Release Notes

Cross-conversation recall. Chat can search your earlier conversations and reopen one to read what was actually said, backed by a hybrid semantic and keyword index — so a follow-up that assumes shared context resolves against your own history instead of guessing.

Company provisioning wizard. Administrators create an organization in five steps — company, workspace, features, employees, review — resolving to a single write. Nothing is written before the review step, so an abandoned wizard leaves no partial organization behind.

Portfolio site plans. A property page can render a bird's-eye site plan: building footprints and roads from OpenStreetMap, tenant positions from Overture Maps, tenants labelled with square footage. The rent roll stays authoritative for who a tenant is and how much space they hold — map data only supplies where — and a tenant that cannot be matched confidently is listed rather than pinned to a building.

Tenant intelligence views. Tenant Watch (distress and closure events carrying an evidence tier and source citation), Renewal Radar (expirations across roughly a five- to eighteen-month horizon, beyond the lease-expiry watchdog's window), and tenant sales performance against a center baseline each have a view and an API. The background scans that populate them are not yet dispatched on a schedule.

Work-order helpfulness. A tenant rating page reachable from a signed link without signing in, photo strips on work-order detail, a reports dashboard with status / priority / category / property breakdowns and satisfaction figures, and a work-order attention section in the morning briefing.

Deeper lease records. An admin-dispatched batch pass reads each lease's indexed source document and fills the deeper lease fields — renewal notice deadline, renewal options, rent schedule and escalations, CAM structure, co-tenancy, exclusives, security deposit — writing only where a field was empty unless a re-read is explicitly forced, and recording the source document for provenance.

Google calendar and inbox freshness. The Google-only calendar view reads live from the Google Calendar API, matching the Microsoft experience, and falls back to the synced cache on any failure. The Google inbox refreshes on open rather than waiting for the background sweep; a fully live Gmail list stays quota-prohibitive. Applies when a Google account is connected and the corresponding scopes are granted.

Document rows are organization-scoped. The documents table gained an organization column and strict row-level security. Its tenant key previously lived in a polymorphic user field, which left document names and source paths without a database-level isolation backstop.

Work orders in chat. Create and track work orders conversationally — permission-gated (workorders:write, checked at staging and again at confirmation), confirm-first, with status history and selected audit events.

Google Drive connector. Live document retrieval from Google Drive joins Egnyte and direct upload, with cross-organization isolation verified in CI.

Document tools. PDF redaction and structural page edits (delete, rotate), spreadsheet generation, and chart reading directly in chat.

Proactive agents with honest health reporting. Morning briefing, email triage, and lease-expiry agents run org-scoped scans; every run records a health status (healthy / degraded / unavailable / disabled) to the audit trail.

Custom roles. Organization-defined roles join the base viewer / member / admin set, assignable from the admin console and enforced per endpoint.

In-app notifications. A notifications panel on the masthead with typed, deep-linked alerts for email, chat, work-order, calendar, and agent events.

GDPR erasure, end to end. DSAR export and erasure endpoints are live; audit logs are pseudonymised (never deleted) via a database-enforced carve-out in the append-only trigger, covered by real-Postgres tests.

Tenant portal hardening. The tenant self-service portal (photo-attached requests via expiring magic links) is in a dedicated security-hardening pass — stored-XSS, quota, and token-handling fixes landed — ahead of general enablement.

Verified backups. Backup and restore-drill tooling was exercised against an isolated historical database snapshot using the then-current plaintext backup path. That result is not evidence for the current encrypted/checksummed backup path, app-role access, full-service recovery, provider backup settings, or the present schema head; those require a current isolated restore drill.

Faster, cleaner chat. Instant new-chat rendering with provisional titles, adaptive reasoning shown live, accurate reasoning labels, and a smoother signed-in loading experience.

Automatic model tiering. Queries are routed automatically across a fast tier (quick lookups), a standard tier (standard analysis), and a deep tier (deep reasoning) based on complexity — no manual switching.

Cost tracking. Supported user inference paths schedule token/cost usage records when provider usage metadata is available. Provider health checks, administrative calls, and failed/best-effort logging paths are not represented as universally recorded.

