The original title is: "How retrieval diversity and citation discipline beat fine-tuning in a RAG analyst system"

Short summary

A team built an AI analyst-memo system for Nigerian economic data, targeting a strict acceptance bar: six queries scoring mean ≥8/10 across three runs. They improved retrieval diversity, added rule-based intent classification with memo templates, enforced primary-source citation preferences, and ran a 24-item hallucination audit achieving 24/24 trap passes with zero hallucinations. The fine-tune wasn't the key driver — retrieval breadth, citation discipline, and structural router fixes did the heavy lifting.

• •Five-phase improvement of a RAG-based analyst memo system for Nigerian economic data
• •Hallucination audit: 24/24 traps pass, 0 halls, 20/20 production queries cite-clean
• •Retrieval diversity and citation preference did more than fine-tuning to hit the quality bar

Generated with AI, which can make mistakes.