Logo
Adnan
Back to Research
Published Paper & Benchmarks

Retrieval-Augmented Generation (RAG) in Legal Technologies

Abstract

An academic analysis of hallucination rates and retrieval accuracy when applying LLMs to complex, domain-specific legal corpora.

Abstract

Abstract

The application of Large Language Models (LLMs) in legal technology presents significant risks regarding hallucinated citations. This paper explores the efficacy of Retrieval-Augmented Generation (RAG) architectures in mitigating these risks.

Methodology & Design

Methodology

We indexed 10,000 public court rulings using OpenAI text-embedding-3-large and implemented a hybrid search approach (Dense Vector + BM25).

Evaluation & Benchmarks

Results

Accuracy IncreaseHybrid search improved retrieval accuracy by 22% over pure vector search.
Hallucination ReductionStrict prompting combined with accurate context retrieval reduced hallucination rates to <0.5%.

Figures & Benchmarks

Click any diagram to view in high resolution.

Retrieval-Augmented Generation (RAG) in Legal Technologies diagram 1