitions. Task Definition The task is statutory question answering, requiring systems to retrieve relevant legal provisions and generate grounded responses that accurately interpret applicable sections of Indian tax law. Baselines TaxFlow is compared against representative baselines, including keyword-based retrieval, dense-only RAG, sparse-only retrieval with language models, and generic hybrid RAG systems. All baselines are evaluated under identical constraints using the same dataset and metrics. Metrics Performance is measured using accuracy, precision, recall, F1-score, BLEU, and ROUGE-L. These metrics evaluate both legal correctness and linguistic quality. Results On the held-out test set, TaxFlow achieves 96.5% accuracy and an F1 Score of 95.0%, demonstrating consistent and substantial improvements over representative baseline systems. BLEU and ROUGE-L scores show notable improvements, indicating enhanced response coherence and legal phrasing. These results demonstrate the effectiveness of hybrid retrieval combined with domain-specific adaptation for statutory interpretation. summarizes TaxFlow’s performance against representative baseline systems, evaluated under identical conditions on the Indian Tax Law QA benchmark. The results indicate that while generic hybrid RAG approaches deliver strong performance, TaxFlow achieves substantial improvements across all metrics. In particular, the 10–15% gain in F1-score highlights the effectiveness of combining hybrid retrieval, domain-specific adaptation, and temporal filtering for statutory interpretation. All baseline methods were re-implemented and evaluated by the authors on the same dataset under identical experimental settings. Discussion The results in highlight clear performance differences between representative retrieval-based baselines and the proposed TaxFlow framework under identical experimental conditions. Keyword-based retrieval, while effective for exact-term matching, demonstrates limited contextual understanding, resulting in lower overall accuracy. Dense-only and sparse-only retrieval approaches improve semantic coverage but struggle to balance recall and precision when dealing with highly structured statutory language. Generic hybrid RAG methods offer a stronger baseline by combining lexical and semantic retrieval; however, their performance remains constrained by the absence of legal-domain adaptation and temporal awareness. In contrast, TaxFlow consistently outperforms all baselines across accuracy, precision, recall, and F1-score. This improvement can be attributed to three key design choices: (i) hybrid dense–sparse retrieval optimized for legal text, (ii) domain-specific adaptation of the language model for statutory reasoning, and (iii) temporal filtering that ensures retrieval of legally effective provisions. These findings indicate that improvements in statutory question answering are not driven solely by larger language models but by careful integration of retrieval strategies and domain knowledge. The results also demonstrate that fair comparison under a unified benchmark is essential for drawing reliable conclusions about system performance in legal AI. Expert Validation In addition to automatic evaluation, qualitative feedback was obtained from domain experts with experience in Indian tax practice. Experts reviewed a subset of system responses to assess legal correctness, citation relevance, and practical usability. The feedback confirmed that the generated responses were consistent with statutory interpretation practices and supported the reliability of the proposed approach. Conclusion This paper introduced TaxFlow, a retrieval-augmented legal assistant tailored for statutory interpretation in Indian tax law. By integrating hybrid dense–sparse retrieval, legal-domain adaptation, and temporal validity filtering, the proposed system addresses key challenges, including citation accuracy, factual grounding, and legislative evolution. Evaluation on a unified benchmark demonstrates that TaxFlow consistently outperforms strong baseline systems in both legal accuracy and linguistic quality. The results highlight the importance of jurisdiction-specific design and controlled evaluation for deploying large language models in high-stakes legal domains. Future work will focus on expanding judicial coverage, supporting multilingual Indian languages, and incorporating explainable AI techniques to enhance transparency and practitioner trust further. Future Work Several directions can further extend the scope of this research. Expanding the corpus to include judgments from the Supreme Court and High Courts would strengthen performance in precedent-driven cases. Developing multilingual support for Indian regional languages can improve inclusivity and reduce jurisdictional bias. Incorporating explainable