Back to AI Advisor

Release notes

AI Advisor 1.7.3  ·  28 Sep 2026

New features
  • Ask about equipment on your other controllers. Previously the assistant answered "there are no such points" — wrongly, because the equipment was there. Nothing to set up: if the controller is on your Niagara Network and connected, it uses the login you already have.
  • Newer AI models can now use the station. The latest OpenAI reasoning models can look things up while working through a question, so answers are better informed and take fewer steps.
  • Full screen per user. A new option in each user's web profile opens the AI Advisor without Niagara's top bar and side panels, so the chat and the view get the whole window.
Improvements
  • Answers about other controllers are complete — not just the top level, with the rest silently missing.
  • The assistant says when it has not looked everywhere, instead of returning "no results" and letting it read as "this does not exist".
  • Equipment on another controller is marked as such when the assistant lists your site.
  • Attaching a document no longer slows the conversation — a PDF is read once and summarised, not carried in full behind every later message.
  • Trends return in full rather than stopping partway.
  • No unasked-for navigation — the assistant opens a view only when you ask.
  • One chat with Reflow Redirect. Sites that show Niagara views inside Reflow pages now get a single chat window at the top of the page, instead of one per frame.
  • Faster searches on large driver networks — searching for other controllers no longer walks through every device's points and histories.
Bug fixes
  • Fixed reading the wrong controller when two pieces of equipment shared a name.
  • Fixed equipment filed in a folder being reported as "not on your network".
  • Fixed moving up a level from another controller landing back on the local one.
  • Fixed search results for other controllers pointing at the wrong place.
  • Fixed Workbench occasionally hanging on start-up.
  • Fixed mistyped commands being ignored instead of reported.
  • Fixed an unclear failure when a controller on the Niagara Network has no connection; the assistant now says which controller it could not reach.
  • Fixed "Report problem" saying "Nothing to send yet" — it now sends the conversation you are in.
  • Fixed conversation logs missing the user's name.
Upgrading
  • Licensing now ships as a single module, bdrtLicense-rt.jar. Delete bdrtLicense-wb.jar from the Niagara modules folder when you copy in the new jars.

AI Advisor 1.5.2  ·  14 Aug 2026

Covers all changes since 1.0.30.

New features
  • Tell us when an answer is wrong — a Report problem button under every answer sends that one conversation to Baudrate with your note. Separately, the station can share conversation logs with us daily to improve the product — off by default, asked once in the chat, and changed any time under Configuration > Log Sharing.
  • Licensing is production ready — the licence now enforces what your subscription covers, and the subscription itself is managed from Workbench (Tools > Manage Baudrate Subscription).
  • System prompt split into editable parts — the single ^ai/prompt.md is replaced by ^ai/main.md (the skeleton), ^ai/user-preferences.md (the assistant's persona and your house rules), and ^ai/station.md (written by /init). user-preferences.md is the one to edit — e.g. "Always include units when reporting values". Edits to any of the three take effect on the next message; no engine or station restart. See the Update Guide for what to do with the old prompt.md.
  • Answer length control — new Configuration > Verbosity property: brief, normal (default), or detailed.
  • Skills — the assistant now loads instruction files on demand through a new get_skill tool instead of carrying everything in the system prompt, starting with Haystack tag-based queries. /init also records the tags your station actually uses, so tag queries work against your naming rather than a generic assumption.
Improvements
  • Deeper, faster station access — the assistant now explores the station through one flexible read-only interface rather than a handful of fixed lookups. It can combine and filter several reads in a single step, drill into what it just found without going back for another round, and reach corners of the station the fixed lookups never exposed — including the tags actually applied on your station rather than what a dictionary says should exist. The practical effect is fewer AI calls per question, quicker answers, and much less that the assistant simply cannot see. Anything that would modify the station is still refused.
  • Current values everywhere, virtuals included — 1.0.30 made BQL queries wait for fresh readings; that now covers every way the assistant reads the station, including listings, single point reads and the filters inside a query, so a where-clause is evaluated against current values rather than whatever was last cached. Virtual components are brought online before they are read, so points under virtuals return real data instead of blanks. Large reads refresh up to a capped number of points and say so, rather than stalling.
  • Voice playback controls — text-to-speech playback can be started and stopped properly from the chat (ESC stops reading).
  • Readable tables in chat — markdown tables returned by the assistant are now styled to match the station page instead of rendering unformatted.
  • Cleaner first run — the /init prompt in the chat only appears until the station profile exists, so it stops nagging once the station is set up.
  • Documentation — the user guide gains chapters on log sharing, licence limits and Azure OpenAI, and an explicit table of which tool edits need a station restart (adding, deleting, renaming, enabling/disabling a tool, or editing its Description or Schema needs one; Query edits apply immediately). Azure OpenAI is configured in ^ai/application.yml — endpoint, deployment name and API version — not from the Provider dropdown as previously stated.
Bug fixes
  • Fixed a crash that ended the chat turn when Gemini returned an unrecognised finish reason.
  • Fixed BQL/NEQL parsing of quotes and ORDs containing pipes, which could make a query fail or target the wrong component.
  • Fixed the screenshot tool firing before the view had finished rendering, and its argument parsing.
  • Fixed a duplicate chat widget and a lingering floating button when the station page contained sub-frames.
  • Fixed a chat history freeze and message wrapping in narrow windows.
  • Fixed a Fox session timeout NullPointerException and web-profile scope injection.
  • Stopped the /init agent emitting NEQL queries during the station survey.

