Interact-RAG: Interactive Retrieval Paradigm

• Interact-RAG is a retrieval-augmented generation paradigm that redefines retrieval as a manipulable process for enhanced information synthesis.
• It introduces a Corpus Interaction Engine featuring action primitives for multi-faceted retrieval, anchored matching, and context shaping.
• Empirical results show significant improvements with an average Exact Match of 52.8 on diverse QA benchmarks, emphasizing its efficiency and robustness.

Interact-RAG is a retrieval-augmented generation paradigm that treats retrieval not as a black-box query interface but as an object of direct reasoning and manipulation. Introduced as “reason and interact with the corpus, beyond black-box retrieval,” it equips an LLM agent with a Corpus Interaction Engine, a reasoning-enhanced workflow for zero-shot execution and trajectory synthesis, and a training pipeline that combines supervised fine-tuning with reinforcement learning. In reported experiments on six question-answering benchmarks, it achieves the best Exact Match (EM) scores across all datasets, reaching an average EM of 52.8 versus 43.1 for the strongest listed baseline, thereby positioning retrieval control itself as a first-class locus of agentic competence (Hui et al., 31 Oct 2025).

1. From black-box retrieval to corpus interaction

The conceptual point of departure for Interact-RAG is the progression from static RAG to iterative RAG and then to agentic RAG. Static RAG performs a single retrieval per query; iterative RAG adds multi-step retrieval chains; agentic RAG allows an LLM-driven agent to decide when and how to retrieve and synthesize information. Interact-RAG argues that these forms remain constrained by a shared limitation: retrieval is still treated as a black-box querying operation. Under that regime, the agent can only issue queries and receive returned chunks, without inspecting or controlling the mechanics of retrieval itself (Hui et al., 31 Oct 2025).

This limitation is presented as more than an implementation detail. It restricts strategic control, encourages inefficient trial-and-error query loops, and impairs performance on complex, multi-hop, or compositional information-seeking tasks. The paper’s motivating example asks which film was released first, “The Jaws of Death” or “Failure to Launch”; a conventional agent may repeatedly paraphrase queries about release dates without resolving a retriever failure rooted in phrasing or noise. Interact-RAG therefore redefines retrieval as an interactive substrate that can be shaped, constrained, and combined during reasoning rather than merely queried (Hui et al., 31 Oct 2025).

A common misconception is to equate Interact-RAG with ordinary iterative retrieval. The distinction is narrower and more technical: iterative retrieval changes the sequence of queries, whereas Interact-RAG changes the agent’s action space over the corpus. The central novelty is not only that retrieval happens multiple times, but that the agent gains fine-grained control over retrieval modes, anchoring strategies, document inclusion and exclusion, and retrieval scale (Hui et al., 31 Oct 2025).

2. Corpus Interaction Engine and action primitives

The core systems contribution is the Corpus Interaction Engine, which exposes a set of action primitives denoted as $\mathcal{A}_{\text{CI}}$. These primitives are organized into three categories: multi-faceted retrieval, anchored matching, and context shaping. Multi-faceted retrieval includes semantic_search(query_s) for dense retrieval using embedding similarity, exact_search(keywords_e) for sparse retrieval using BM25-style keyword matching, and weighted_fusion(w_s, w_e) for adaptive blending of semantic and keyword retrieval with agent-specified weights. Anchored matching is implemented through entity_match(entity), which centers retrieval on segments associated with a named entity. Context shaping includes include_docs(doc_ids), exclude_docs(doc_ids), and adjust_scale(n) to preserve, remove, or resize the working retrieval context (Hui et al., 31 Oct 2025).

Category	Primitive	Function
Multi-faceted retrieval	semantic_search, exact_search, weighted_fusion	Combine dense and sparse evidence
Anchored matching	entity_match	Focus retrieval around an entity
Context shaping	include_docs, exclude_docs, adjust_scale	Retain, filter, or resize context

These primitives are designed to be operationally lightweight. exact_search and entity_match are realized with SQLite Full-Text Search using BM25, while semantic_search uses a vector database built with ChromaDB and E5 embeddings. The fusion primitive normalizes and combines top-$k$ results under agent-set weights, and context shaping relies on efficient document-ID filters. The engine does not require heavyweight LLM invocations during primitive execution; instead, the LLM reasons over structured tool outputs and decides subsequent actions (Hui et al., 31 Oct 2025).

This architecture shifts the semantics of retrieval. Instead of returning a single opaque list of documents, the retrieval substrate becomes decomposable into searchable, anchorable, and filterable operations. A plausible implication is that retrieval errors become more diagnosable because the agent can fail at a specific primitive—such as weighting, entity anchoring, or context shaping—rather than at an undifferentiated search step.

