RAG-Based Legal Research Assistant for Finding Similar Past Cases

This paper presents the design, architecture, and evaluation of a Retrieval-Augmented Generation system that assists new legal assistants in locating connected and similar past cases for new filings. The solution addresses Job 1, Legal Assistant, and leverages a curated Knowledge Base of 10 structured research session logs spanning five practice areas at a fictional law firm. Generative AI is assessed as capable of handling approximately 80 percent of the task, with the remaining 20 percent requiring human legal judgment, case validity verification, and jurisdiction-specific reasoning. The RAG architecture pairs a HuggingFace sentence transformer embedding model with FAISS vector search and GPT-4o-mini for grounded generation. Three query enhancement techniques improve retrieval precision beyond the baseline. Evaluation across eight metrics covering retrieval quality, generation quality through the RAG Triad, and operational performance demonstrates that the solution meets or exceeds the 0.80 target threshold on seven of eight dimensions. The paper documents limitations in cost, latency, case law currency, and the irreplaceable need for attorney oversight.

@article{ijter2026212604234510,
  author = {Shivanand R Koppalkar},
  title = {RAG-Based Legal Research Assistant for Finding Similar Past Cases},
  journal = {International Journal of Technology &  Emerging Research },
  year = {2026},
  volume = {2},
  number = {4},
  pages = {125-134},
  doi =  {10.64823/ijter.2604015},
  issn = {3068-109X},
  url = {https://www.ijter.org/article/212604234510/rag-based-legal-research-assistant-for-finding-similar-past-cases},
  abstract = {This paper presents the design, architecture, and evaluation of a Retrieval-Augmented Generation system that assists new legal assistants in locating connected and similar past cases for new filings. The solution addresses Job 1, Legal Assistant, and leverages a curated Knowledge Base of 10 structured research session logs spanning five practice areas at a fictional law firm. Generative AI is assessed as capable of handling approximately 80 percent of the task, with the remaining 20 percent requiring human legal judgment, case validity verification, and jurisdiction-specific reasoning. The RAG architecture pairs a HuggingFace sentence transformer embedding model with FAISS vector search and GPT-4o-mini for grounded generation. Three query enhancement techniques improve retrieval precision beyond the baseline. Evaluation across eight metrics covering retrieval quality, generation quality through the RAG Triad, and operational performance demonstrates that the solution meets or exceeds the 0.80 target threshold on seven of eight dimensions. The paper documents limitations in cost, latency, case law currency, and the irreplaceable need for attorney oversight.},
  keywords = {Retrieval-Augmented Generation, RAG, legal research, pre-trained models, FAISS, groundedness, faithfulness, knowledge base, GenAI evaluation, legal assistant},
  month = {Apr},
}

Copyright © 2025 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/) which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.