Towards Retrieving Interaction Spaces for Agentic Search

Abstract: Retrieval for search agents is still inherited from non-agentic information retrieval: a retriever ranks the corpus and the agent reads a small set of returned documents. Recent direct corpus interaction (DCI) work shows that agents can instead interact with the raw corpus through shell tools such as grep and file reads. But unbounded interaction does not scale: every broad shell command is a scan over the whole corpus, and latency degrades sharply as the corpus grows. We argue that the role of retrieval for agentic search is not just to select documents that fit in the LLM context window, but to construct an interaction space: a bounded subset of the corpus the agent can explore with associated tools. Two design consequences follow. The space needs a boundary supplied by retrieval, and the objects within it should be processed for interaction. As a proof of concept, we propose RISE (Retrieving Interaction SpacE): we use BM25 to construct the interaction space; meanwhile, its documents are processed during indexing for shell-style navigation. On BrowseComp-Plus, RISE matches the pure-shell DCI baseline at 78% accuracy with gpt-5.4-mini at roughly one quarter of the per-query cost. At 1M documents, RISE-BM25 reaches 81% on gpt-5.4-mini, whereas DCI on gpt-5.4-nano degrades to 60% with 33 of 100 wall-clock failures.

• The paper presents RISE, which reconceptualizes retrieval by constructing bounded interaction spaces with explicit workspaces for agentic search.
• It employs BM25 for workspace construction and LLM-based TOC extraction to enable efficient, targeted sub-document navigation.
• Experimental results show that RISE achieves comparable accuracy to traditional methods at lower computational cost and improved scalability.

Retrieving Interaction Spaces for Agentic Search: A Technical Overview

Problem Formulation and Motivation

This work refines the core IR interface for agentic LLM-driven search. Traditional retrieval—whether in RAG pipelines, browser-assisted QA, or ReAct-style agents—treats retrieval as snippet selection: feed the LLM a pre-ranked set of evidence fragments fitting within its context window. However, direct corpus interaction (DCI) approaches have shifted this paradigm by granting agents shell access to the corpus filesystem, enabling iterative exploration through compositional queries and programmatic inspection. Such methods expose stark limitations: truly unbounded shell-based corpus interaction does not scale due to corpus-size linear traversal costs, while reverting to snippet retrieval loses the benefits of iterative exploration and evidence synthesis.

The critical hypothesis advanced is that retrieval's output should not be raw context or unrestricted corpus access, but a bounded interaction space: a persistent, explorable subset of the corpus coupled to agent shell tools and augmented for efficient sub-document navigation. Two key requirements are formalized:

1. Explicit Boundary: Retrieval must define an explicit, persistent working set distinct from context windows or implicit rankings, ruling out both unbounded DCI and classical snippet-based RAG.
2. Processed Objects: Returned documents require structural augmentation, enabling direct sub-document access (e.g., via line-numbered TOCs), eschewing raw-text-only or destructively pre-chunked approaches.

RISE Methodology

The RISE system instantiates these requirements as follows:

• Bounded Workspace Construction: BM25 supplies the top-$K$ results across compositional sub-queries, hardlinking the union into a dedicated working directory. This per-query workspace persists throughout the agent’s turn, enabling shell-based commands (rg, grep, cat, read) with uniform file access semantics, and is monotonic across time.
• Navigational Augmentation: At indexing time, each document undergoes a TOC extraction phase using an LLM-based section boundary prompter. The result is a body-preserving structural augmentation: YAML frontmatter, TOC blocks with section line ranges, and anchor-inserted headings. Consequentially, the agent can efficiently enumerate or directly access relevant sections via deterministic line offsets, reducing the cost of evidence localization.

  Figure 1: RISE workflow—BM25 retrieval bounds the interaction set, documents are indexed with line-numbered TOCs, and agents iteratively search, shell, and read within the bounded workspace.

Experimental Results and Numerical Validation

Evaluation is conducted primarily on BrowseComp-Plus, leveraging a 100-query subset with two corpus scales (100k and 1M documents, the latter augmented with FineWeb-Edu distractors). Agent configurations include gpt-5.4-mini, gpt-5.4-nano, and mimo-v2.5-pro, with comparison to both DCI and snippet-retrieval baselines.

Key outcomes:

• Cost-Accuracy Tradeoff: On 100k, RISE with gpt-5.4-mini achieves $78\%$ accuracy at \$0.28 per query, matching DCI accuracy but with roughly one-quarter of the per-query compute budget and API spend. Mimic performance is consistent across model families.
• Scalability: At 1M corpus size, RISE-BM25 maintains or slightly improves accuracy (up to $83\%$), while DCI experiences pronounced degradation ($71\% \to 60\%$ accuracy on nano, with 33/100 wall-clock failures). Cost and tool-invocation statistics highlight the sharply better efficiency and robustness of workspace-bounded interaction.

  Figure 2: Accuracy-cost frontier for RISE and DCI on BrowseComp-Plus shows RISE maintains equivalent accuracy at a fraction of computational cost, especially as corpus scale increases.

• Ablation: Document structural augmentation via TOCs confers $1$–$4$ point accuracy improvements on top of the boundary mechanism over multiple models, confirming its additive value beyond mere workspace bounding.

Agent Trajectory Case Studies

Qualitative analysis demonstrates the two-layered approach’s efficacy. For obfuscated queries, agents leveraging RISE-BM25 compose multi-faceted BM25 sub-queries to populate the workspace, then verify answers using shell tools. With full RISE, agents exploit TOC structure to directly jump to informative spans, drastically reducing unnecessary document reads.

Figure 3: Case study—(a) RISE-BM25 trajectory relies on multi-clue sub-queries and workspace verification; (b) RISE leverages TOC to map high-level cues to targeted, minimal-range reads.

Theoretical and Practical Implications

The results fundamentally challenge the dichotomy between snippet retrieval and unrestricted corpus access. By formalizing the interaction space as a retrieval output, RISE reframes corpus interface design: the retriever’s role broadens from document ranking to workspace curation, and document pre-processing becomes critical to agent efficiency.

Practically, this approach enables scalable, robust, and cost-efficient agentic search, especially as corpus sizes expand beyond the thresholds feasible for naive DCI. Theoretically, this paradigm may motivate new benchmarks and metrics focusing on interaction-space quality and exploration efficiency, rather than just retrieval accuracy or context R@K.

Future Directions

Obvious next steps include:

• Retriever Generalization: Substituting BM25 for dense, hybrid, or late-interaction architectures to test generalizability of the interaction-space design.
• Structural Variants: Augmenting or replacing TOC-based navigation with hierarchical segmentations, learned anchor graphs, or paragraph-level pointers.
• Benchmark Evolution: Designing evaluation datasets that stress the need for interaction-space quality rather than context window fit.
• Deeper Agent Integration: Studying co-adaptation of agent plans with alternative interaction-space structures, possibly incorporating agent-driven workspace expansion and self-pruning.

Conclusion

RISE advances the agentic search paradigm by decoupling retrieval from context constraints and unscalable corpus-level shell interaction. By bounding the agent's workspace via explicit retrieval and structurally augmenting candidate objects for shell-explorable evidence localization, RISE offers a scalable, cost-effective, and robust framework for LLM-based research agents. This framework foregrounds the design of interaction spaces as a core architectural variable in retrieval-driven agent research, underscoring the need to reevaluate retrieval benchmarks and agent-corpus interfaces in this light.