TenderMatch for Evidence Grounded Tender Recommendation with Fixed and Adaptive GraphRAG

Public procurement recommendation requires recovering relevant opportunities from sparse observations while preserving the distinction between recorded business relations and demonstrated supplier capability. We present TenderMatch, a graph retrieval augmented generation framework that requires every graph based ranking request to retrieve permitted evidence before recommending procurement lots. A fixed controller follows a prescribed history schedule; an adaptive controller selects one additional tool or stops after the mandatory retrieval. The study evaluates seven methods on TenderKG, comprising 6,126 procedures, 29,220 lots and 15,340 companies, using a grouped retrospective holdout with evaluation labels separated from retrieval. Matched validation favors adaptive retrieval, but the frozen 20-company test shows no consistent advantage: adaptive and fixed Recall@10 are 0.20435 and 0.20217, while nDCG@10 is 0.10722 and 0.11212. BM25 with E5 cosine reranking achieves the highest test nDCG@10, 0.11379. One adaptive length failure remains in the evaluation denominator. Every nonempty graph output passes identifier and evidence-path checks; these structural checks do not establish supplier eligibility or semantic sufficiency. The contribution is an explicit evidence boundary and a controlled comparison of retrieval policies, with separate analysis of candidate reach, ranking quality, controller behavior and failure coverage. The findings identify candidate retrieval and capability evidence as priorities for subsequent study, subject to incomplete award labels, uncertain historical text availability and small warm-history cohorts.