v0.3.2 — Qdrant & Retrieval Observability

[Rcloudso]

Product impact

Adds an optional Qdrant production vector backend while keeping SQLite authoritative and memory as the fresh-clone default. Administrators can build, quality-gate, atomically activate, inspect, and roll back versioned indexes from a bilingual Retrieval Operations page.

Implementation scope

• asynchronous memory/Qdrant VectorStore with safe payloads and SQLite hydration
• recoverable index jobs, Quality Lab backend targets, alias activation and rollback
• eight-stage retrieval traces with bounded retention and authorized hydration
• admin APIs and responsive Retrieval Operations UI
• pinned Qdrant 1.18.2 Compose/CI integration, bilingual docs, screenshots and demo

Verification

• full regression and TypeScript checks passed
• FAQ/document/mixed/Quality/OCR/triage evaluations passed
• OCR worker tests: 6/6
• Playwright API/Web: 48/48
• production build, JSON/YAML parsing and git diff check passed

Remaining risks

Docker is unavailable in the local verification environment. The dedicated GitHub Actions Qdrant service job is required before merge. Old collections are retained intentionally; clustering, snapshots, automatic cleanup, runtime provider switching, traffic shadowing and Agentic Retrieval remain out of scope.