Customer support and compliance teams spend hours manually cross-referencing relational order histories, payment records, and complex return policies to resolve refund disputes.
End-to-End Reproducibility: The entire project (IBM Db2 12.1.5 + Streamlit + LangChain + FastEmbed) runs via Docker Compose (docker compose up -d) and includes an automated submission validator and unit test suite covering SQL security guardrails, native Db2 VECTOR_DISTANCE retrieval, and QUERY_AUDIT logging.
Neuro-Symbolic Zero-Hallucination Design: Instead of letting the LLM guess date math or policy thresholds, Python deterministically computes exact elapsed days (days_elapsed) from Db2 relational rows and pairs them with policies retrieved via Db2's native VECTOR(384) COSINE distance before synthesizing the final verdict via LangChain LCEL.
Verified Test Cases Included: Tested and verified across standard approvals (Order 101: 4 days <= 7-day damaged window -> APPROVED), VIP loyalty exceptions (Order 102: Gold tier, 39 days <= 60-day VIP window -> APPROVED), out-of-window denials (Order 103: 47 days > 30-day window -> DENIED), multi-table aggregations (LISTAGG), and adversarial SQL injection/DML (DELETE FROM ORDERS -> REJECTED & logged in QUERY_AUDIT).
All dataset records are 100% synthetic and licensed under MIT.