3. Reasoning-enhanced workflow

Interact-RAG does not rely solely on end-to-end prompting of an action-capable agent. It first introduces a reasoning-enhanced modular workflow with three collaborative modules: a Global-Planner, an Adaptive-Reasoner, and an Executor. The Global-Planner decomposes the initial user query into a stepwise execution plan. The Adaptive-Reasoner examines the evolving task state, including prior actions, gathered documents, and the plan, and decides whether to proceed or to reflect and refine when retrieval has failed or stalled. The Executor converts the reasoner’s directive into concrete action primitives and, after all subtasks, produces the final answer (Hui et al., 31 Oct 2025).

The workflow serves two purposes. First, it supports stable zero-shot performance by separating planning, reflection, and execution. Second, it generates structured agent trajectories that can later be used as training data. Those trajectories contain states, reasoning thoughts, and action sequences, and are retained only when the final answer matches ground truth. This retention rule is important because it constrains synthesis toward successful interaction patterns rather than generic tool-use traces (Hui et al., 31 Oct 2025).

The modularity also sharpens the contrast with prior agentic RAG methods. Prompt-only agents often collapse planning and retrieval into a single autoregressive stream, which can obscure whether failure originates in decomposition, evidence acquisition, or answer synthesis. In Interact-RAG, the workflow makes that decomposition explicit. This suggests that the system’s contribution is as much about policy factorization as about tool design.

4. Supervised fine-tuning and reinforcement learning

The training pipeline proceeds in two stages. In the first stage, the model is supervisedly fine-tuned on 4.8K synthesized agent trajectories. The objective is imitation of high-quality thought-and-action sequences in autoregressive form. Retrieved content tokens are masked during this phase so that the model does not overfit superficial artifacts of specific retrieval outputs (Hui et al., 31 Oct 2025).

In the second stage, the SFT model is refined with reinforcement learning using Group Relative Policy Optimization (GRPO) on 7.1K QA pairs. The reward function is reported as

$$R(\tau) = -1 + \mathbb{I}\{\tau_{\text{valid}}\} + \mathbb{I}\{\tau_{\text{valid}}\}\cdot \mathbb{I}\{y_{\text{ans}}\},$$

so malformed trajectories incur a penalty, syntactically valid trajectories receive a bonus, and correct answers earn an additional bonus. The emphasis is therefore on both well-formed interaction behavior and final correctness, rather than on answer accuracy alone (Hui et al., 31 Oct 2025).

This two-stage design matters because the action space is richer than that of ordinary query-only RAG agents. SFT supplies behavioral priors for planning, reflection, and tool invocation, while RL refines those priors into a more selective policy. The reported behavioral analysis states that, as training progresses from untrained to SFT to RL, the agent interacts more frequently and more efficiently with the primitives, uses entity anchoring and document shaping more adaptively, and achieves higher yield with fewer interactions after RL (Hui et al., 31 Oct 2025).

The training procedure places Interact-RAG within a broader family of process-aware agentic RAG methods. SIRAG, for example, inserts a Decision Maker and a Knowledge Selector between retriever and generator, uses an LLM-as-a-Judge for process-level rewards, explores reasoning trees, and trains with PPO in an end-to-end manner (Wang et al., 17 Sep 2025). Interact-RAG differs in where it intervenes: its primary intervention is to expose and learn over the internal controls of retrieval itself.

5. Empirical performance and observed behavior

The empirical evaluation spans six datasets: Natural Questions and PopQA for single-hop QA, and HotpotQA, 2WikiMultiHopQA, MuSiQue, and Bamboogle for multi-hop QA. The corpus is a preprocessed 2018 Wikipedia dump of roughly 280K text chunks, and the primary backbone LLM is Qwen3-8B. Against listed baselines including Direct, Standard RAG, IR-CoT, MA-RAG, Search-R1, R-Search, and S-DeepSearcher, Interact-RAG obtains the highest EM on every dataset (Hui et al., 31 Oct 2025).

Dataset	Interact-RAG EM	Best listed non-Interact baseline
HotpotQA	51.6	45.2
2WikiMultiHopQA	69.6	58.8
MuSiQue	34.8	25.5
Bamboogle	54.0	46.2
NQ	50.9	45.3
PopQA	56.0	49.3
Average	52.8	43.1

The average EM improvement over the strongest listed baseline is 9.7 points, reported as a 22.5% relative gain. The gains are especially pronounced on multi-hop tasks, including a reported 36.4% relative EM improvement on MuSiQue. The paper also reports out-of-distribution robustness: the agent is trained on HotpotQA, MuSiQue, and NQ, yet evaluates well on PopQA, 2Wiki, and Bamboogle (Hui et al., 31 Oct 2025).

