RGAR Framework: Factual & Conceptual Retrieval

• RGAR is a dual-source methodology that iteratively refines factual extraction and conceptual retrieval to enhance answer precision.
• It employs multiple zero-shot query expansions and recurrence rounds to separate patient-specific evidence from general domain knowledge.
• Empirical evaluations indicate RGAR improves accuracy by up to 26 points, showcasing robustness in long-context, high-stakes scenarios.

The RGAR framework refers primarily to Recurrence Generation-Augmented Retrieval, an advanced methodology in Retrieval-Augmented Generation (RAG) that explicitly models the interplay and iterative refinement between factual and conceptual knowledge sources. The approach was first instantiated for high-stakes medical question answering, where the separation and recursion between patient-specific factual evidence (e.g., extracted from Electronic Health Records, EHRs) and conceptual/encyclopedic knowledge (e.g., biomedical corpora) are essential for system reliability and answer precision (Liang et al., 19 Feb 2025). RGAR has also influenced other paradigms where regulated, governance-focused, or graph-based retrieval-augmented generation architectures require more granular orchestration of retrieval and evidence assembly (Ray, 22 Oct 2025, Cao et al., 2024).

1. Motivations and Foundational Principles

RGAR was developed to address the limitations of standard RAG methods, which process all external text uniformly. Such uniform treatment is suboptimal when the information sources differ significantly in character and relevance, as in medical QA: EHRs provide case-specific, factual observations, whereas large medical corpora encode general, conceptual domain knowledge. Traditional retrievers fail in scenarios where noisy, lengthy factual records dilute useful context, and generation models may hallucinate or be distracted by irrelevant details.

RGAR targets the following principles:

• Separate extraction of relevant factual spans from long, noisy primary sources (e.g., EHRs), minimizing irrelevant context.
• Iterative, bidirectional refinement between factual extraction and conceptual retrieval: factual evidence sharpens conceptual retrieval, and conceptual context guides further factual selection.
• Zero-shot operation: RGAR achieves its performance improvements through orchestrated prompting and retrieval, not through additional model fine-tuning or task-specific supervision.
• Explicit consideration of the dual-source nature of evidence, enabling maximal exploitation of each modality's strengths.

This approach has shown that deep interaction between evidence types increases robustness and enables smaller models to outperform much larger baselines in specialized tasks.

2. System Architecture and Workflow

The RGAR pipeline is defined by an explicit recurrence loop bridging factual extraction and conceptual retrieval. The major components and their algorithmic flow are as follows (Liang et al., 19 Feb 2025):

1. Conceptual Knowledge Retrieval (CKR): Assemble a basic query by concatenating the user's question $\mathcal{Q}$ with the full EHR $\mathcal{F}$: $q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$. Use a pretrained dense retriever $E(\cdot)$ to retrieve the top-$N$ chunks $\mathcal{C}$ from a medical corpus $\mathcal{K}$. Enhance retrieval through multiple zero-shot query expansions (answer-focused, context-focused, title-focused) generated via LLM prompting. Final retrieval score is a softmax-normalized sum across expansion similarities.
2. Factual Knowledge Extraction (FKE): Prompt an LLM to extract only spans $\mathcal{F}_s \subset \mathcal{F}$ directly relevant to $\mathcal{Q}$ given the retrieved conceptual evidence $\mathcal{C}$. Optionally, the LLM further summarizes or enriches $\mathcal{F}$0 into $\mathcal{F}$1 for interpretability or aggregation.
3. Recurrence Pipeline: Update the next query with the enriched factual evidence, forming $\mathcal{F}$2. Rerun CKR and FKE for a fixed number of rounds (typically $\mathcal{F}$3). Each round refines the retrieval and extracted facts in light of new evidence.
4. Final Answer Generation: After $\mathcal{F}$4 rounds, pass the final conceptual set $\mathcal{F}$5, the final filtered/enriched factual evidence $\mathcal{F}$6, the question $\mathcal{F}$7, and answer options $\mathcal{F}$8 to the LLM for answer synthesis. All LLM steps use zero-shot prompting and deterministic decoding; no additional training is performed.

A generic block diagram of the RGAR process is as follows:

$E(\cdot)$0

3. Mathematical Formulations

RGAR formalizes its retrieval and extraction procedure as follows:

• Retrieval Scoring: For each expansion $\mathcal{F}$9 (answer/context/title), compute similarity with candidate chunk $q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$0: $q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$1. The final normalized score:

$q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$2

Top-$q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$3 chunks are selected accordingly.

