BeamRAG: Query-Guided Heuristic Beam Search with Macro-Gating on Knowledge Graphs

Dense retrieval over isolated passages can miss evidence distributed across multiple documents. Graph-based retrieval offers complementary relational structure, but unguided expansion can introduce irrelevant neighbors and cross-document semantic drift. We present BeamRAG, a query-guided retrieval pipeline combining bounded heuristic beam search, degree attenuation, hop decay, document-provenance gating, and graph–text mention-alignment reranking. The evaluated configuration uses beam width B=6, a maximum of four hops, node similarity threshold 0.400, and document relevance threshold 0.65. Relational, vector, and metadata assets are stored in PostgreSQL; traversal is orchestrated in the application tier. Unified storage is an operational design choice, not an exclusive capability of BeamRAG. The evaluation encompasses 500 curated queries across four UltraDomain corpora, evaluated across four retrieval architectures and two complementary generator–judge configurations (Track 1: Gemma 26B generator → Qwen 27B judge; Track 2: Qwen 27B generator → Gemma 26B judge). The benchmark comprises 4,000 generated system responses (500 queries × 4 systems × 2 tracks) and 3,000 pairwise judge evaluations (500 queries × 3 baseline comparisons × 2 tracks). Of the 500 queries, 125 MIX queries served as a diagnostic development set, while 375 queries across Computer Science, Agriculture, and Legal represent frozen holdout evaluations. Reported tie-adjusted win rates are 73.6% and 80.4% against dense-only retrieval, 60.4% and 63.5% against shallow BFS, and 77.7% and 85.2% against the evaluated LightRAG deployment. These rates measure comparative synthesis preferences under length-controlled rubrics, not source-verified factual accuracy. Baselines differ in native retrieval mechanics and context budgets; component-level causal attribution is not established. The reported BeamRAG retrieval-pipeline time is approximately 68 ms, excluding answer generation; its 16.35 ms search/traversal sub-timer reflects the core traversal stage and is not directly comparable to LightRAG's complete multi-stage retrieval time.