PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Retrieving Interaction Spaces for Agentic Search

Shengyao Zhuang, Yuansheng Ni, Hengxin Fun, Jimmy Lin, Xueguang Ma

arXiv:2606.06880v1cs.IR

TL;DR

Agentic search needs a bounded, navigable interaction space rather than either context-limited snippets or unrestricted corpus access. RISE constructs this space with BM25 and indexed navigation metadata, matching DCI accuracy at substantially lower cost while remaining stable at 1M documents.

  • Problem

    Unbounded corpus interaction scales poorly, while conventional snippet retrieval limits evidence resolution, motivating retrieval that constructs a bounded interaction space.

  • Method

    RISE uses BM25 to define an interaction-space boundary and offline structural metadata to make selected documents shell-navigable.

  • Results

    78% accuracy matches pure-shell DCI at one quarter of its per-query cost, while RISE-BM25 remains stable at 1M documents as DCI degrades.

  • Takeaways & Limitations

    An interaction space offers an intermediate retrieval output that lets agents re-explore bounded candidates with shell tools and efficiently reach sub-document evidence.

  • Takeaways & Limitations

    The processing consequence is evaluated only on the 100k corpus, so the 1M scaling result primarily measures the boundary mechanism.

Abstract

from arXiv · show

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.

1 Introduction

RISE reframes retrieval for agentic search as constructing a bounded, tool-equipped interaction space rather than merely selecting documents for the LLM context. It addresses the scaling costs of unbounded direct corpus interaction while matching DCI accuracy at substantially lower cost and remaining stable on a 1M-document corpus.

  • Background: Agentic search has traditionally inherited non-agentic retrieval, where a retriever ranks the corpus and returns a small document or snippet set for the agent’s context [Lewis et al., 2020; Nakano et al., 2021; Yao et al., 2023].Recent work instead enables raw-corpus shell interaction, iterative lexical search, or file-based manipulation [Li et al., 2026b; Cao et al., 2026; Sen et al., 2026].
  • Motivation: Unbounded direct corpus interaction does not scale because broad shell commands scan the whole corpus, increasing tool use, latency, and cost while reducing accuracy [Li et al., 2026b].On a 100-query BrowseComp-Plus subset, expanding from 100k to 200k documents raises average tool calls from 38.5 to 86.9, drops accuracy by 13.6 points, and reaches 37.5% accuracy with 20 of 100 maximum-budget queries at 400k documents.
  • Framing: RISE constructs an interaction space: a bounded corpus subset that agents can explore with associated tools, with its boundary supplied by retrieval.This framing differs from retrieval that only returns documents or snippets fitting the LLM context window.
  • Results: 78% accuracy lets RISE match the pure-shell DCI baseline with gpt-5.4-mini at one quarter of the per-query cost on BrowseComp-Plus.At 1M documents, RISE-BM25 remains stable, whereas DCI falls to 60% with 33 of 100 wall-clock failures.
  • Results: Removing document processing lowers RISE-BM25 accuracy by 1–4 points across model tiers at 100k documents, supporting the value of processing documents for interaction.This isolates the contribution of the interaction-processing consequence at the evaluated scale.

2 Related Work

Related work frames agentic search as iterative interaction with external information, evaluates it through fixed-corpus browsing benchmarks, and increasingly adapts retrieval and corpus interfaces to agents. This paper focuses on which corpus interface agents should receive after retrieval selects an initial evidence set.

  • Retrieval-augmented and deep research agents: Agentic search extends retrieval-augmented generation and action–observation loops into longer-horizon research workflows, with agents repeatedly searching and inspecting evidence [Lewis et al., 2020] [Nakano et al., 2021, Yao et al., 2023].
  • Retrieval-augmented and deep research agents: BrowseComp and BrowseComp-Plus evaluate browsing-style agents, with BrowseComp-Plus using a fixed corpus that makes retrieval choices comparable [Wei et al., 2025, Chen et al., 2025].
  • Retrieval and ranking for agentic search: AgentIR and Agentic-R adapt retrieval to agent behavior by using reasoning traces or combining local passage relevance with global answer correctness [Chen et al., 2026] [Liu et al.,
  • Corpus interfaces for agentic search: Corpus-interface research distinguishes ranked snippets delivered through the context window from approaches that remove the candidate limit, while retrieval success also depends on the agent harness and information delivery [2026].