AI (XAI) modules would help make the reasoning process transparent, thereby improving practitioner trust. Finally, adopting low-latency and resource-efficient techniques such as quantization and model distillation will be essential for real-time professional deployment. Together, these avenues provide a roadmap for advancing retrieval-augmented generation as a practical tool in high-stakes legal environments. Practical Implications The outcomes of this study are relevant not only to academia but also to practitioners and policymakers. For tax professionals and accountants, TaxFlow can serve as a rapid assistant for interpreting statutory provisions, reducing the burden of manually scanning lengthy gazettes and notifications. For lawyers and law firms, the framework provides a reliable mechanism for validating compliance and improving the accuracy of legal drafting when statutory and precedent interact. For regulators and policymakers, TaxFlow provides a scalable platform for assessing the clarity, consistency, and potential ambiguity of tax law, thereby supporting more transparent legislative processes. Researchers in AI and law can also adopt the proposed architecture as a template for building domain-specific retrieval-augmented systems in other high-impact areas such as financial compliance and healthcare regulations. By combining corpus curation, hybrid retrieval, and adaptive generation, TaxFlow delivers a modular pipeline that can be adapted to new jurisdictions and emerging legal frameworks. Overall, this study lays the groundwork for bridging the gap between static legal repositories and the need for dynamic, context-aware, practitioner-focused decision-support systems. Disclosure Statement No potential conflict of interest was reported by the author(s). Data Availability Statement The data supporting the findings of this study are derived from publicly accessible legal materials, including Indian statutory texts, government gazettes, notifications, and case-law repositories, as given below in a and b. Certain source materials were consulted from standard academic and professional reference texts for interpretation and contextual understanding. Due to copyright, licensing, and redistribution restrictions, the original legal texts and textbook materials cannot be redistributed in full. To support transparency and reproducibility, all derived data generated as part of this study – including processed text segments, question – answer pairs, ground-truth annotations, evaluation splits, configuration files, and metric outputs – have been made publicly available. These materials are sufficient to replicate all reported experiments and results. No human participants, personal data, or sensitive information were involved in this study. The Income-tax Act, 1961. Government of India.https://www.incometaxindia.gov.in. ICAI. Direct Tax Laws and International Taxation. Institute of Chartered Accountants of India. https://www.icai.org/. References - Aggarwal, P., S. Banerjee, and R. Singh. 2025. “Evaluation of Retrieval-Augmented Generation in Legal QA: Human vs. Automated Judgments.” arXiv preprint arXiv:2501.04567. - Ajay Mukund, S., and K. S. Easwarakumar. 2025. “Optimizing Legal Text Summarization Through Dynamic Retrieval-Augmented Generation and Domain-Specific Adaptation.” Symmetry 17 (5): 1–13. https://doi.org/10.3390/sym17050633. - Ariai, F., J. Mackenzie, and G. Demartini. 2025. “Natural Language Processing for the Legal Domain: A Survey of Tasks, Datasets, Models and Challenges.” ACM Computing Surveys 57 (5): 1–42. - Borgeaud, S., A. Mensch, J. Hoffmann, T. Cai, E. Rutherford, J. Millican, G. Van Den Driessche, et al. 2022. “Improving Language Models by Retrieving from Trillions of Tokens.” Proceedings of the 39th International Conference on Machine Learning (ICML), Proceedings of Machine Learning Research 34:22091–22104. https://doi.org/10.48550/arXiv.2112.04426. - Boytsov, L., T. Lin, and J. Ma. 2022. “Hybrid Sparse–Dense Retrieval for Large-Scale Question Answering.” arXiv preprint arXiv:2205.08918. - Chalkidis, I., M. Fergadiotis, P. Malakasiotis, N. Aletras, and D. Kloetzer. 2021. “LexGLUE: A Benchmark Dataset for Legal Language Understanding in English.” arXiv preprint arXiv:2110.00976. - Dettmers, T., M. Lewis, L. Zettlemoyer, and M. Riedel. 2022. “LLM.Int8(): 8-Bit Matrix Multiplication for Transformers at Scale.” arXiv preprint arXiv:2208.07339. - Es, Y., R. Sharma, and M. Chen. 2023. “RAGAS: Framework for Retrieval-Augmented Generation Evaluation.” arXiv preprint arXiv:2312.02912. - Formal, B., M. Hentschel, S. Goyal, and E. Smirnova. 2021. “Weaviate: An Open-Source Vector Search Engine.” Weaviate.io White Paper. - Frantar, J., A. Ashkboos, and D. Alistarh. 