Cross-organization admin isolation. Admin users are now restricted to managing users within their own organization. Platform-wide operations require super-admin privileges.

Open redirect prevention. OAuth flows now validate redirect URLs before navigation.

Financial anomaly detection. A March preview explored NOI drift, CapEx irregularities, and rent anomalies. The corresponding automated anomaly agents are now retired/dormant and are not represented as active monitoring.

Rent reconciliation dashboard. A preview compared lease data with payment inputs. This is not a currently mounted continuous-reconciliation dashboard.

API Reference documentation. Complete reference for 50+ API endpoints with request/response examples. View API docs.

Help Center expanded to 22 FAQs. New answers covering AI agents, data export, dashboard, critical dates, conversation sharing, and templates.

Morning briefing. Time-aware greetings, portfolio health score, and quick-action buttons to jump directly into chat with prefilled queries.

AI answer accuracy. Reduced false hallucination warnings on calculated financial figures — derived amounts (occupancy rates, pro rata calculations, weighted averages) are now correctly recognized.

Search precision. Tenant name matching now uses word boundaries to prevent false positives (e.g., "Ross" no longer matches "across").

Historical critical-date creation/export and agent-run endpoints received runtime fixes in that release. The standalone critical-date router is no longer mounted.

A historical Dashboard urgency badge consumed critical-date counts. The legacy Dashboard UI and standalone critical-date feature were later retired; current dashboard summaries do not provide those live counts.

Multi-tenant organizations. B2B customer isolation with organization-scoped data, user management, and role-based access. Admin endpoints for creating and managing organizations.

Row-Level Security (RLS). PostgreSQL row-level policies enforce data isolation at the database level — defense-in-depth against cross-organization data leaks.

Subscription tiers. Free trial (100 queries/month), Professional (2,000/month), and Enterprise (unlimited). Usage tracking and quota enforcement.

Critical date tracking. Historical feature: a standalone critical-date tracker covered expirations, renewal deadlines, and option dates. Its router is no longer mounted. The current lease-expiry agent only checks recorded lease end dates inside 30 days when enabled.

Query templates. Historical feature: 19 CRE query templates were introduced, but the dedicated template router and custom-template panel are not currently mounted. Current slash entries are disclosed as ordinary one-turn prompt shortcuts.

Scheduled reports. The mounted API can save supported report definitions and execute them on demand. No current background dispatcher automatically runs due definitions or delivers recurring report email, so the March recurring-delivery description is historical.

Portfolio analytics dashboard. Selected authenticated dashboard-analysis APIs remain mounted, including tenant risk, rollover, concentration, escalation, and data-quality views. The legacy Dashboard UI is retired, and critical-date portions may be empty because that feature is not mounted.

Structured logging. JSON logs now include organization ID, correlation IDs propagate to background tasks, and slow database queries (>500ms) are flagged.

Scroll-to-bottom button in chat now appears correctly when scrolling up during a conversation.

Streaming errors no longer replace partial AI responses — errors append below the content you already received.

Escape key now properly closes the evidence panel.

Long conversation titles no longer overflow the chat header.

Search-first document retrieval. Redesigned RAG pipeline: search the connected document library first, cache document IDs, then fetch full text on demand — reducing full-document reads on lookup-shaped questions. Retrieval has since become provider-agnostic, and it now combines this search-first path with a semantic (embedding) index rather than replacing it.

Real-time thinking narration. See what the AI is doing as it works — each search and document read shows a descriptive status update with discovery summaries.

Efficiency. Retrieval can inspect search snippets before reading complete documents. Tool-call count and latency vary with the request, provider, available evidence, and pipeline state; no universal 2–5-call benchmark is claimed.

LOI Generator agent. Historical preview of a multi-step LOI workflow. That agent is now archived and is not available as a deterministic production pipeline.

Clause deviation review. Historical LOI comparison preview; not a currently mounted automated workflow.

Large context windows. Handle longer documents and multi-document analysis in a single query using our AI provider's large-context models.

Reasoning orchestration. A March preview classified queries as simple, moderate, or complex and routed them to different reasoning strategies. That standalone reasoning package is no longer wired into any production path; complexity-based routing today is handled by the automatic fast / standard / deep model tiering described in the March 22 entry.