3 Method

RISE constructs an agent interaction space with two layers: retrieval supplies a persistent, bounded candidate set outside the context window, and offline processing makes its documents navigable with shell tools. RISE-BM25 isolates the boundary layer, while full RISE combines both consequences.

  • Method overview: RISE combines a persistent candidate workspace with navigable document representations, while RISE-BM25 implements only the bounded-workspace layer.The workspace is re-traversable with shell tools, and documents are processed so tools can reach sub-document evidence without scanning whole files.
  • Bounded workspace: The workspace boundary rules out full-corpus shell scans, context-window snippet limits, and candidate lists that cannot be re-traversed with native tools.RISE-BM25 treats the imported candidates as a filesystem outside the model context window.
  • Navigable objects: RISE processes retrieved documents offline into line-numbered tables of contents with section anchors, enabling broad browsing and targeted line-range reads.At inference time, structured counterparts are hardlinked into the workspace; the agent uses the TOC for navigation and verifies committed facts against body text.
  • Bounded workspace: The bounded workspace is a per-query filesystem populated by the union of top-K documents retrieved for one or more natural-language BM25 sub-queries.Only search can import documents; the workspace grows monotonically, while full matches remain available beyond the returned top-10 previews.
  • Navigable objects: The navigable-object layer avoids raw end-to-end document scans and retrieval-time chunking or snippet delivery that preselects the evidence visible to the agent.TOCs are used because they are a cheap in-place metadata form supporting both broad browsing and precise reads.

4 Experimental Setup

Experiments evaluate RISE and controlled baselines on a fixed 100-query BrowseComp-Plus sample across 100k- and 1M-document corpora. The study compares agent models, judged accuracy, efficiency, tool use, evidence coverage, and stopping behavior under specified computational budgets.

  • Execution environment: Runs execute separately on a 24GB Apple M4 MacBook Air with four workers, allowing at most four concurrent queries per run.Methods are not co-scheduled, reducing direct contention between systems during execution.
  • Compared systems: The comparison includes full RISE, boundary-only RISE-BM25, DCI, and a retrieval-agent baseline, with RISE variants retrieving 1,000 documents per sub-query and showing top-10 previews.RISE uses BM25 bounding plus offline TOC processing, whereas RISE-BM25 retains original documents in the workspace; both use bm25s defaults k1=1.5 and b=0.75.
  • Budgets and stopping conditions: RISE, RISE-BM25, and the retrieval-agent baseline receive 100 model calls and a one-hour cap, while DCI receives 300 calls and a 1.5-hour cap.The reproduced baselines follow their original budget and stopping settings; DCI receives a larger allowance because full-corpus shell scans are I/O-bound.

5 Results

RISE achieves strong accuracy with lower cost by bounding the agent’s interaction space and processing documents for shell navigation, while remaining robust as the corpus scales. Its best configuration reaches 82% accuracy, and the 1M-document results show that boundary-only RISE-BM25 scales better than DCI.

