A chat assistant in the bottom-right corner of every page, backed by a language model you run — Ollama, LM Studio, vLLM, the llama.cpp server, LocalAI or Open WebUI — or an Azure AI Foundry GPT deployment. Ask it about your devices, alerts, networks and events, have it build a report you can download, ask it to help correlate an outage, or ask how something in Polaris works.

It sees only what you can see. Every lookup the assistant makes runs with your permissions — the assistant can never show you more than your role already can. It can only look things up: it never changes, acknowledges, pushes or deletes anything. See rule 95.

---

1. Integrations → + Add Integration → AI Assistant.
2. Fill in the model server:

Field
Assistant name	what people see in the chat window — its title, the button's tooltip and the greeting (blank = "Assistant"). The model is told to answer to it
Provider	Local / OpenAI-compatible (the default) for a server you run, or Azure AI Foundry — see Azure AI Foundry below for that form
Host / IP, Port	where the model server listens — Ollama's default port is 11434
API path	/v1 for Ollama, LM Studio, vLLM and llama.cpp; /api for Open WebUI
Use HTTPS / Verify TLS	for a server behind TLS; untick Verify for a self-signed lab certificate
API key	only if the server requires one; stored encrypted
Model	pick from the server's own list — Load models (or Test Connection) fills it. Each model is marked ✓ tool calling, ✗ no tool calling or tool calling unverified. Leave it on Auto to use the server's first tool-calling model
Allow loopback	tick only when the model server runs on the Polaris host itself (localhost). In a container, use the host's LAN address instead

3. Test Connection, then Create.

Pick a model that supports tool calling — for example Qwen 2.5 / 3, Llama 3.1 or newer, or Mistral Small. A model without it can chat, but cannot look anything up. Larger models follow instructions and choose lookups better; on modest hardware a 7–14B model is a reasonable start.

How Polaris knows which models call tools. The model list itself never says. On Ollama, Polaris reads each model's reported capabilities, so the marks are the server's own answer, and embedding-only models are left out of the list. On other servers every model starts as unverified; Check tool calling asks the selected model, once, to call a dummy tool and records whether it did. That costs one model request, so it runs only when you ask.

Every save checks it for you. After you create or save the integration, Polaris asks the model it will actually use (the one you picked, or the Auto choice) to call a dummy tool, then shows the result in a notification and on the card's Tool Calling row: ✓ Verified, ✗ Not supported — chat only, no lookups, or Could not tell, with the model's name and when it was checked. Changing the provider, host, port, API path or model clears the result until the next save checks again.

Choose Azure AI Foundry as the Provider to use a GPT deployment from an Azure OpenAI or Foundry resource, or a Claude deployment (Claude Haiku, Sonnet or Opus from the Foundry model catalog). The form then shows:

Field
Endpoint	the resource's endpoint from its Keys and Endpoint page, e.g. https://my-resource.openai.azure.com (a …services.ai.azure.com endpoint works too). For Claude, paste the deployment's Target URI. Paste it whole — anything from /openai or /anthropic on is ignored
API shape	for GPT, v1 (the default — /openai/v1, no API version needed) or Deployments for the older dated path, which also asks for an API version such as 2024-10-21. For a Claude deployment choose Claude (Anthropic Messages API)
Authentication	API key — Key 1 or Key 2 from Keys and Endpoint — or Entra ID app: the app registration's Tenant ID, Client ID and Client secret. Give the app's service principal the Cognitive Services OpenAI User role on the resource (for Claude, Foundry User or Cognitive Services User). Keys and secrets are stored encrypted. Token scope is advanced; leave it blank for the default (https://cognitiveservices.azure.com/.default for GPT, https://ai.azure.com/.default for Claude)
Deployment name	the name you gave the deployment in Foundry, not the model's name
Omit temperature	GPT only: tick for reasoning models (o-series, gpt-5), which refuse it. If you forget, Polaris retries once without it. Claude never gets a temperature, so the box is hidden
Context window	under Assistant Behaviour — set it to the deployment's window: 1000000 for Claude Haiku, Sonnet or Opus 5.5. Left at the 8192 default, Polaris trims history and lookups as if the model were small

Test Connection sends one short chat message to the deployment, the only way to prove the endpoint, the credential and the deployment name are all right. The background connection check repeats that every 10 minutes while the integration is enabled, so it costs a few tokens per check. Everything else — lookups as the person asking, read-only, the role and token it creates — is the same as for a local server.

Creating an AI Assistant integration also creates, for the model server's own use:

• a read-only role named llm-<integration name> — it reads inventory, networks, alerts, events and maps, and cannot read credentials, API tokens, users, roles, sign-in settings, automation scripts or server settings;
• an API token bound to that role.

The token is shown once, with the Polaris API base URL. Give both to the model server if you want it to call Polaris itself (an Open WebUI tool, an MCP server, a script — see REST API). The chat in Polaris does not use this token. Use Regenerate Token on the integration card if it leaks; deleting the integration deletes the token and the role.

