PaperlayerView exact revision on arXiv

Source-linked AI summary

Reimagining Enterprise RAG for Multi-Agent Systems: A Knowledge Layer That Organizes Before It Retrieves

Garima Agrawal, Prasun Das, Akshay Sarvade, Priyanka L, Minisha N, Sravani Joshna, Hemath Manivannan, Lavanya Korada, Sambhrit Pattanayak, Srashti Gupta, Siddu R, Prasad Kalyanasundaram

arXiv:2610.32851v1cs.AI

TL;DR

Machine-generated enterprise requirements can be hard to retrieve precisely when shared vocabulary masks operational differences across services. KB Cataloging organizes artifacts by their production context and constrains retrieval to scoped artifact identifiers. The paper reports support for more than 100 production-defect analyses and resolutions under developer verification, while limiting its claim to machine-generated artifacts that retain provenance.

  • Problem

    Shared vocabulary makes rules across services textually similar but operationally distinct, while flat vector retrieval can return a wrong-service rule without signaling the mismatch.

    Source: Introduction · p. 1

  • Method

    KB Cataloging organizes provenance-bearing artifacts in a catalog, registry, and hypergraph, then constrains vector retrieval to artifact identifiers selected through that structure.

    Source: Introduction · p. 2

  • Results

    More than 100 production defects were analyzed, with accepted results reviewed by the affected service’s developers and written back into the issue tracker.

    Source: Deployment and Evaluation · p. 5

  • Takeaways & Limitations

    The deployed layer serves production SDLC agents and practitioners directly, and its artifact-level organization excludes adjacent-service rules from the candidate set rather than merely ranking them lower.

    Source: Limitations, Path Forward, and… · p. 6

    The approach depends on machine-generated artifacts retaining provenance; it claims no generality for heterogeneous human-authored material, and retrieval precision is bounded by placement quality.

    Source: Limitations, Path Forward, and… · p. 6

Abstract

from arXiv · show

Enterprises are deploying agentic AI across the software development lifecycle, and they need the behavioral logic of the systems they act on, not stakeholder documentation. A seven-agent reverse-engineering pipeline recovered this logic, processing 241 repositories across 18 services into thousands of artifacts describing scope, journeys, rules and acceptance criteria. Recovery does not confer usability. Entities like offer, eligibility, and activation recur across services under a shared vocabulary, related in name but operationally distinct, so one end-to-end journey implicates over a thousand near-identical rules that must not be conflated. Flat vector retrieval fails, returning an identical rule from the wrong service at high similarity with no signal of error. Content-level ontology, the traditional remedy, did not exist, and relating every occurrence of recurring entities would have cost more than the engagement and not survived a release cycle. We place the ontology at the artifact level. In this machine-generated corpus, each artifact retains its producer, component, service, domain and type, provenance that disambiguates before content is read. We present a knowledge layer that organizes before retrieving: a self-describing catalog, a sparse metadata registry, and a navigation hypergraph defined over artifacts rather than extracted entities. Queries resolve to artifact identifiers that constrain vector search, so the competing rule from another service is absent rather than outranked, making disambiguation structural, not statistical. Built in under a week and admitting new services by convention, the layer serves SDLC agents, developers, quality engineers and analysts. It supported root-cause analysis and resolution of 100+ production defects with developer verification and attributed 35 release regression failures to specific rules and services.

Introduction

Machine-generated requirements artifacts capture software behavior, but shared vocabulary makes precise retrieval across services difficult. The paper proposes organizing artifacts before retrieval, using their production context to disambiguate them without modeling every entity occurrence.