  • 100k results: RISE-BM25 is close behind full RISE on the larger model tiers, showing that bounding the interaction space supplies most of the gain while indexing-time processing adds a smaller improvement.On gpt-5.4-mini, DCI ties RISE but costs 4× more per query; on mimo-v2.5-pro, DCI has 18 of 100 wall-clock failures and reaches 60% accuracy.
  • 5 Results: 82% accuracy makes RISE with gpt-5.4 medium the most accurate configuration overall, while RISE with gpt-5.4-mini and mimo-v2.5-pro reaches 78% at lower cost than retrieval-agent and DCI.The medium configuration has cov_mean=92.4%; its gain over gpt-5.4-mini is modest (+4 points) and costs $1.25 versus $0.28 per query, or 4.4× more.
  • Scaling: At 1M documents, RISE-BM25 improves from 75% to 83% on mimo-v2.5-pro, from 77% to 81% on gpt-5.4-mini, and from 64% to 65% on gpt-5.4-nano.These results measure the boundary mechanism only because offline document processing was not extended to the added 900k FineWeb-Edu distractors.
  • Scaling: At 1M documents, DCI on nano drops from 71% to 60% and wall-clock failures rise from 7 to 33 queries, whereas RISE-BM25 does not degrade.DCI often spends its budget triaging broad shell outputs before answering, so its lower reported API cost is not an efficiency gain.
  • Top-K ablation: K=1000 is the best BM25 top-K setting for gpt-5.4-mini, within one point of the best mimo-v2.5-pro setting, and tied on nano up to one point.It yields accumulated workspaces of roughly 7.6k–10.4k files; larger K does not monotonically improve accuracy, and the main tradeoff is recall versus local navigation overhead.

6 Case Studies

The case studies show RISE agents combining retrieval-shaped workspaces with indexing-time document structure for in-place verification. On obfuscated multi-clue questions, agents repeatedly expand candidate workspaces before using shell tools to verify targeted evidence.

  • Case Studies: Figure 3 shows two correct gpt-5.4-mini trajectories on the 100k corpus for obfuscated, multi-clue BrowseComp-Plus questions whose answer entities are never named.Both trajectories follow the intended workflow end to end.
  • Case Studies: In the RISE-BM25 case, the agent decomposes an unknown-target question into fifteen paraphrased sub-queries across five search calls, unions the results, and verifies within the workspace using rg and read.Retrieval repeatedly shapes the workspace, while shell tools verify evidence in place.
  • Case Studies: In the RISE case, an LLM-generated, line-numbered table of contents lets the agent map a clue to the Acknowledgements entry, read the relevant lines, and confirm the journal from the title header.This trajectory illustrates how indexing-time processing turns long documents into navigable structure rather than requiring full-body reading.

7 Conclusion

The conclusion frames retrieval for agentic search as constructing a bounded, navigable interaction space rather than extracting snippets or exposing the unrestricted corpus. RISE instantiates this design with BM25 boundaries and offline document processing, achieving a favorable accuracy–cost tradeoff and greater scalability than pure-shell DCI.

  • RISE design: RISE constructs an interaction space by using BM25 to supply a boundary and offline-augmenting documents with line-numbered tables of contents for shell navigation.This is presented as a proof-of-concept instantiation of retrieval that returns an explorable space rather than only context.
  • Empirical results: RISE matches pure-shell DCI at a fraction of its per-query cost, remains stable from 100k to 1M documents, while pure-shell DCI degrades with wall-clock failures.On BrowseComp-Plus, the interface delivers the reported accuracy–cost tradeoff; snippet retrieval-agent performance is weaker on larger model tiers.
  • Broader conclusion: The broader design target is an intermediate retrieval output: bounded candidates that agents can re-explore with shell tools and whose structure makes sub-document evidence cheap to reach.This combines the boundary consequence with the processing consequence, making corpus interaction part of retrieval-system design.

8 Limitations

The study instantiates its design consequences with BM25 boundaries and line-numbered TOCs, while leaving alternative retrieval and metadata schemes unexplored. Its evaluation is also narrow, covering one benchmark family, 100 queries, closed-weight agents, and one judge, so small accuracy gaps are directional rather than precise.

  • 8 Limitations: RISE evaluates the boundary mechanism only with BM25 and document processing only with line-numbered TOCs, leaving broader design-space coverage for future work.Natural follow-ups include dense, late-interaction, or hybrid retrievers and section graphs or paragraph-level anchors; the processing consequence is evaluated only on the 100k corpus.
  • 8 Limitations: The evaluation covers only BrowseComp-Plus, 100 queries, closed-weight agents, and a single judge, gpt-5.1, which shares a model family with most agents and the structured-corpus generator.These choices narrow the scope of the empirical evidence.
  • 8 Limitations: Per-tier accuracy gaps of a few points should therefore be interpreted as directional rather than precise.This qualification follows from the evaluation’s narrow benchmark, agent, query, and judging setup.

