Orchestrating Specialized Agents for Trustworthy Enterprise RAG

• Enterprise reporting requires completeness, citation stability, and section coverage, while complex queries often need iterative decomposition and cross-document synthesis.Source: §1 · p. 1
• ADORE differs from outline-only convergence by constraining generation to section-scoped evidence and using coverage audits to drive retrieval and stopping.Source: §1 · p. 2
• Memory-locked synthesis stores explicit claim–evidence links and section-level admissible evidence to ground report claims and citations.Source: §1 · p. 2
• Evidence-coverage audits identify weak or missing sections, trigger targeted retrieval, and provide an evidence-driven stopping criterion.Source: §1 · p. 2
• Section-packed grounding packs relevant evidence, prunes redundancy, and compresses sources into citation-preserving summaries for context-limited synthesis.Source: §1 · p. 2
• ADORE achieves state-of-the-art results on both public and internal benchmarks.Source: §1 · p. 2

• The Orchestrator selects low-latency retrieval for simple informational queries and multi-agent deep research for complex analytical tasks.Source: §2 · p. 2
• ADORE replaces linear retrieval with an Orchestrator-governed agentic workflow for complex tasks.Source: Figure 1 · p. 2, §2 · p. 2
• The structured Memory Bank links claims to evidence and constrains generation to support traceable enterprise reporting.Source: Figure 1 · p. 2

• The Orchestrator classifies requests by complexity, including ambiguity, expected depth, sub-question count, and traceability needs.Source: §2.1 · p. 2
• Complex requests move through clarification, user-editable planning, and iterative retrieval, synthesis, and verification.Source: §2.1 · p. 3
• Coverage and stability signals determine whether ADORE retrieves more, refines the outline, or finalizes, rather than following a fixed iteration count.Source: §2.1 · p. 3

• The Grounding Agent clarifies underspecified prompts, while the Planning Agent turns the refined brief into a user-editable plan.Source: §2.2 · p. 3
• The Execution Agent adapts search queries to intermediate findings and evidence gaps to support multi-hop recovery.Source: §2.2 · p. 3
• The Report Generation Agent creates and revises outlines before producing citation-traceable reports.Source: §2.2 · p. 3
• The WebSearch Agent filters high-authority public sources when external context is needed.Source: §2.2 · p. 3

2.3 Execution Workflow

• The Report Generation Agent first proposes an outline aligned with the user-confirmed plan.The outline establishes the sections for the subsequent research loop.Source: §2.3 · p. 3
• The Self-Evolution Engine iteratively searches internal and external sources, while reflection identifies weakly supported or missing sections and refines queries.The Report Generation Agent then revises the outline using newly retrieved evidence.Source: §2.3 · p. 3
• The Execution Agent checks evidence coverage and citation stability against the plan, returning to retrieval if they are insufficient and otherwise triggering final report generation.Source: §2.3 · p. 3

2.4 Synthesis and Traceability

• ADORE’s departure from prior approaches is to make the Memory Bank a hard generation constraint and use section-level evidence coverage to drive retrieval and convergence.Source: §2.4 · p. 3

2.4.1 Memory Bank (Claim–Evidence Graph).

• The Memory Bank stores claim–evidence links, section-specific admissible sources and excerpts, and coverage requirements, so claims can be traced to supporting documents.With a fixed Memory Bank, each section is generated only from its admissible evidence, supporting citation audits against stored sources and quoted spans.Source: §2.4.1 · p. 3, §2.4.2 · p. 3

2.4.2 Memory-locked synthesis.

• After each retrieval–reflection cycle, section-level coverage audits trigger targeted retrieval and focused rewriting where evidence is insufficient; the loop stops when plan coverage requirements are satisfied.Source: §2.4.3 · p. 3

2.4.3 Evidence-coverage–guided execution Criterion.

• Section-packed grounding uses section-relevant evidence, prunes redundant or low-salience text, and compresses long sources into citation-preserving summaries to fit context budgets.These steps are described as keeping grounded long-form synthesis practical at enterprise scale.Source: §2.4.4 · p. 3

3 Evaluation Methodology

• The evaluation uses both reference-based metrics and reference-free preference judgments.Source: §3 · p. 3

3.1 Reference-Based: The RACE Framework

• RACE benchmarks Readability, Instruction Following, Comprehensiveness, and Insight using task-specific weights and generated sub-criteria.For the SPM market-analysis task, Table 1 presents the comprehensiveness sub-criteria and their weights.Source: §3.1 · p. 3, Table 1 · p. 4

3.2 Reference-free: Side-by-Side Evaluation

• The side-by-side evaluation compares ADORE reports with competitor reports using an LLM judge’s win, tie, or lose judgments.The setup estimates how often judges or users would prefer ADORE when viewing reports side by side.Source: §3.2 · p. 4

3.3 Evaluation Datasets

• The evaluation datasets are DeepResearch Bench, an expert-authored public set of 100 PhD-level tasks; a proprietary internal enterprise benchmark; and the public Deep Consult business and consulting dataset.The internal benchmark tasks are derived from high-quality internal documents using reverse prompt engineering.Source: §3.3 · p. 4

4.2 Side-by-Side Preference

• 52.65 on DeepResearch Bench ranked ADORE #1; it scored 64.11 on the Internal Enterprise Benchmark and won 77.21% against ChatGPT Deep Research on DeepConsult.The DeepConsult lose rate was 4.41%.Source: §4.1 · p. 4, §4.2 · p. 4

4.1 RACE Performance

• The reported component scores were 52.22 for Comprehensiveness, 54.37 for Insight, 51.11 for Instruction Following, and 52.18 for Readability.Source: §4.1 · p. 4
• The reported component scores were 65.91 for Comprehensiveness, 66.76 for Insight, 60.35 for Instruction Following, and 61.95 for Readability.Source: §4.1 · p. 4
• 14.11% improvement over human baselines was reported for the specific enterprise tasks, with 50.0 calibrated as the human-written reference score.Source: §4.1 · p. 4

5 Conclusion

• ADORE reports state-of-the-art results across public and internal benchmarks while maintaining traceable, section-level evidence grounding.Source: §5 · p. 4
• Its operational novelty comprises Memory-locked synthesis, evidence-coverage-guided retrieval with a stopping rule, and section-packed long-context grounding.Source: §5 · p. 4
• Future work would extend planning to propose decision options and trade-offs, beyond synthesis.Source: §5 · p. 4
• The authors also propose granular citation-accuracy metrics and automatic audits for technical domains.Source: §5 · p. 4
• Multimodal future work would generate charts and plots aligned with evidence-backed report sections.Source: §5 · p. 4