Graph-Augmented Retrieval and Supersession-Based Chunk Promotion for Technical Infrastructure RAG: A Phase 4 Graph Study

The prior companion study [2] to the base study [1] named two future-work extensions: completing live-state ingestion for the storage domain, and graph-augmented retrieval targeting compositional questions and a recurring entity-fabrication pattern. This study implements and empirically tests both. A Neo4j property graph (19 node types, 20 relationship types) was built over the original 451-chunk static corpus from a direct audit of the source documents, and graph facts were surfaced as ordinary text chunks through the unchanged top-k vector pipeline — deliberately introducing no new retrieval mechanism. Benchmarked on the original 30-question set, the graph-derived chunks produced a genuine embedding-backend × model crossover rather than a uniform gain: one model improved on the dense backend and declined on the hybrid backend while the other did the reverse (+2.0 / −3.3 / −5.3 / +8.0 points across the four configurations), with the nearly flat +0.4-point average actively concealing the interaction. Live-state ingestion was then completed across six infrastructure domains, including the storage domain the prior study identified as its coverage gap. direct device verification resolved the base study’s longest-standing corpus ambiguity — both candidate parallel filesystems are real, live, and running on different hardware tiers, making the ambiguity a factual underspecification rather than an irreducible retrieval limit — and closed the monitoring-tool corpus gap; it also corrected a corpus-wide design figure, with two independent live methods converging on 29 connected GPU nodes / 232 GPUs against the documented design inventory of 31 / 248. Three retrieval strategies — plain similarity, the prior 3×-candidate-pool authority weighting, and a newly designed supersession-judge mechanism that promotes a higher-authority chunk only after a small local language model confirms the two chunks describe the identical real-world entity — were evaluated across the full 45-question benchmark and all four model/embedding configurations (540 manually scored answers). Plain retrieval won or tied in three of four configurations. Diagnostic tracing identified pool expansion before re-ranking as the structural failure mode of tier-multiplier weighting, mechanistically explaining the prior study’s corroborated authority-weighting failure. The supersession judge consistently outperformed blanket weighting but did not demonstrate a clear win over plain retrieval, and one judge promotion was confirmed as a false positive by direct chunk inspection. Judge-model selection itself reversed under a deliberately hardened evaluation set, with the best small model on an easy 10-pair test falling to worst-of-four once structural near-twins were added. Finally, the pipeline’s generation stage was confirmed at source to run at a non-zero default sampling temperature, identifying the concrete mechanism behind the prior study’s measured ±6-point single-trial variance band