A RISE-BM25 System Prompt

The RISE-BM25 system prompt frames the agent as answering research questions over a large unseen corpus using iterative search, shell exploration, and file reading. It instructs the agent to accumulate retrieved documents, inspect them with relative-path tools, and return a structured final answer when confident.

  • The agent answers research questions over a large document corpus it cannot see directly.
  • The prompt provides three tools—search, bash, and read—for retrieving documents, exploring the working directory, and inspecting files.
  • Search matches bag-of-words queries, returns per-query top-10 previews, and adds the full match sets to an accumulating working directory.Queries should use several distinctive terms and can be issued in complementary groups for broader coverage.
  • Bash supports commands such as rg, grep, ls, find, cat, and head only over the relative-path working directory, while read inspects a file by its full path.
  • The agent should iterate between search and file inspection, then stop calling tools after outputting an explanation, exact answer, and confidence percentage.

B Structured Documents: Prompt and Example

RISE augments retrieved documents with navigable TOCs and a prompt that treats those summaries as indices rather than evidence. The structured format combines metadata, section ranges, and body text to support shell-style exploration while preserving factual verification in the document body.

  • B Structured Documents: Prompt and Example: RISE’s prompt explains the TOC-augmented format and requires agents to verify facts in document bodies rather than TOC summaries.The system uses the same search and shell interface as RISE-BM25, with added instructions for navigating structured files.
  • B Structured Documents: Prompt and Example: 94.5% of 100,195 documents received at least one validated TOC entry, while 99.3% of 808,062 proposed section anchors were located.Documents with entirely invalid anchors or unparseable output fall back to original text with an empty TOC.
  • B.1 RISE System Prompt (with Document Processing): Search returns previews while adding each query’s full match set to the working directory, enabling iterative exploration with bash and read.Multiple complementary queries can broaden coverage, and retrieved documents accumulate across turns.
  • B.1 RISE System Prompt (with Document Processing): Each structured document contains YAML metadata, a line-ranged Table of Contents, a body sentinel, and body headings matching the listed sections.The format supports section-level navigation while retaining the literal document body for verification.
  • B.1 RISE System Prompt (with Document Processing): Agents must read relevant sections fully, expand the read range when necessary, and avoid committing to facts based only on plausible TOC entries.TOC summaries may omit details, lose nuance, or be slightly inaccurate; short or single-topic documents are read normally.
  • B.2 Example Structured Document: The example structured document for Bath Spa University shows frontmatter, section summaries with line ranges, body headings, and corresponding historical text.Its sections include institutional heritage, the Bath School of Art era, and the Locksbrook Campus expansion.
  • B.2 Example Structured Document: The example demonstrates that TOC entries navigate diverse body sections, including Bath Spa’s 1852 art-school origins and 2019 Locksbrook Campus opening.The body supplies the underlying historical details that the TOC summarizes.

C Structured Corpus Generation Prompt

The prompt instructs an agent to propose navigable section boundaries for an unchanged document, producing headings, locators, and descriptions for downstream insertion and table-of-contents generation. It enforces exact, unique anchors and adaptive sectioning based on document structure and type.

  • Prompt purpose: The prompt proposes section boundaries, short content-descriptive headings, and one-sentence descriptions without rewriting the original document.A downstream script inserts headings into the original text and builds a line-numbered table of contents from the output.
  • Anchor rules: Anchors must be unique 6–12-word substrings copied exactly from the document and listed in section order.The first anchor begins the document body, anchors cannot be paraphrased, and sections lacking a verbatim anchor should not be proposed.
  • Sectioning guidelines: Sectioning follows natural topic shifts, with ordinary prose divided moderately and catalogs divided aggressively by independent item or grouping.Dependent procedural steps remain under one umbrella section, while frontmatter and insubstantial trailing boilerplate are excluded.
  • Output quality: Descriptions must be concrete 10–25-word summaries that mention key entities, dates, places, or numbers when relevant.Titles should describe section content rather than use generic labels.
Loading 2606.06880v1…