Cache and Nova

Prepare your tenant cache#

Open Settings > Data. All supported resources are selected by default. Choose Preload all to collect every page of each selected resource, including available log history. You can deselect resource types or search the list. Search and Needs attention filter the list without changing your selections.

The page shows resource counts, collected pages, snapshot timestamps, required read permissions and refresh errors. Retry incomplete retries selected resources that are missing, partial or failed. Cancel refresh stops new requests and keeps previous complete snapshots for unfinished collections. Resources that already finished remain updated.

“All” means the app's supported resource catalogue. It does not include every Microsoft Graph endpoint, property, relationship or action. Microsoft retention, licensing and permissions determine what is available. Collection limits and failures are shown explicitly. A fresh cache does not imply a recent device check-in.

Automatic refresh preloads all supported resources while the app is running, this device is awake and the tenant is signed in. Choose an hourly, six-hourly or daily schedule. The cache is shared with Chat and Agent Team.

Talk to Nova#

Nova knows the selected tenant before you preload anything. For device and other tenant questions it asks the app to retrieve missing or stale relevant data. An empty cache does not mean there are no devices. Permission failures and partial coverage still need to be resolved through the tenant connection and the Data section in Settings.

Voice and reasoning have separate setup. The OpenAI key enables hosted speech and public web research; tenant answers use the reasoning provider and model shown in Nova. Connect that provider in Settings. Local voice needs a local reasoning model plus Whisper for recognition and Kokoro for speech output. Use Check voice setup to check key/model access or local service availability. Starting hosted voice separately checks the actual Live connection and billing, so a restricted key need not allow model listing.

Large cache collections stay in the local database. Nova gets selected evidence and short answers, with full results available through Open evidence in Chat. Preload before a presentation to reduce retrieval time; it is not required for Nova to recognize your tenant. Basic device counts, reported encryption and OS-version summaries come directly from the snapshot, avoiding a model round trip. More complex questions use the selected reasoning model. Freshness and partial coverage remain visible in the answer.

Nova carries recent conversation context into follow-up questions. If a question needs too much context, narrow it or continue in Chat. Stop cancels pending tenant work as well as audio; a new delegated question replaces the previous pending one. Greetings, jokes and “still there?” do not replace a pending tenant investigation. Nova keeps the current question separate from earlier spoken turns. Changing the reasoning model or its local/hosted status ends the voice session.

Use Full screen to fill the app window with Nova. During a conversation, only the animated orb remains after three seconds without interaction. Move the pointer, tap the view or use the keyboard to reveal the controls. Enable Conversation to see your words and Nova’s replies in chat bubbles, with live updates for cache reads, tenant queries, web research and answer preparation. The panel appears on the right in wide windows and below the orb in narrow windows. It starts hidden and remembers your preference between compact and full-window views. Exit full screen returns to the compact panel without interrupting the conversation. Setup and errors stay visible; Escape still ends the session. Mute mic keeps the conversation connected while disabling microphone audio; Stop ends the session and releases the microphone. Hiding Conversation leaves audio running. Activity cards distinguish finished answers, failures, stopped sessions and questions replaced by a new request. Expand View result for the backend answer, or use Open evidence for the full Chat record. The orb responds to microphone and hosted playback audio, with distinct waiting and speaking states. Reduced-motion settings disable animated movement. Open Voice settings while idle to change providers or the greeting name.

Choose Talk to Nova from any app page, or press Alt+V. Choose your voice provider, select a tenant, and click the orb to start the microphone. The orb responds to your voice volume; the optional panel lets you follow the conversation. Stop, Escape, closing the panel or changing the tenant/provider ends the session. Backgrounding the app also stops microphone use.

You can say “Hey Nova” during a running session. Set a greeting name in the panel. The app does not listen for a wake word in the background.

Ask about your tenant or say “Open Data settings”, “Open Agents” or “Open Settings”. Tenant questions use the existing Chat tools and the active agent model. Preload data before asking questions that need broad coverage. Open evidence in Chat to inspect the answer and its sources. Missing data is not proof of a healthy state. Voice does not approve or execute tenant writes; use the existing visual review.

OpenAI voice#

Choose OpenAI · GPT-Live-1, enter your API key and choose Save key. The key is kept in OS-secured storage. GPT-Live API access and billing are separate from ChatGPT or Codex CLI subscriptions.

