Inspectable Python/FastAPI and TypeScript reference system bridging structured relational KPIs (SQLite) and unstructured operational SOPs with pre-query data-freshness gates, AST-validated read-only semantic SQL templates, deterministic numerical grounding verification, and SHA-256 digest-bound HITL workflow approvals.
artifacts/copilot_benchmark.json)Evaluated across 14 relational SQL KPI queries, 10 unstructured SOP policy queries, 10 multi-source SQL+RAG compliance synthesis queries, and 6 stale-partition data-freshness gate queries.
| Architecture | SQL Exec Accuracy | Numerical Parity Rate | Freshness Gate Precision | Routing Accuracy | Mean Latency (ms) | Est. Tokens / Query | Est. Cost / 1k Queries |
|---|
Select an operational or financial stakeholder question below to inspect its pre-query freshness check, AST-validated read-only SQL execution plan, retrieved SOP passages, and token-by-token numerical parity verification.
DQA-FRESH-24H / DQA-NULL-5PCT)
Before executing any SQL query or hybrid synthesis, src/copilot/data_quality.py inspects every table's schema columns, row-count bounds, column null ratios, and refresh timestamp age. Switch evaluation clock to observe partition blocking.
HITL-STATE-4 / HITL-DRIFT-SHA)
Step through the transactional SQLite state machine (DRAFT -> PENDING_APPROVAL -> APPROVED / REJECTED) bound to
SHA256(normalized_question || sql_snapshot_sha256 || sorted_passage_sha256s || answer_sha256), and simulate an upstream SQL table refresh that automatically invalidates a prior approval.