An Agent Cascade for Explainable Relevance Assessment of Procurement Procedures in the Tenders Electronic Daily Environment
2227-7390
Abstract
This paper considers the task of assessing the relevance of a procurement procedure for a supplier under conditions of growing volumes of open procurement data and increasing complexity of electronic publication services. It is shown that, for a supplier, relevance cannot be reduced to a match between a procurement notice and a search query or classifier code. In the European procurement context, procedure assessment requires consideration of the procurement object, lot structure, selection and award criteria, European Single Procurement Document requirements, procedure language, place of performance, participation through a specific legal entity, contract terms, data-protection requirements, service obligations, risk signals, and retrospective context. A formal mathematical representation of relevance assessment is developed, and on this basis a cascade agent model is proposed in which relevance is defined as the result of matching a procurement procedure, a supplier profile, and a set of evaluation rules. The model includes specialized agents for query normalization, procedure search, primary filtering, documentation preparation, fact extraction, constraint identification, procurement-object identification, customer analysis, contract-terms analysis, supplier-profile matching, dossier checking, relevance calculation, result interpretation, and final-card generation. The key intermediate result is an analytical procurement dossier containing structured facts, constraints, risk signals, matching results, and evidential links to sources. The proposed model is implemented as a research prototype and evaluated on an active stream of procurement procedures published through TED. The prototype evaluation shows that the cascade can transform a broad and heterogeneous set of retrieved procedures into a multilevel operational funnel comprising a shortlist, comparison pool, manual-review routes, and a monitoring layer. Within the reported case, these outputs demonstrate execution of the intended cascade sequence and generation of explainable user-specific assessments; they are not presented as evidence of ranking accuracy, superiority over simpler alternatives, or generalizability.