Because it creates a role and a token, adding an AI Assistant integration needs Read-Write on Roles and on API Tokens as well as on Integrations.

Anyone whose role has AI Assistant at Read (every built-in role does, the read-only role included) — once at least one AI Assistant integration is enabled. To hide it from a role, set AI Assistant to None under Users → Roles.

The same people get a Chat tab in the phone app, in the slot Networks otherwise takes (Networks moves to More). It opens the same conversations as the desktop.

---

Click the round button in the bottom-right corner. The chat window opens:

• Move it by dragging the title bar; resize it from the top-left corner. Double-click the title bar to send it back to the corner, or the top-left corner to restore the default size. Its top edge stops just below the page's top bar, so the title bar and the corner are always in reach; a window too tall for the space is shortened to fit.
• It stays put when you change page, and an answer that is still being written carries on — the next page shows Still answering… and then the answer. Stop is the only thing that stops it.
• Esc or the – button collapses it back to the button. If an answer finishes while it is collapsed, the button shows a dot.
• Enter sends; Shift+Enter starts a new line.
• Answers appear as they are written. While it works you see what it is doing — looked up assets, checked alerts, searched the Polaris help. Stop cuts an answer off; what was written so far is kept.
• Before the first word, the reply shows how long it has been working — Thinking… · 12s. A thinking model (Qwen 3, DeepSeek-R1 and similar) reasons privately before it answers, which can take a minute on modest hardware; while it does, the line reads Reasoning… 2,340 characters · 25s so you can see it is making progress. The reasoning itself is never shown or saved.

Talk to it normally — follow-up questions use the earlier conversation as context. Some things to try:

• What's down right now, and what do those devices hang off?
• What else alerted within an hour of fw-nsh-01 going down last night?
• Which networks are more than 80% full?
• Build a report of every FortiSwitch with its serial and firmware.
• How do I set up a maintenance window that repeats monthly?
• How many devices are down in my region?

Four kinds of question follow a set procedure, so the answer is built the same way every time: why did a device go down (the device, then alerts and events around the time it changed, then the device it hangs off), what changed since a time, is a device healthy, which networks are filling up, and who has an address (look up 153.66.102.165) — which also finds a device that held the address on another interface or in the past, and says when, and a firewall whose IPsec tunnel or VPN peer is at that address. Each answer keeps what the data shows apart from what the assistant infers. A device named in an answer is a link: click it to open its details.

It knows your role and access, and the regions and tags assigned to you (from your role, your account and your sign-in groups) — so "my region" means your regions, and "how do I…" answers say whether your role can do it. That knowledge only shapes the answer: every lookup is still checked against your role.

Decommissioned assets and deprecated (retired) networks are left out of answers, counts and reports unless you ask about them ("list the decommissioned switches"). Searching for one by name still finds it.

When you ask for a report, list or export, the answer carries a report card with a preview and CSV, PDF and Markdown download buttons. Polaris runs the query itself (up to 5,000 rows), so the rows are an extract from the database, never typed by the model. The report is saved with the conversation: reopening it later downloads the same figures you saw then.

Questions about using or configuring Polaris are answered from this wiki — the assistant searches it first and links the page it used. If the help has nothing on it, the assistant says so. Links in an answer can only point to a real page of this wiki or to a page inside Polaris; any other link the model writes is shown as plain text.

---

Type / in the message box to see them. Keep typing to filter; ↑ / ↓ to move, Tab or Enter to pick, Esc to close.

Command	What it does
/clear	Clear this conversation and start over in the same thread
/new	Start a new conversation (this one stays in history)
/resume	Reopen the conversation set aside after a spell without activity
/history	Open your past conversations
/retry	Ask for the last answer again
/report <what>	Build a downloadable report, e.g. /report switches down in the last 24h
/docs <question>	Answer from the Polaris help, e.g. /docs how do maintenance windows work
/export [md|pdf]	Download this conversation, reports included
/rename <title>	Rename this conversation
/delete	Delete this conversation permanently
/model [name]	Show the model in use, or switch to another AI Assistant integration
/memory	See and edit what the assistant remembers about you
/remember <text>	Save something about you for future conversations, e.g. /remember I look after the Nashville sites
/help	List these commands

---

• Conversations are saved and reopen where you left off. The history button (or /history) lists them; rename or delete from there.
• After 30 minutes without activity by default — set it with Start a fresh chat after on the integration; 0 turns it off — (no question, answer or opened conversation in this browser) the window starts a fresh chat the next time it loads or opens, or within a minute on a page left open. Nothing is deleted: /resume reopens the conversation it set aside, and it is still in History.
• Conversations follow you between the desktop and the phone. Opening the chat continues the conversation you used most recently on either device, as long as it was within that same idle time — unless you have used this device more recently since. A chat open on both at once does not update live; reopen it, or open it from History.
• They are private to you — no other user, administrators included, can read them.
• A conversation untouched for 90 days is deleted automatically. Change the period with Keep conversations for on the integration (it applies to every assistant on the install; changing it needs Server Settings → System at Read-Write, and is recorded in the event log).
• Only the most recent turns are sent to the model with each question — at most Messages of history sent (20 by default), and fewer when they would not fit the model's context window — so a very long conversation forgets its beginning; /new starts fresh.
• The event log records that you asked something and which lookups ran — never what you asked or what was answered.

