DualRAG Framework Overview

• DualRAG is an iterative, dual-process architecture that integrates explicit reasoning with dynamic retrieval to address multi-hop question answering challenges.
• The framework combines Reasoning-Augmented Querying (RaQ) with Progressive Knowledge Aggregation (pKA) to continually update and refine a structured knowledge outline.
• Empirical results show DualRAG outperforms traditional single-branch RAG approaches, achieving higher accuracy on benchmarks like HotpotQA and specialized tasks such as medical decision support.

The DualRAG framework designates a class of retrieval-augmented generation (RAG) architectures that explicitly partition reasoning and retrieval into iterative, interacting processes. This “dual-process” paradigm aims to address longstanding challenges in multi-hop question answering (MHQA) and complex knowledge-intensive domains by tightly coupling chains of reasoning with targeted, evolving knowledge acquisition. Variants extend the principle to specialized settings such as medical decision support, where retrieval by analogy complements structured knowledge base access. DualRAG systems consistently outperform single-branch RAG approaches and approach the performance upper bound set by oracle knowledge access (Cheng et al., 25 Apr 2025, Lu et al., 26 May 2025).

1. DualRAG Architectural Principles

DualRAG is fundamentally an iterative framework consisting of two tightly coupled modules: Reasoning-augmented Querying (RaQ) and Progressive Knowledge Aggregation (pKA). RaQ maintains both a reasoning history and a dynamic knowledge outline for each input $x$, assessing at each step whether external knowledge is needed and generating entity-centric queries to address knowledge gaps. pKA, acting as an assistant, ingests retrieved documents, filters and summarizes them, and integrates their substance into the evolving knowledge outline. Operationally, the closed-loop interaction follows:

$$\begin{aligned} (r^t, D^t) &= \mathrm{RaQ}(K^{t-1}, x, R^{t-1}) \ K^t &= \mathrm{pKA}(K^{t-1}, D^t) \end{aligned}$$

where $r^t$ denotes the next reasoning step, $D^t$ the newly retrieved data, and $K^t$ the updated knowledge outline. The process repeats for $t = 1,...,T$, culminating in an answer generation step conditioned on the final knowledge sketch $K^T$ and complete reasoning trace $R^T$:

$\hat{a} = M_A(K^{T}, x, R^{T})$

This design ensures that query generation is continually informed by current knowledge and reasoning state, with retrieval acting as a dynamic, need-driven process rather than a static pre-processing stage (Cheng et al., 25 Apr 2025).

2. Reasoning-Augmented Querying & Entity-Centric Retrieval

The RaQ module is bifurcated into:

• Reasoner: Detects knowledge gaps and generates substeps in the reasoning chain using explicit representations of prior knowledge and reasoning steps. It also sets a binary flag $f^t$ indicating whether external retrieval is required.
• Entity Identifier: When $f^t = \text{True}$, extracts key entities $E^t$ that are critical for answering the current subquestion and formulates contextualized queries $Q^t(e)$ for each entity.

Retrieval employs dense representations (embeddings), where queries and passages are mapped via functions $\psi(q)$ and $\phi(p)$, and passage scores are computed via softmax-normalized dot-product similarity:

$$\mathrm{score}(p, q) = \frac{\exp(\phi(p) \cdot \psi(q))}{\sum_{p'} \exp(\phi(p') \cdot \psi(q))}$$

Top-$N$ passages are aggregated and reranked per entity, enabling fine-grained, context-sensitive knowledge collection across multiple hops (Cheng et al., 25 Apr 2025).

3. Progressive Knowledge Aggregation and Outline Construction

Upon retrieval, pKA converts the document sets for each entity into condensed knowledge fragments $k_e$ using a learned summarizer $M_{KS}$. These fragments are inserted into an entity-indexed outline:

$$K^t(e) = K^{t-1}(e) \cup \{k_e\}, \quad \forall e \in E^t$$

The overall outline thus incrementally accumulates supporting facts relevant to each entity mentioned in the reasoning process. A more neural representation treats this as iterative updates to a joint state vector incorporating both reasoning and retrieval context. This structured aggregation controls information overload, counteracts noise accumulation, and enables targeted reasoning on the expanded knowledge base (Cheng et al., 25 Apr 2025).

4. DualRAG Specialization: Medical Domain and Dual-Branch Retrieval

DoctorRAG, a specialized instantiation in the medical domain, exemplifies “dual” retrieval integration by fusing explicit clinical knowledge base access with analogical retrieval from a de-identified patient case base. The retrieval process is partitioned:

• KB branch: Retrieves ICD-10–concept-tagged declarative sentences by concept-constrained cosine similarity, enforcing semantic gating on retrievals.
• PB branch: Retrieves patient cases with unconstrained cosine similarity for analogical reasoning.

