A Retrieval-Augmented Generation Chatbot for Accessible, Hallucination-Resistant Access to the Constitution of Nepal

Abstract

The constitution and laws of Nepal remain limited
to static PDF documents, leaving users to interpret dense legal
language largely unaided. This paper presents ArticleWise,
a retrieval-augmented generation (RAG) system for question
answering over the Constitution of Nepal (2015/2072) that is
designed to be accessible and hallucination-resistant. The system
combines a deterministic citation router that provides O(1)
lookup of statutes with a dense semantic search pipeline: the
full set of 308 numbered constitutional provisions was extracted
through layout-aware parsing, represented in a 1024-dimensional
semantic space with multilingual-e5-large and asymmetric task
prefixes, and indexed in PostgreSQL with pgvector using an exact
flat nearest-neighbor scan. The top-ranked article is supplied
to Google Gemini 3.5 Flash-Lite for a succinct, plain-language
explanation generated under closed-domain prompting, while
the statutory text itself is always returned verbatim from an
immutable database record and never passed through the generative
model. On a stratified benchmark of N = 50 questions
spanning paraphrases, citizen phrasing, underspecified keywords,
and multi-article workflow queries, the system achieves Top-1
retrieval accuracy of 62.0% (95% CI 48.2–74.1%) and Top-3
accuracy of 84.0% (95% CI 71.5–91.7%), with median retrievalonly
latency of 72.2 ms and average end-to-end latency of
372.5 ms. A faithfulness evaluation shows a 0.0% hallucination
rate relative to whatever article is retrieved, with unanimous
inter-annotator agreement — a result that speaks to the fidelity of
the explanation step but not to the correctness of article selection,
a distinction we treat as a first-class limitation of the system
rather than a footnote to the headline result