Lexical Fusion and Cross-Encoder Reranking for Technical Infrastructure RAG: A Phase 5–6 Companion Study

The second companion study (DOI: 10.5281/zenodo.21636834) left the corpus frozen and named lexical/exact-identifier fusion, motivated directly by a confirmed hexadecimal-versus-decimal indexing mismatch, as the next retrieval-side lever. This study reports two sequential phases delivering on that plan. Phase 5 fused BM25 lexical retrieval with the project's nomic-embed-text dense backend via Reciprocal Rank Fusion across all three retrieval strategies (plain, authority-weighted, supersession-judge) and both language models, on the fixed 45-question benchmark. BM25 fusion's effect proved model-dependent rather than strategy-dependent: it improved falcon3:7b in every strategy tested (+1.8, +7.6, +2.7 points) while helping llama3.1:8b only under weighted (+6.2), with plain and supersession both regressing (−4.0, −4.9). The gain for falcon3:7b traced almost entirely to the live/graph-derived question set rather than the original benchmark, consistent with BM25's exact-token strength. Cumulative evidence — nomic+BM25 now matching or exceeding BGE-M3's hybrid retrieval on the two exact-identifier question types it was originally retained for, without inheriting BGE-M3's independent, unrelated failure modes — closed a three-phase-old open question and confirmed BGE-M3's removal from the project. Phase 6 then tested a cross-encoder reranker (BAAI/bge-reranker-v2-m3) on top of Phase 5's fused retrieval, using a mechanism-first methodology: before any reranker code was written for the authority-weighted strategy, its own diagnosed failure mode was directly tested against real reranker scores on all eight known cut cases and found already resolved in seven of eight by reranker-primary-sort alone — evidence that no reranker mechanism was needed there, and none was added. Where the reranker was applied (plain and supersession), it produced individually verified correct-answer wins — including the project's first-ever exact match to a documented ground-truth figure — but a net-negative result overall: falcon3:7b regressed in all three strategies (−6.6, −3.5, −8.4 points) and llama3.1:8b was roughly flat (+0.4, +0.9, −1.8). The single best result across both phases — llama3.1:8b at 80.0% combined — was achieved with the reranker completely uninvolved, on Phase 5's unchanged authority-weighted code. One question-specific, non-infrastructure model failure was confirmed by direct re-run, and two previously reliable facts were found wrong for both models specifically under the reranker-modified supersession strategy. The reranker is retained as a documented, evidence-backed but currently shelved lever; Phase 7 proceeds on Phase 5's validated BM25-fused baselines.