Bias-Aware Information Retrieval Using Multi-Agent LLM Evaluation and Fairness-Constrained Reranking

Abstract

Bias-aware information retrieval has become a central concern for search engines, recommendation platforms, enterprise knowledge systems, and public-sector digital services. Conventional retrieval pipelines optimize relevance while treating fairness as a secondary constraint, which allows historical inequalities, representation gaps, and feedback loops to persist across rankings. This paper examines a system-level architecture that combines multi-agent large language model evaluation with fairness-constrained reranking. The proposed perspective treats retrieval as a socio-technical infrastructure rather than a purely algorithmic task. It analyzes how multiple evaluator agents can surface divergent judgments about relevance, harm, representational balance, and contextual sensitivity, and how those judgments can inform a reranking layer that enforces explicit fairness constraints. The discussion emphasizes architectural trade-offs among relevance, diversity, transparency, latency, cost, and maintainability. It also considers governance requirements such as auditability, model documentation, escalation pathways, and policy alignment. Particular attention is given to deployment constraints in resource-limited settings, the risk of fairness gerrymandering, the instability of opaque evaluator ensembles, and the need for continuous monitoring after release. Through cross-domain examples from web search, academic discovery, hiring systems, and public information access, the paper argues that fairness-aware retrieval is most credible when it is embedded in institutional processes rather than appended as a final technical filter. The conclusion outlines a research agenda for robust, contestable, and sustainable bias-aware retrieval systems.