Introduction

  • Shared terms such as offer and eligibility recur across services, making end-to-end journeys span over a thousand textually similar but operationally distinct rules.The estate comprises 14 workflows, and one workflow may invoke up to 112 endpoint-level requirements documents.Source: Introduction · p. 1
  • Flat vector retrieval can return a high-similarity rule from the wrong service without signaling that its evidence is misattributed.The returned text may be well-formed and topically correct, leaving the consuming agent unaware of the mismatch.Source: Introduction · p. 2
  • Content-level ontology would require aligning recurring entity occurrences across services, an effort the authors say would exceed the engagement and quickly become outdated.The cited approaches are Gruber 1993 and Noy and Musen 2000.Source: Introduction · p. 2
  • The corpus contains structured artifacts with known producers, components, services, domains, types, and source references, so disambiguating structure is available before reading content.The upstream pipeline processed 241 repositories across 18 services into artifacts covering scope definitions, journeys, rules, acceptance criteria, and gap registers.Source: Introduction · pp. 1–2
  • KB Cataloging relocates ontology to the artifact layer and constrains vector retrieval to artifact identifiers, excluding competing rules from adjacent services before similarity is computed.The approach is presented as an artifact-level alternative to modeling the entities in the corpus.Source: Introduction · p. 2
  • The framework combines a catalog, sparse registry, artifact hypergraph, and identifier-constrained retrieval, and reports production use supporting defect resolution and regression attribution.The contribution description reports more than 100 defects reviewed under owning-developer verification and 35 regression failures attributed to specific rules and services.Source: Introduction · p. 2

Design: The KB Cataloging Framework

The framework settles artifact classification, placement, relationships, and metadata at ingestion, then navigates that structure to scope retrieval before similarity search. A confidence gate can widen retrieval when the scoped evidence is insufficient.

Design: The KB Cataloging Framework

  • At query time, intent classification and hypergraph navigation select artifact identifiers before constrained vector retrieval and a confidence check.Figure 1 depicts ingestion updating the artifact store, registry, and navigation hypergraph in synchrony, followed by query-time scoping.Source: Design: The KB Cataloging Framework · p. 2, Figure 1 · p. 3
  • Artifacts are renamed, registered with producer-supplied metadata, placed in a service-domain-feature node, and stored in a matching folder hierarchy.These steps keep the artifact store, registry, and navigation structure in correspondence.Source: Design: The KB Cataloging Framework · p. 3
  • Classification does not require content inference because producing agents already establish service, domain, and feature, while node relations follow service and repository structure.The layer retains structure already established upstream rather than reconstructing it afterward.Source: Design: The KB Cataloging Framework · p. 3
  • The registry combines creation-time identity and provenance fields with usage-derived task, intent, co-retrieval, and use-frequency fields; unknown fields are absent rather than null.The schema caption states that this lets the schema grow across services without migration.Source: Design: The KB Cataloging Framework · p. 3, Table 1 · p. 4
  • Hyperedges connect the set of service nodes participating in a shared workflow, while service-scoped nodes keep another service’s similarly named rules out of candidates.The order-submission example is one n-ary relation among ordering, catalog, and billing, rather than three pairwise links.Source: Design: The KB Cataloging Framework · p. 4, Figure 2 · p. 4
  • Identifier filtering removes the rest of the corpus from similarity computation, so a near-identical rule from another service is absent rather than outranked.The filter applies to the vector query itself, and retrieval cost scales with the scoped set.Source: Design: The KB Cataloging Framework · p. 4
  • The confidence gate widens to full-corpus retrieval when scoped results are insufficient, guarding against errors from intent classification, aliases, or artifact placement.The framework verifies scoped evidence rather than trusting the scope unconditionally.Source: Design: The KB Cataloging Framework · p. 4
  • Runtime context graphs record inferred intent, navigation, consulted artifacts, contributing chunks, and source provenance so developers can verify proposed root causes.The authors describe these graphs as important to adoption and as an organizational audit trail.Source: Design: The KB Cataloging Framework · p. 4

Deployment and Evaluation

The framework was built and extended without schema work, while deployed agents used its artifacts in developer-reviewed defect analysis and regression attribution. Retrieval accuracy remained under evaluation through widening-rate and adjudicated ambiguity-query measurements.