• Factual Span Extraction: $q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$4, where $q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$5 is the span-extraction prompt. Summarization/enrichment: $q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$6.
• Recurrence Update: Each iteration uses $q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$7 for subsequent retrieval.
• Dual-Source Objective (Intuition): The framework aims to maximize:

$q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$8

with score and quality defined implicitly by pretrained retriever and extraction prompts.

No additional model fine-tuning or task-specific loss is introduced.

4. Empirical Evaluation and Performance

RGAR has been extensively evaluated on three factual-aware medical QA benchmarks (Liang et al., 19 Feb 2025):

• MedQA-USMLE: Physician-exam style questions, moderate EHR length. RGAR achieves +8.63 points accuracy improvement over a custom zero-shot baseline.
• MedMCQA: Multi-subject medical MCQA, short EHRs. RGAR achieves +1.01 point gain.
• EHRNoteQA: Authentic discharge-note QA, long EHRs. RGAR improves by +26.09 points over baseline, largest among compared methods.

Across all benchmarks, mean improvement is +11.91 points, outperforming previous advanced methods such as i-MedRAG and GAR. Qualitative analysis shows RGAR's gains increase with longer EHRs, confirming the value of targeted factual extraction in complex settings. Model scaling results demonstrate that for LLMs above 3 billion parameters, relative performance improvements via RGAR become more pronounced.

Ablation studies indicate that eliminating either multi-query expansion in CKR or guided factual extraction degrades performance, and that at least two recurrence rounds are required for maximal benefit.

Method	MedQA-USMLE Acc	MedMCQA Acc	EHRNoteQA Acc	Average Gain
Custom	50.20	50.01	47.19	49.13
RAG	53.50	50.54	61.12	55.05
GAR	57.97	50.42	65.48	57.96
i-MedRAG	56.24	44.94	74.22	58.47
RGAR	58.83	51.02	73.28	61.04

In tests with Llama-3.1-8B-Instruct, RGAR surpasses RAG-enhanced GPT-3.5-turbo in accuracy.

5. Comparison to Related Paradigms

RGAR extends and generalizes earlier hybrid schemes:

• GAR (Generation-Augmented Retrieval): Generates additional context for the query and then retrieves. RGAR refines this by focusing generated context on factual span selection, and by iteratively feeding improved evidence to the retriever.
• RAG (Retrieval-Augmented Generation): Retrieves documents as context and then generates answers. RGAR injects recurrence and bidirectionally leverages extraction and retrieval.
• i-MedRAG and MedRAG: These methods combine elements of query expansion and retrieval but lack explicit recurrent interplay found in RGAR.
• Policy-Governed RGAR (Regulated-Governance Augmented Retrieval): In regulatory workflows, RGAR-inspired architectures enforce contracts, manifest cryptographic trails, and generate signed receipts for audit—characteristics aligned with regulated domains (Ray, 22 Oct 2025).
• RGAR for graph-based workflows: In knowledge graph settings, RGAR decomposes into modules for subgraph extraction, path filtering and refinement, supporting modular design and best-practice selection (Cao et al., 2024).

6. Extensions, Design Insights, and Best Practices

RGAR demonstrates that fine-grained orchestration between distinct knowledge streams enhances factual correctness, especially under resource constraints or long-context demands. Best-practice insights include:

• Employ explicit span extraction over long factual sources where irrelevant detail may confound retrieval.
• Use multiple LLM-generated query expansions to promote retrieval diversity and coverage.
• Restrict recurrence rounds (typically $q_b^{(0)} = \mathcal{Q} \oplus \mathcal{F}$9) to avoid over-inference while capturing most iterative benefits.
• Exploit RGAR's train-free nature for settings where supervised data is scarce or zero-shot domain adaptation is required.

In regulated or graph-based domains, RGAR-inspired frameworks further introduce mechanisms such as policy gating, cryptographic evidence anchoring, and modular graph reasoning stages, generalizing the core recurrence and dual-source principles.

---

In summary, RGAR establishes a principled, modular approach for high-fidelity retrieval-augmented generation, centered on recurrent, interacting extraction and retrieval modules. This dual-source, recurrence-based orchestration achieves state-of-the-art results in factual-aware domains and provides a foundational architecture for future extensions in regulated, scientific, and highly structured reasoning workflows (Liang et al., 19 Feb 2025, Ray, 22 Oct 2025, Cao et al., 2024).