@misc{misc,
  author = {Bruce Mainza and Dalitso Zulu and Maliseni Chavula and Lighton Phiri},
  title = {NAZ Order Papers System: A Questions and Motions Management System with Hybrid AI-Assisted Duplicate and Similarity Detection for the National Assembly of Zambia},
  abstract = {The National Assembly of Zambia (NAZ) handles parliamentary questions and notices of motion through largely manual procedures. Clerks check each new submission against earlier records by hand to decide whether the matter is a duplicate or has already been addressed, and then place approved items on the Order Paper. The process is slow and difficult to audit, and it misses repeated matters that are worded differently. This project designed, built and evaluated the NAZ Order Papers System, a web application that digitises the full lifecycle of a question or motion, from submission and clerk review to scheduling, Order Paper generation and archiving. It adds locally hosted artificial intelligence that helps clerks detect duplicate and previously addressed matters. Following a design science approach and an iterative, test-driven development process, the team implemented a Next.js interface, a FastAPI service and a PostgreSQL 16 database with the pgvector extension. Access is controlled by fifteen permissions grouped into configurable roles, and a nine-state lifecycle governs every record. Retrieval combines weighted PostgreSQL full-text search with 768-dimensional EmbeddingGemma vectors through Reciprocal Rank Fusion, run entirely on locally hosted infrastructure through Ollama, and a clerk always makes the final decision. This hybrid lexical-semantic retrieval, used for record search and for two-stage exact and near-duplicate detection, is the project{\textquoteright}s delivered artificial intelligence contribution; a further generative, retrieval-augmented explanation layer is scaffolded in the codebase but is deliberately treated as future work and is not evaluated in this report. All 224 backend tests and 21 frontend source tests pass. Lexical search over a 100,000-record corpus returned 20 ranked results with a 95th-percentile latency of 266.5 ms, against a 3,000 ms budget. In a controlled evaluation with reworded queries, hybrid retrieval ranked the correct earlier record higher than keyword search alone (Chapter 7). The results show that a locally deployed, human-supervised hybrid retrieval system is a feasible way to modernise Order Paper preparation at NAZ. Formal user acceptance testing, calibration of the similarity thresholds on real NAZ records, and the future generative explanation phase remain for subsequent work.},
  year = {2026},
  pages = {112},
  publisher = {University of Zambia},
  address = {Lusaka, Zambia},
}