The context for generation is formed by concatenating $k_{\mathcal{K}}$ top-ranked KB statements and $k_{\mathcal{P}}$ analogous patient cases. This dual source context is then fed into a standard RAG generator, followed by Med-TextGrad multi-agent iterative refinement to align the output with both factual and analogical evidence (Lu et al., 26 May 2025). Mathematically:

$$s_{\mathcal{K}}(q, d_i) = \begin{cases} \frac{v_q^\top v_{d_i}}{\|v_q\| \|v_{d_i}\|}, & c_q \cap c_i \neq \emptyset \ -\infty, & \text{otherwise} \end{cases}$$

$$s_{\mathcal{P}}(q, p_j) = \frac{v_q^\top v_{p_j}}{\|v_q\| \|v_{p_j}\|}$$

This partitioning enables richer, more nuanced reasoning in clinical tasks, directly addressing the limitations of knowledge-base-only approaches.

5. Iterative Refinement via Multi-Agent Textual Gradients

DoctorRAG introduces Med-TextGrad, a module that refines candidate answers by simulating gradient descent in textual space. Critic agents compute two losses:

• Context Criterion $KC_\psi(A_t, C)$ for factual consistency with retrieved context
• Patient Criterion $PC_\psi(A_t, q)$ for relevance to the actual query

The aggregated textual loss is:

$$\mathcal{L}(A_t \mid q, C) = KC_\psi(A_t, C) + PC_\psi(A_t, q)$$

Two levels of gradients—answer-level and prompt-level—are then computed by specialized LLMs. The Textual Gradient Descent (TGD) agent fuses these, iteratively updating the generation prompt and driving the answer toward strict alignment with evidence and query requirements. This multi-agent, iterative optimization yields measurably more accurate and complete outputs (Lu et al., 26 May 2025).

6. Training, Fine-Tuning, and Inference Protocols

Component training in DualRAG frameworks proceeds via supervised losses attached to each submodule:

$$\mathcal{L} = \mathcal{L}_{\text{ret}} + \lambda_1 \mathcal{L}_{\text{trigger}} + \lambda_2 \mathcal{L}_{\text{EI}} + \lambda_3 \mathcal{L}_{\text{KS}}$$

• Retrieval: contrastive loss favoring ground-truth supporting passages
• Trigger: binary cross-entropy for retrieval-needed decisions
• Entity Identification: multi-label entity selection
• Summarization: sequence-to-sequence LLM loss on factoid condensation

Teacher model–generated full trajectories create datasets permitting efficient fine-tuning for compact models without sacrificing core dual-process capabilities (Cheng et al., 25 Apr 2025). DoctorRAG operates with fixed components and zero- or few-shot prompting, relying on post-hoc evaluation rather than end-to-end gradient updates (Lu et al., 26 May 2025).

7. Empirical Performance and Design Insights

DualRAG delivers strong quantitative gains on MHQA datasets (HotpotQA, 2WikiMultihopQA, MuSiQue), achieving EM and F1 scores substantially above single-process RAG and rivaling systems with oracle document access. For example, on HotpotQA:

Model	EM	F1	Oracle F1
Direct	26.0	36.4	–
NativeRAG	46.4	60.3	79.6
DualRAG	49.7	65.7	79.6

DoctorRAG attains the highest accuracy on disease diagnosis benchmarks and outperforms strong RAG and graph-based baselines in medical question answering and generation tasks across English, Chinese, and French. Ablation results demonstrate that both branches—conceptual knowledge and patient analogy—are essential, each yielding 1–3% accuracy in isolation, and concept tagging further boosts precision by enforcing semantic alignment (Lu et al., 26 May 2025).

8. Dual-Process Synergy and Implications

The key advantage of DualRAG frameworks is the closed-loop synergy between targeted knowledge gap identification and structured evidence accumulation. RaQ’s explicit reasoning trace enables focused retrieval, reducing the risk of unnecessary or irrelevant document inclusion. pKA structures supporting facts around entities, maintaining coherence across hops and controlling noise. This interaction yields more faithful, complete chains of reasoning, substantiating the system’s answers and minimizing hallucination and error propagation. A plausible implication is increased robustness in domains characterized by complex, multi-source evidence chains—where isolated retrieval or reasoning quickly becomes brittle (Cheng et al., 25 Apr 2025, Lu et al., 26 May 2025).

In summary, DualRAG architectures establish a rigorous dual-process paradigm for combining reasoned query generation and progressive evidence synthesis. This approach demonstrates empirical superiority in multi-hop and analogy-driven reasoning tasks, underlining the importance of dynamic, context-sensitive retrieval and iterative refinement in contemporary retrieval-augmented generation systems.