Ablation results are structurally informative. Removing the interaction primitives, removing SFT, or removing RL all causes clear performance drops. In addition, the modular “Interact-RAG-Workflow” without end-to-end training still outperforms strong prompt-based agentic baselines, indicating that the workflow itself carries substantial value even before policy learning. Efficiency analysis further reports that Interact-RAG requires fewer retrieval or action iterations to reach an answer, which is consistent with the claim that fine-grained control helps avoid unproductive query loops (Hui et al., 31 Oct 2025).

6. Relation to adjacent interactive RAG paradigms

Interact-RAG is part of a broader shift toward treating retrieval as an active reasoning process. In graph-centric settings, GRAIL defines three actions—Explore Entity, Choose Relation, and Finish—and trains a policy that balances retrieval breadth and precision over large knowledge graphs; it reports an average accuracy improvement of 21.01% and an $F_1$ improvement of 22.43% on three knowledge-graph QA datasets (Chang et al., 7 Aug 2025). MC-RAG reformulates retrieval for multi-constraint queries as subgraph matching over a knowledge graph, combining semantic and structural embeddings with path-level R*-Tree indexing and generating answers only after constraint verification (Zhang et al., 11 Jul 2026). These systems differ in corpus structure, but they share with Interact-RAG the view that retrieval is not merely document lookup.

A second, complementary line of work uses “interactive” in a user-facing rather than agent-facing sense. SMARTFinRAG provides a fully modular financial RAG benchmark with five pipeline components, centralized YAML configuration, a Streamlit-based GUI, runtime swapping of retrievers and LLMs, document-centric QA generation, and real-time visualization of retrieval and response-quality metrics (Zha, 25 Apr 2025). XGraphRAG offers interactive visual analysis for GraphRAG systems, with linked views for QA traces, topic exploration, entity exploration, and LLM invocation tracing to identify missing or unexpected recalls (Wang et al., 10 Jun 2025). iRAG brings interactivity to video RAG through incremental, query-driven extraction, reporting 23x to 25x faster video-to-text ingestion while maintaining response latency and quality comparable to traditional full-upfront conversion (Arefeen et al., 2024).

This suggests that “interactive RAG” now spans at least two axes: agent-corpus interaction, where the model actively manipulates retrieval; and human-system interaction, where practitioners inspect, configure, or debug the retrieval pipeline. Interact-RAG belongs primarily to the first axis, though its explicit primitive structure also makes it more amenable to the second.

7. Evaluation, robustness, and emerging research directions

The rise of Interact-RAG intensifies evaluation demands. DataMorgana targets the benchmark-construction problem by generating highly customizable and diverse synthetic Q&A benchmarks with control over user and question categories and their distributions, and reports superior lexical, syntactic, and semantic diversity over existing tools (Filice et al., 22 Jan 2025). CRAG-MM extends the evaluation frontier to multi-modal, multi-turn RAG, with 6.5K single-turn triplets and 2K visual-based multi-turn conversations across 13 domains; straightforward RAG approaches achieve only 32% and 43% truthfulness on its single- and multi-turn settings, respectively, while state-of-the-art industry solutions are reported at 32% and 45% (Wang et al., 30 Oct 2025). For Interact-RAG-like agents, these benchmarks indicate that richer action spaces will have to be tested under multi-turn, noisy, and multi-modal conditions.

Robustness and continual adaptation are equally central. AL4RAG uses active learning over conversation records to build preference data that teaches models to reject hallucination-prone “incapables” and answer manageable “capables,” and introduces a retrieval-augmented similarity metric tailored to RAG sample selection (Geng et al., 13 Feb 2025). TempPerturb-Eval studies the interaction between retrieved-context perturbations and decoding temperature, finding that high-temperature settings consistently amplify vulnerability to perturbations (Zhou et al., 1 Dec 2025). RAGShaper addresses agentic training-data scarcity by synthesizing distractor-rich tasks and constrained teacher trajectories that explicitly demonstrate error correction and noise rejection (Tao et al., 13 Jan 2026). Together, these results imply that Interact-RAG is not only a control-interface innovation; it is also a call for process-level supervision, adversarial training conditions, and evaluation regimes that expose failure modes beyond final-answer accuracy.

A further implication concerns retriever optimization. R3 trains retrievers inside the live RAG environment through reinforced contrastive learning, improving RAG performance by 5.2% over the original retriever and by 4.9% over state-of-the-art retrievers while requiring only 4 GPUs and a single day of training (Zhou et al., 28 Oct 2025). Interact-RAG changes the policy over retrieval operations; R3 changes the retriever itself. A plausible future synthesis would combine adaptive retriever training with explicit corpus-interaction policies, yielding systems that are jointly better at finding evidence and at deciding how evidence should be sought, filtered, and composed.