v1.0.0 - FAQ-First RAG Retrieval

Highlights

FAQ-First Priority Retrieval — the flagship retrieval upgrade of this release:

• FAQ intent classifier routes questions to a priority path: 98.3% R@1 within FAQ domain (multi-gold dataset, 120 cases, 6 categories all ≥ 91.67%)
• Raw hybrid retrieval baseline: 47.5% R@1 / 62.5% R@5 (single-gold)
• P95 retrieval latency: 181 ms

What's included

• Adaptive RAG pipeline: dense vector (Milvus) + BM25 hybrid with RRF fusion, CrossEncoder reranking (candidate pool 200)
• Semantic cache to cut repeat LLM calls
• Multi-model routing and cost governance
• Observability: trace retrieval decisions end-to-end
• QA classifier + FAQ data cleaning pipeline

Quality

• 372 tests, JaCoCo coverage gate 64%
• CI: CodeQL + Maven build
• Full evaluation evidence: docs/RAG_EVALUATION_SUMMARY_2026-08-27.md
• Reproducible evaluation datasets (silver-v2) and snapshots included

License

Apache-2.0