Mustafa3946/rca-retrieval-bench: Artifact v1.0.0 — Quality–Latency Benchmarking of RCA Template Retrieval (BGL)

Paper This release is the archived research artifact for "Quality–Latency Benchmarking of RCA Template Retrieval for GenAI-Driven Network Operations". What this release contains Reproducible benchmarking harness for RCA template retrieval with effectiveness (MRR@10, nDCG@10, MAP, Recall@10) and QoE latency percentiles (P50/P95/P99) reported jointly. Implementations/configs for BM25, Dense (SBERT + FAISS ANN), Hybrid (RRF), and LAR-Rerank (semantic + temporal/frequency/topology priors) as described in the paper. Headline results on BGL (40 incidents) Dense is the best quality–latency operating point (MRR@10=0.690, nDCG@10=0.760, MAP=0.128; P95=772 ms). Contextual re-ranking shows no measurable effectiveness gain on this benchmark and worsens tail latency (P95≈6.9 s). Data & licensing Processed artefacts, query sets, relevance judgements, and analysis scripts are provided under CC BY 4.0 in this repository; original BGL logs are available via LogHub. How to reproduce (high-level) Run the pipeline to build indexes/embeddings offline, then execute evaluation to generate ranking metrics and latency percentiles (see repo README/scripts). Zenodo If Zenodo is enabled for this repo, publishing this GitHub Release will mint a DOI for this exact snapshot (recommended for citation).