Orchestrating Specialized Agents for Trustworthy Enterprise RAG

TL;DR

Enterprise knowledge tasks with RAG often fail in high-stakes settings due to underspecified prompts, incomplete grounding, and lack of traceability. The paper proposes ADORE, an agentic framework that replaces linear retrieval with an adaptive orchestration of specialized agents and a Memory Bank that constrains generation to section-level admissible evidence. Its core contributions are memory-locked synthesis via a Claim–Evidence Graph, evidence-coverage guided execution with an evidence-driven stopping rule, and section-packed long-context grounding to maintain grounding under context limits. Across public and internal benchmarks, ADORE achieves state-of-the-art results and demonstrates strong preference over commercial baselines, signaling practical impact for trustworthy enterprise RAG workflows.

Abstract

Retrieval-Augmented Generation (RAG) shows promise for enterprise knowledge work, yet it often underperforms in high-stakes decision settings that require deep synthesis, strict traceability, and recovery from underspecified prompts. One-pass retrieval-and-write pipelines frequently yield shallow summaries, inconsistent grounding, and weak mechanisms for completeness verification. We introduce ADORE (Adaptive Deep Orchestration for Research in Enterprise), an agentic framework that replaces linear retrieval with iterative, user-steered investigation coordinated by a central orchestrator and a set of specialized agents. ADORE's key insight is that a structured Memory Bank (a curated evidence store with explicit claim-evidence linkage and section-level admissible evidence) enables traceable report generation and systematic checks for evidence completeness. Our contributions are threefold: (1) Memory-locked synthesis - report generation is constrained to a structured Memory Bank (Claim-Evidence Graph) with section-level admissible evidence, enabling traceable claims and grounded citations; (2) Evidence-coverage-guided execution - a retrieval-reflection loop audits section-level evidence coverage to trigger targeted follow-up retrieval and terminates via an evidence-driven stopping criterion; (3) Section-packed long-context grounding - section-level packing, pruning, and citation-preserving compression make long-form synthesis feasible under context limits. Across our evaluation suite, ADORE ranks first on DeepResearch Bench (52.65) and achieves the highest head-to-head preference win rate on DeepConsult (77.2%) against commercial systems.

Figures (2)

• Figure 1: The ADORE framework. The Orchestrator routes complex tasks to an agentic workflow that replaces linear retrieval with Memory-locked synthesis. Unlike standard RAG, ADORE constrains generation to a structured Memory Bank (Claim--Evidence Graph), enabling iterative, evidence-coverage--guided execution for traceable enterprise reporting.
• Figure 2: Self-evolving execution with Memory-locked synthesis. The workflow iterates through retrieval and reflection, storing evidence in a structured Memory Bank. The process uses section-level evidence coverage audits to localize weak or missing sections, triggering targeted follow-up retrieval and terminating when coverage criteria are met.