AI Advisor 1.0.30

Covers all changes since 1.0.22.

New features
  • Azure OpenAI provider — Configuration > Provider now offers OpenAI, Anthropic, Google, Ollama, and Azure OpenAI. Pick the provider and model on the property sheet; no application.yml editing required.
  • Always-fresh query data (live subscription) — BQL queries used by both the chat tools and the dashboard pipelines now detect stale proxy-point values, subscribe to those points to force the driver to re-poll, wait for fresh readings, then return. Previously a query could hand the assistant stale values, because resolving BQL alone does not trigger station subscription. The assistant and dashboards now act on current data.
  • Node Manager dashboard — seven example pipelines — the dashboard now ships with seven worked examples instead of three. ForgottenOverrides, AlarmPatterns and NetworkHealth work on any station with no configuration.
  • Four per-equipment and whole-building pipelines — Air Handling Unit AHU1 [CONF], Electric Meter QUPSB [CONF], Chiller CH1 [CONF] and Overall System Summary [CONF], driven by domain-expert prompts (HVAC, electrical, refrigeration, building supervisor). Pipelines whose name ends in [CONF] must be pointed at your own equipment before use; each one's Description block lists exactly which query ORD and alarm-source filter to change. Duplicate a [CONF] pipeline once per piece of equipment you want to monitor.
Improvements
  • Rate-limit aware dashboard — when an AI node hits a provider rate limit, the dashboard queue pauses instead of hammering the API, then resumes, avoiding cascade failures and wasted calls.
  • Broader JDK auto-detection — the AI Engine launcher now finds Adoptium, Microsoft, Semeru, and Azul/Zulu JDKs (in addition to the previous set), so the engine starts reliably across more host configurations.
  • Licensing in the palette — a LicenseService now ships as part of the AI Advisor palette, so licensing drops in with the rest of the service.
  • User guide overhaul — restructured sections, a single per-model table, an Ollama setup chapter, an AI Pipelines (Node Manager) chapter, and a proper PDF (title page and table of contents).
Bug fixes
  • Fixed a station NullPointerException hang.
  • Fixed a license-fetch deadlock.
  • Fixed the thinking indicator not clearing during tool calls.
  • Fixed Hx profile CSS leaking into the host page.
  • Various fixes — internal reliability work (Fox connection-registry and login-pipeline refactors, an RPC bridge for integrations) plus assorted stability and correctness fixes.

AI Advisor 1.0.22

New features
  • Simplified upgrade process — from 1.0.22 onward, every upgrade is the same three steps: (1) replace the JARs, (2) restart station and Workbench, (3) right-click AiAdvisorService > Reset Defaults. No more property-type migrations or manual prompt copy-paste.
  • /init station profiler — run /init in chat once after install (or after major navigation-file changes). The init agent surveys your station and writes a profile to ^ai/station.md that is auto-injected into every assistant request. Makes answers dramatically more accurate, faster, and cheaper.
  • Node Manager dashboard — proactive station monitoring with AI pipelines. Pipelines run on schedule and surface as RAG (Red/Amber/Green) status cards on a single dashboard view. Ships with three example pipelines (BooleanOverrides, AlarmPatterns, NetworkHealth); supports user-defined pipelines for any use case.
Improvements
  • Provider and Model moved to the property sheet — previously edited in ^ai/application.yml; now live on Configuration > Provider (OpenAI / Anthropic / Google / Ollama — new) and Configuration > Model.
  • Navigation file property moved — was Tools > nav_tree > Query; is now Configuration > Navigation File. The nav_tree tool itself was removed and replaced by a smarter navigation tool that handles large stations.
  • Tool results paginate automatically — fixes truncation and timeout problems on large stations.
  • Reset Defaults moved and renamed — was Configuration > forceDefaultConfig; is now AiAdvisorService > Reset Defaults. One click restores all ^ai/* config files, repopulates Tools, and restarts the AI Engine.
  • Better debugging via conversation logs — logs now under ^ai/conversations/, tagged by agent, with tool-call results visible in summary mode. Configurable in application.yml (off / summary / full).
  • System prompt moved to a file — edit ^ai/prompt.md in any text editor instead of a giant property-sheet field.
  • Longer chat memory — assistant remembers more of the conversation before older messages drop off.
Bug fixes
  • Various stability and correctness fixes across navigation, query handling, and chat sessions.

Running an older build? Copy the whole set of module JARs from the newest release into your Niagara modules folder, then restart the station and Workbench. Questions: info@baudrate.io.