OPG's Deep Agentic Retrieval Pipeline beats naive RAG for regulatory compliance

A new paper from Ontario Power Generation (OPG) and academic collaborators traces the evolution of a production retrieval pipeline for regulatory compliance and rate case analysis under Ontario Energy Board reporting requirements. Starting from naive RAG that fails as corpora scale, they moved through hybrid retrieval with re-ranking, then agentic function-calling retrieval, and finally a deep multi-agent architecture with code-based tool synthesis and explicit planning. The mature system, formalized as Progressive Evidence Acquisition with Cost-Aware Escalation (PEA-CAE), begins with low-cost, high-precision retrieval and only escalates to full-document reads when the expected evidence gain justifies the additional latency and cost.

The paper's key insight: context engineering—carefully orchestrating retrieval, reformulation, and summarization—is a more tractable and economically viable path than domain-specific fine-tuning for large, evolving regulatory corpora. The progression toward deep agentic retrieval mirrors classical IR concepts: adaptive query reformulation, progressive document discovery, and hierarchical subagent summarization. Operational traces confirm that iterative evidence acquisition and adaptive planning increasingly replace single-pass retrieval as the foundation for enterprise-scale question answering. Accepted at the 2026 IEEE SEGE conference, this work provides a practical blueprint for any organization managing complex regulatory document landscapes.