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

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The constitution and laws of Nepal remain limitedto static PDF documents, leaving users to interpret dense legallanguage largely unaided. This paper presents ArticleWise,a retrieval-augmented generation (RAG) system for questionanswering over the Constitution of Nepal (2015/2072) that isdesigned to be accessible and hallucination-resistant. The systemcombines a deterministic citation router that provides O(1)lookup of statutes with a dense semantic search pipeline: thefull set of 308 numbered constitutional provisions was extractedthrough layout-aware parsing, represented in a 1024-dimensionalsemantic space with multilingual-e5-large and asymmetric taskprefixes, and indexed in PostgreSQL with pgvector using an exactflat nearest-neighbor scan. The top-ranked article is suppliedto Google Gemini 3.5 Flash-Lite for a succinct, plain-languageexplanation generated under closed-domain prompting, whilethe statutory text itself is always returned verbatim from animmutable database record and never passed through the generativemodel. On a stratified benchmark of N = 50 questionsspanning paraphrases, citizen phrasing, underspecified keywords,and multi-article workflow queries, the system achieves Top-1retrieval accuracy of 62.0% (95% CI 48.2–74.1%) and Top-3accuracy of 84.0% (95% CI 71.5–91.7%), with median retrievalonlylatency of 72.2 ms and average end-to-end latency of372.5 ms. A faithfulness evaluation shows a 0.0% hallucinationrate relative to whatever article is retrieved, with unanimousinter-annotator agreement — a result that speaks to the fidelity ofthe explanation step but not to the correctness of article selection,a distinction we treat as a first-class limitation of the systemrather than a footnote to the headline result