Trace

Trace turns engineering drawings into a queryable evidence graph: a PDF is ingested, structured into objects, relations, and source page regions, and every chat answer is bound to evidenceIds and highlightIds that link back to the exact region it came from — when evidence is insufficient the system says so rather than guessing. Production is backed by an AI cost-control layer (unified router, per-request and daily budgets, circuit breakers, audited usage events) and multi-tenant infrastructure (RBAC, credit ledger, cross-tenant isolation, deletion pipelines). A research track (Pipeline v3) ran a 30-drawing bootstrap benchmark — the gold annotations are still pending human double-review, not yet frozen — and the quality gate correctly returned rollback_required on the extraction scores, so production stays on the v2 pipeline until accuracy clears the bar: a fail-closed rollout decision documented rather than hidden.