BeamRAG v2: Alias-Aware Adaptive Beam Traversal with Thresholded Hub Regularization for Multi-Hop GraphRAG

Knowledge Graph-augmented Generation (GraphRAG) is a promising approach for overcoming flat vector retrieval limitations in multi-hop question answering over complex enterprise corpora. While first-generation BeamRAG demonstrated substantial gains over Breadth-First Search (BFS) and dense baselines through bounded heuristic beam search and document-level macro-gating, empirical profiling identified key bottlenecks in alias redundancy, discontinuous degree regularizers, and coarse document-level gating. In this paper, we propose BeamRAG v2, an integrated second-generation retrieval engine designed for auditable, fine-grained multi-hop grounding. BeamRAG v2 introduces four foundational contributions: (1) pre-traversal document-scoped alias-cluster mapping, which indexes multimember alias components as implicit hyper-nodes under canonical representatives for seed deduplication and per-path loop pruning; (2) query-complexity-adaptive beam width allocation 𝐵(𝑞) ∈ [4, 8], scaling search breadth based on syntactic clause density; (3) thresholded continuous power-law hub regularization with explicit leaf dampening, which leaves nodes of degree at most 20 unattenuated by the hub factor while smoothly suppressing high-degree connector hubs; and (4) fine-grained edge-to-chunk syntactic provenance tracking with macro-coherence-gated cross-document bridge traversal. We formulate a standardized 500-query benchmark across four heterogeneous enterprise domains (Computer Science, Agricultural Science, Contractual Legal, and Cross-Domain MIX) operating on an authentic enterprise database (177 documents, 67,772 text chunks, 17,613 resident graph nodes, and 53,628 unique resident edges). We evaluate BeamRAG v2 in a blind, randomized comparative protocol against first-generation BeamRAG v1, baseline BFS, and a deployed LightRAG baseline. BeamRAG v2, BeamRAG v1, and BFS are evaluated on the same enterprise graph and text-chunk collection under identical generator configurations and track-specific prompt templates. BeamRAG v1 and BFS responses are reused from benchmark logs produced under these matched conditions, preserving the original generated outputs without resampling. LightRAG uses its native index constructed from the same source knowledge-graph export. Across two inverted 500-query generation tracks, BeamRAG v2 demonstrates statistically significant superiority over baseline BFS (Track 1: 63.45% decisive win rate, 𝑝Holm = 4.42 × 10−8‡ ; Track 2: 72.13%, 𝑝Holm = 4.31 × 10−20‡ ) and over the evaluated LightRAG deployment (Track 1: 81.21%, 𝑝Holm = 7.50 × 10−45‡ ; Track 2: 89.45%, 𝑝Holm = 2.16 × 10−77‡ ). In direct head-to-head comparison with BeamRAG v1, performance shows no statistically detectable difference in Track 1 (50.23%, 𝑝 = 0.9615ns) while reaching statistically significant superiority in Track 2 (55.29%, 𝑝Holm = 0.0327∗ ). —