Two business rules can use almost identical words while belonging to different services. Reimagining Enterprise RAG for Multi-Agent Systems studies that problem in a corpus recovered from 241 repositories across 18 services by a seven-agent reverse-engineering pipeline.
Use the structure already attached to the artifact
The corpus contains scope descriptions, journeys, business rules and acceptance criteria. Terms such as offer, eligibility and activation recur across services, but their operational meaning is not interchangeable. A high similarity score can therefore return a plausible rule from the wrong service.
The proposed knowledge layer retains the artifact's producer, component, service, domain and type. A self-describing catalog, sparse metadata registry and navigation hypergraph organize these artifacts before content retrieval begins.
Constrain the candidate set before ranking
A query first resolves to artifact identifiers, which constrain the vector search. The competing rule from another service is excluded from the candidate set instead of merely receiving a lower similarity score. The paper places the organizing relationships at the artifact level rather than attempting to infer and maintain every recurring content-level entity relationship.
This approach depends on useful provenance being available. It addresses a particular ambiguity in a machine-generated corpus; it does not establish that metadata will resolve every ambiguous question or that all remaining candidates within a service are correct.
Read the production findings as a case study
The authors say the layer supported root-cause analysis and resolution of more than 100 production defects with developer verification, and attributed 35 release regression failures to specific rules and services. These are reported operational uses, rather than a controlled estimate of how many defects the architecture prevents.
For another organization, the first test is whether service and artifact boundaries are reliable enough to constrain retrieval without hiding necessary evidence. Preserve links back to the source code and verify the resulting rule with the owning team. The contribution is a practical way to use existing corpus structure before asking similarity search to choose among near-identical text.
Comments