Adaptive Retrieval-Augmented Generation Pipelines for Dynamic Document Landscapes in the Insurance Industry: A Comparative Study of Retrieval Strategies for Compliance-Aware Response Generation

Insurance organizations rely on policy contracts, compliance guidance, product manuals, and regulatory filings that change frequently. Many Retrieval-Augmented Generation (RAG) systems remain built around a mostly static index, so retrieval can return passages that look relevant while reflecting an older version of the source material. In a regulated domain, such as insurance, this exposes users to outdated coverage language, exclusions, and regulatory obligations. This study presents an adaptive RAG pipeline that combines drift detection, version-aware query routing, hybrid retrieval (dense, semantic, and Best Matching 25), and compliance validation in a LangGraph workflow. The pipeline was compared against three v0-only static baselines (a Facebook artificial intelligence similarity search-based chatbot, a hybrid Azure AI Search pipeline, and a multi-modal multi-agent RAG text proxy) and two fair-comparison baselines: Static-Reindexed, in which the full corpus is re-embedded after every amendment, and Partial-Update, in which 30% or 60% of stale-flagged chunks are refreshed per cycle. Evaluation used a versioned synthetic corpus of eight insurance policy forms with 30 amended provisions across four amendment types (numeric replacements, scope shifts, definitional changes, and invariant controls) and five US jurisdictions, together with 106 version-sensitive queries (85 version-divergent, 21 invariant). Metrics separated stale chunk rate (SCR), semantic hit rate (SHR), and version-aware retrieval precision (VRP) from a corrected version-sensitive answer metric, VA-RAcc-strict, reported on the divergent subset. All comparisons report paired-bootstrap 95% confidence intervals, two-sided p-values, and Cohen's h effect sizes. The corpus remains synthetic and the contrasts sharp; the headline scores therefore reflect controlled-protocol evidence rather than production performance. A metric-coherence issue in the original VA-RAcc, under which v0-only baselines could score around 73% on version-invariant queries despite zero version-aware precision, was diagnosed and replaced by VA-RAcc-strict. Under the corrected metric, the v0-only baselines correctly score 0.000 and the proposed system scores 0.965 on the divergent subset (paired bootstrap, p < 0.001, Cohen's h = 2.80). The proposed system is statistically equivalent to Static-Reindexed (p = 1.00, h = 0) and strictly outperforms both Partial-Update configurations (30%: VRP difference = 0.736, h = 1.79; 60%: 0.245, h = 0.76; both p < 0.001). The architectural contribution is therefore the drift-aware trigger that determines when re-indexing is required, rather than retrieval gains over an operator with perfect amendment knowledge. A standalone drift-detector evaluation, a component ablation, a drift-threshold sensitivity sweep, and end-to-end latency (P50, P95, P99) per system are also reported. The results do not settle the broader production question. They indicate that document freshness should be handled inside retrieval design through a drift-aware routing trigger, rather than left to manual re-indexing cycles.