2023. “QLoRA: Efficient Finetuning of Quantized Large Language Models.” arXiv preprint arXiv:2305.14314. - Henderson, P., J. Krass, K. Zheng, D. A. Sorensen, and D. Ho. 2022. “Pile of Law: Learning Responsible Data Filtering from the Law and Legal Corpora.” arXiv preprint arXiv:2207.00220. - Holzenberger, N., D. G. McDuff, D. Card, and A. Nenkova. 2022. “LegalBench: A Collaborative Benchmark for Legal Reasoning.” arXiv preprint arXiv:2212.01361. - Hou, Z., Z. Ye, N. Zeng, T. Hao, and K. Zeng. 2025. “Large Language Models Meet Legal Artificial Intelligence: A Survey.” arXiv preprint arXiv:2509.09969. https://doi.org/10.48550/arXiv.2509.09969. - Izacard, G., E. Grave, L. Hosseini, F. Petroni, L. Scholten, M. Riedel, and F. Guzmán. 2022. “ATLAS: Few-Shot Learning with Retrieval Augmented Language Models.” Proc—International Conference on Machine Learning (ICML), 9525–9555. - Kshitij, A., R. Gupta, and M. Singh. 2024. “Caselaw-RAG: Retrieval-Augmented Generation for Citation Fidelity in Legal QA.” Proceedings of the 38th AAAI Conference on Artificial Intelligence, Vancouver, British Columbia, Canada, 14567–14575. - Lewis, P., E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, et al. 2021. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” Advances in Neural Information Processing Systems (NeurIPS) 33:9459–9474. https://doi.org/10.48550/arXiv.2005.11401. - Lin, J., and X. Ma. 2021. “SPLADE V2: Sparse Lexical and Expansion Retrieval.” Information Retrieval Journal 24 (6): 573–597. - Longpre, S., L. Hou, D. Zellers, J. Wu, and R. Fergus. 2023. “The FLAN Collection: Designing Data and Methods for Effective Instruction Tuning.” arXiv preprint arXiv:2301.13688. - Meta, A. I. 2024. “LLaMA 3 Technical Report.” Meta Research. - Muennighoff, N., T. Reimers, C. He, R. Gurevych, and I. Beltagy. 2023. “MTEB: Massive Text Embedding Benchmark.” arXiv preprint arXiv:2306.04617. - OpenAI. 2023. “GPT-4 Technical Report.” arXiv preprint arXiv:2303.08774. - Quevedo, E., T. Cerny, A. Rodriguez, P. Rivas, J. Yero, K. Sooksatra, A. Zhakubayev, and D. Taibi. 2022. “Legal Natural Language Processing from 2015–2022: A Comprehensive Systematic Mapping Study of Advances and Applications.” IEEE Access 12:145286–145317. https://doi.org/10.1109/ACCESS.2023.3333946. - Rackauckas, Z. 2024. “Rag-Fusion: A New Take on Retrieval Augmented Generation.” International Journal on Natural Language Computing 13:37–47. https://doi.org/10.5121/ijnlc.2024.13103. - Ramakrishnan, A., R. Sen, and P. Rao. 2022. “Retrieval-Augmented Generation for Compliance QA.” In Proceedings of the Workshop on LegalAI, 88–95. - Rashkin, H., E. Holtzman, A. Celikyilmaz, and Y. Choi. 2023. “Measuring Attribution in Language Models.” arXiv preprint arXiv:2305.14324. - Research, B. A. A. I. 2023. “BGE Embeddings.” GitHub Repository. https://github.com/FlagOpen/FlagEmbedding(open in a new window). - Research, Z. 2022. “Milvus 2.0: Open-Source Vector Database for Billion-Scale Retrieval.” - Richmond, K. M. G. 2024. “Explainable AI and Law: An Evidential Survey.” Artificial Intelligence and Law 32 (1): 45–72. - Sharma, R., P. Iyer, and V. Srinivasan. 2024. “Domain-Specific Retrieval-Augmented Generation for Indian Statutory Question Answering.” In Proceedings of the International Conference on Computational Linguistics (COLING 2024), 2215–2227. Turin, Italy. - Shi, W., S. Min, M. Lewis, and L. Zettlemoyer. 2023. “Self-RAG: Learning to Retrieve, Generate, and Critique for Improved Language Modeling.” arXiv preprint arXiv:2310.11511. - Shuster, K., M. Roller, S. Riedel, and J. Weston. “RARR: Assessing and Improving Faithfulness of Generation with Retrieval-Augmented Models.” arXiv preprint arXiv:2101.11794, 2021. - Tang, Y., X. Chen, L. Yang, and J. Ma. 2022. “ANCE-PRF: Document Expansion and Query Reformulation for Dense Retrieval.” In Proc. Empirical Methods in Natural Language Processing (EMNLP), 1245–1257. Abu Dhabi, United Arab Emirates. - Touvron, H., L. Martin, K. Stone, P. Albert, and T. Lavril. 2023. “LLaMA 2: Open Foundation and Fine-Tuned Chat Models.” arXiv preprint arXiv:2307.09288. - Trivedi, H., S. Rajpurohit, J. R. Balachandran, A. Kalyan, and D. Roth. 2023. “Interleaving Retrieval with Chain-of-Thought Reasoning for Robust Multi-Hop Question Answering.” arXiv preprint arXiv:2305.11496. - Xu, Z., J. Liu, H. Guo, and W. Li. 2023. “LAW-MT5: Multilingual Legal Language Understanding via Transfer Learning.” In Findings of the Association for Computational Linguistics: ACL 2023, 505–520. - Zhang, Y., and X. Xiao. 2023. “Contract-RAG: Retrieval-Augmented Generation for Contract Interpretation.” In Proc. Empirical Methods in Natural Language Processing (EMNLP 2023), 10234–10245. Singapore. - Zhong, H., J. Guo, Z. Xiao, and J. Tang. 2023. “Predicting Legal Judgments With Retrieval-Augmented Transformers.” In Proc. International Joint Conference on Artificial Intelligence (IJCAI 2023), 4250–4257. Macao, SAR, China.