Evaluation Scope

  • The framework is evaluated for construction and extension cost, practitioner acceptance, and retrieval accuracy, and is used conversationally by practitioners and programmatically as an agent retrieval substrate.Defect analysis and release-regression attribution have the most operational history among the agent uses described.Source: Deployment and Evaluation · p. 5

Deployment

  • The knowledge layer was built in under a week, and all 18 ingested services were admitted without schema work, ontology maintenance, or manual cataloging.Artifacts arrive classified by their producing agent, from which nodes and hyperedges follow.Source: Deployment and Evaluation · p. 5

Defect Root Cause and Resolution

  • More than 100 production defects were analyzed during defect-fix release cycles, with affected-service developers reviewing each analysis and accepted results written to the issue tracker.The knowledge-layer agent combines issue-tracker and log evidence with retrieved artifacts.Source: Deployment and Evaluation · p. 5

Other Agents Using the Knowledge Layer

  • 35 release regression failures were attributed to a specific rule, journey, or service by a separate agent that required no knowledge-layer modification.Other agents also assess test coverage, retrieve dependencies for changed components, and support requirement brainstorming and story breakdown.Source: Deployment and Evaluation · p. 5

Retrieval Behavior

  • Retrieval accuracy was still under evaluation: the planned measures were how often confidence-gate widening returns to the full corpus and an adjudicated comparison on ambiguity-weighted queries.The widening rate indicates how often scoping is sufficient; the adjudication requires practitioners familiar with the services.Source: Deployment and Evaluation · pp. 5–6

Related Work

The paper places its contribution upstream of vector retrieval: it organizes machine-generated artifacts rather than reconstructing structure from their content. It distinguishes this approach from content-level ontology, entity-graph retrieval, and systems that assume organized specifications already exist.

Retrieval-augmented generation

  • RAG remains the retrieval foundation; the paper adds an organization layer that narrows candidates before similarity search and is compatible with embedding models, rerankers, and hybrid search.The contribution is upstream of the retrieval mechanism, not a replacement for it.Source: Related Work · p. 6

Ontology engineering

  • Rather than rejecting ontology, the paper treats its level as a design choice and classifies artifacts, which producing agents can supply largely without cost in a machine-generated corpus.The related-work discussion notes that vocabulary alignment is semi-automatic at best and scales with concepts requiring disambiguation.Source: Related Work · p. 6

Graph-based retrieval

  • Unlike GraphRAG and LightRAG, which induce entity graphs from text, the paper defines its hypergraph over provenance-fixed artifacts to avoid reconstructing distinctions between colliding entity mentions.It describes knowledge graphs for hallucination reduction as complementary: the organized source could be queried by such approaches.Source: Related Work · p. 6

Multi-agent software engineering

  • The paper complements multi-agent systems that presume specifications or issue descriptions exist by addressing how recovered specifications are organized once they exist at scale.The cited software-engineering systems include ChatDev, MetaGPT, and SWE-agent; the paper positions recovery and organization as complementary stages.Source: Related Work · p. 6

Limitations, Path Forward, and Conclusion

The approach is bounded by machine-generated artifacts with retained provenance, and its deployment is a single estate. The paper concludes that artifact-level organization uses existing provenance to constrain retrieval, while outlining additional agents and quality instrumentation as future development.

Limitations

  • The approach depends on machine-generated artifacts retaining provenance; heterogeneous human-authored material would require content-based classification, and no broader generality is claimed.Precision is bounded by placement quality, and behavior absent from upstream artifacts is absent from retrieval.Source: Limitations, Path Forward, and… · p. 6
  • The deployment covers a single estate, with outcomes adjudicated by practitioners familiar with its services, a necessary condition for competent judgment that may introduce optimism bias.Source: Limitations, Path Forward, and… · p. 6

Path Forward

  • The planned fuller serving layer adds composition, boundary, and curation agents; nearer-term work includes per-agent retrieval-quality instrumentation and re-ingestion after code changes.The design also admits brand as an additional classification axis without structural change if the estate expands to multiple brands.Source: Limitations, Path Forward, and… · p. 6

Conclusion

  • The conclusion argues that when a corpus carries provenance, artifact-level organization can constrain retrieval so a rule from an adjacent service is absent from the candidate set rather than outranked.The deployed estate spans 241 repositories across 18 services, with more than 100 production defects analyzed and resolved under developer verification.Source: Limitations, Path Forward, and… · p. 6

Related papers

Loading 2610.32851v1…