The assistant can remember a few short things about you between conversations — your team, the sites you look after, how you like answers — so you don't have to repeat them.

• It saves what you tell it. Say "I look after the Nashville sites" or "remember I prefer short tables" and it saves a note; a line under the answer says Remembered: …. Say something is wrong or no longer true and it removes it (Forgot: …).
• It only saves what you typed in that message. It cannot save something it looked up, or text from a device name or description — so nothing on the network can plant a note in your memory.
• Never addresses, networks, MAC addresses, links, passwords, keys or tokens — Polaris refuses those, whether the assistant or you try to save them.
• The Memory button in the chat window's title bar (or /memory) lists every note, with who saved it. Add one there or with /remember <text>, remove one, or Forget everything. Up to 25 notes of 200 characters each.
• Remember things about me in that drawer turns it off: the assistant stops reading and saving notes, and the ones you have are kept until you turn it back on or remove them.
• Your memory is private to you — no other user, administrators included, can read it — and is deleted with your account. The event log records that your memory changed, never what it says.
• Notes are background about you, not facts about your network: the assistant still looks up devices, alerts and networks with your permissions every time. Notes count against the model's context window, so a small window leaves a little less room for the conversation.

See rule 95.

Field
Temperature	lower = more factual and repeatable; 0.2 is the default
Lookup rounds per answer	how many rounds of lookups one answer may take (default 6)
Rows per lookup	rows one lookup hands the model (default 200); reports go to 5,000 regardless
Messages of history sent	the most earlier turns sent with each question (default 20); older ones are dropped first when they would not fit the context window
Context window	the model server's context size in tokens (default 8192) — set it to match the server. Polaris sizes each question to fit: the conversation, the lookup results (cut down for a small window) and room for the answer. Ollama uses 4096 unless OLLAMA_CONTEXT_LENGTH raises it, and the assistant's own instructions take about 2,700 of those, so raise it to 8192 or more on the server and here. Test Connection warns below 6000. On Azure AI Foundry use the deployment's window (1000000 for Claude 5.5 models); history stays capped by Messages of history sent and each lookup at 24,000 characters, so a large window does not mean large requests
Response timeout	how long the model may go silent before the answer fails
Start a fresh chat after	how long a conversation can sit idle before the chat window saves it, sets it aside and opens fresh (default 30 minutes; 0 = never). /resume or History brings it back
Extra instructions	added to the assistant's instructions — site naming conventions, who to escalate to
Keep conversations for	how long a conversation nobody has touched is kept (default 90 days). One setting for every assistant on the install

Symptom	Likely cause
No button	no enabled AI Assistant integration, or your role has AI Assistant set to None
"Connected, but model … is not on the server"	the model was removed from the server — pick another with Load models, or set it to Auto
Loopback refused	tick Allow loopback, or use the host's LAN address
It chats but never looks anything up	the model does not support tool calling — the card's Tool Calling row says ✗; choose one that does
It forgets the task, or reasons for minutes and answers off-topic	the model server's context window is too small and is silently cutting the start of the question — raise OLLAMA_CONTEXT_LENGTH (or your server's equivalent) and Context window to match
"Thinking…" counts up with no reasoning shown	a non-thinking model loading or working on slow hardware; the first answer after the model loads is the slowest
Answers arrive all at once instead of streaming	a proxy or load balancer in front of Polaris buffers responses; the answer still arrives
"Not permitted" in an answer	your role cannot read that area — the assistant is telling you, not failing
"Help is not available on this install"	the docs/wiki folder is missing from this install
Azure: "refused the API key (HTTP 401)"	the key is wrong or was regenerated — paste Key 1 or Key 2 again
Azure: "Entra ID refused the token request … AADSTS…"	the tenant ID, client ID or secret is wrong, or the secret expired; the AADSTS code says which
Azure: "refused the app registration (HTTP 403)"	the app has no data-plane access — grant its service principal Cognitive Services OpenAI User on the resource (role assignments can take a few minutes to apply)
Azure: "Deployment … was not found (HTTP 404)"	Deployment name holds the model name instead of the deployment's name, or the resource does not support the chosen API shape — try Deployments
Azure: "Unsupported parameter: 'temperature'"	a reasoning model; tick Omit temperature
Azure Claude: refused or not found right after the deployment was created	check API shape is Claude (Anthropic Messages API) and Deployment name matches the deployment exactly

Rule: 95.