Home/ News/ v5.2.3 — Vector backend gate + .env.example
Hotfix 2026-05-27 ~3 min read

v5.2.3 — Issue #162.
Semantic search finally on by default.

Six minutes after v5.2.2 cut, v5.2.3 followed: a hotfix that closes issue #162. The costs.vector_adapter health probe was hardcoded to lancedb while production has defaulted to qdrant since v3, so every cost upsert and reindex silently no-op'd — /match-elements read “AI Semantic Search is offline” even when Qdrant itself was healthy. Backend probe now matches settings.vector_backend, and backend/.env.example got a full rewrite while we were there.

v5.2.3

This is a small surface area — two files in the costs module and a rewritten .env.example — but the user-visible blast radius was significant. Anybody running a default install (Qdrant since v3, never LanceDB) had no working semantic search: BOQ position vectors weren’t indexed, similarity-search lookups in the /match-elements wizard fell through to the keyword path, and the AI-classification hint chips never lit up. The probe failed in a way that logged a warning and continued; the upsert path then no-op'd on every write. It was a silent miss in the dependency-detection layer.

The two changes

Vector-backend probe matches the configured backend

_vector_available() in costs/vector_adapter.py now dispatches on settings.vector_backend: it probes qdrant_client when the backend is qdrant (the default since v3 — config.py:237) and only falls back to probing lancedb when the backend is explicitly lancedb. Symptom was unmistakable once you knew the cause: /match-elements showed “AI Semantic Search is offline” even though Qdrant was healthy and reporting zero indexed vectors. With the probe corrected, the next cost upsert seeds the collection and the AI path lights back up.

backend/.env.example rewritten with structured sections

Restructured into labelled sections (Database, Vector backend, Demo creds, Optional AI keys, etc.), fixed the stale VECTOR_BACKEND=lancedb comment that was lying about the default, made it explicit that DATABASE_URL must be uncommented for PostgreSQL to take effect (otherwise the SQLite fallback wins silently), and exposed the DEMO_*_PASSWORD vars so a fresh install can pin the canonical demo credentials at first boot instead of inheriting an auto-generated token nobody can recover.

Why this matters

Issue #162 is the kind of bug that punishes people for trying the right setup first. The whole point of switching the production default from LanceDB to Qdrant in v3 was to give users a real vector store with persistence, replication, and a UI — the price for that was a separate service to run. Anyone who set up Qdrant exactly as documented, then opened /match-elements, was told their semantic search was offline. The probe was looking for the wrong library. The fix is a five-line dispatch but it restores the headline AI feature on every default install. The .env.example rewrite is in the same release because the two surfaces share the same root cause — configuration that drifted from its defaults without anyone noticing.

Upgrade

pip install --upgrade openconstructionerp

No Alembic migration, no DB change — the Qdrant collection itself is created lazily on first upsert, so after the upgrade the next cost-row save() will populate the index and AI semantic search returns to service on its own. No manual reindex needed. If you customised backend/.env, diff it against the rewritten backend/.env.example and copy in the new DEMO_*_PASSWORD section if you want pinned demo credentials.

Try v5.2.3 today.

Live demo in your browser, or self-host in five minutes. AGPL-3.0, no signup required.

Try the demo View on GitHub