Electrical and Computer Engineering ETDs

Publication Date

Summer 2026

Abstract

VeriBrief is a multi-agent retrieval-augmented generation (RAG) system for evidence-grounded economic policy analysis. The system orchestrates a five-stage LangGraph pipeline, retrieval, research, analysis, synthesis, and critique, to produce cited, structured responses while detecting out-of-scope queries. An empirical evaluation on eight questions drawn from official U.S. macroeconomic releases compared VeriBrief against a single-pass RAG baseline. The multi-agent system achieved 100% refusal precision on unanswerable analytical queries versus 0% for the baseline. Unsupported claims fell substantially on answerable factual questions. A context-propagation defect discovered during evaluation was diagnosed and corrected. Limitations include failure of the evidence gate on policy-counterfactual queries and a regression in unsupported claim rate on policy-template output. Directions for future work are discussed.

Keywords

retrieval-augmented generation, multi-agent systems, LangGraph, hallucination reduction, economic decision support, local LLM, evidence contract

Document Type

Thesis

Language

English

Degree Name

Computer Engineering

Level of Degree

Masters

Department Name

Electrical and Computer Engineering

First Committee Member (Chair)

Manel Martínez-Ramón

Second Committee Member

Ramiro Jordan

Third Committee Member

Christos Christodoulou

Download

Share

COinS