Before starting, acknowledge that audio, shared tenant answers and public research queries are sent to OpenAI. The agent model may use a different provider; its normal hosted-context consent also applies. A local agent model does not make OpenAI voice local. Remove the saved key from the panel whenever you no longer need it.

If the connection fails, check API access, billing, network access and microphone permissions. On macOS, allow OpenAdminOS under Privacy & Security → Microphone.

Public web research#

In OpenAI Voice, Nova can choose tenant data, public web research, or both for a question. For example, ask about your device versions and compare them with current vendor guidance, or ask about a recent product announcement. This is general web research, not a predefined OS lookup. Simple tenant counts still use your cache without a paid search. Natural requests such as “What are the currently installed OS versions on my devices?” and “What is the encryption status of my devices?” also skip the reasoning-model round trip. Speech recognition and voice generation still add latency; cache access is only one part of response time.

The selected reasoning model decides when to search. Public research uses OpenAI's Responses API and GPT-5.4 mini with the same saved Nova key. Your API project needs access to that model and Responses; search and model usage incur API charges. Use a reasoning model capable of tool use. Nova uses its investigation loop for hosted research regardless of Chat's separate investigation preference.

Open evidence in Chat for clickable public sources and What ran for search queries, timing and failures. Public sources describe external facts, not your tenant's state. Missing or failed search evidence is reported explicitly. Each question permits up to three searches; narrow the question if research exceeds that limit or the voice context budget. A failed answer shows a recovery message; OpenAI Voice stays connected so you can retry or ask another question. Open Chat to inspect the failed investigation. Stop cancels pending research.

Research queries should contain public product names and versions, not tenant names, user details, device identifiers or credentials. The app sends the research query, rather than the full cache, to OpenAI. Local voice and ordinary typed Chat do not use this hosted search capability.

Local voice#

Choose Local · whisper.cpp + Kokoro, and select a local agent provider such as Ollama. Local voice uses two separately installed speech services:

• whisper.cpp server, listening on 127.0.0.1:8080, with a downloaded transcription model.
• Kokoro-FastAPI, listening on 127.0.0.1:8880, with the kokoro model and af_heart voice available.

Follow those projects' installation instructions for your operating system. OpenAdminOS connects to these services; it does not bundle or install their runtimes or models. Keep them bound to loopback. Model downloads need internet access during setup; local voice itself has no automatic cloud fallback.

Click the orb, speak, then choose Finish speaking. A recording is limited to one minute. Nova transcribes it locally, asks your local agent model, and speaks the result locally. If a service is unavailable, the panel reports the error so you can start that service and retry.

Send a result or start an agent#

After Nova completes an answer, try “Send this to my WhatsApp,” “Email this to me,” or “Send the report via Teams.” WhatsApp, Outlook/Exchange email, Teams, Slack, Discord and Signal use the connectors configured in Settings > Connectors. WhatsApp self messages go to the linked account; email to yourself uses the connected tenant account. Other requests use the connector's saved destination. Personal Teams messages require an explicit configured chat destination.

Nova displays the full message and destination before sending. Check both and click Confirm send, or cancel. Speaking “yes” does not approve a send. The preview expires after five minutes, and changing tenant or interrupting invalidates it. Nova reports whether the connector accepted the message; acceptance does not prove that the recipient received it. Do not retry an uncertain send until you check the destination. Sending through a connector shares the reviewed content even when local voice is selected.

Say “Run” followed by an installed agent's name to prepare an agent launch. Review the tenant and click Confirm run. Follow the run in Runs. Write plans still require their normal approval; existing saved delivery routes remain in effect.

Understand device answers#

All-tenant Changes lists use reported Intune device fields, with snapshot time and incomplete coverage disclosed. Missing encryption data is unknown, not unencrypted. Lists show up to 50 matching devices. For fleet questions about why devices are non-compliant, Nova reads actual failed compliance policy settings for up to ten matching devices. It reports remaining or unavailable details explicitly. A non-compliant status alone does not explain its cause. Policy details require the connected account to have permission to read device configuration.

In hosted voice, say Stop or click Stop answer to silence the current response and cancel pending investigation while leaving the microphone available. Click Stop to end the session. Result retrieved means the backend lookup is complete; it does not mean every word has been spoken. Local voice remains a recording-based flow and does not listen continuously while playing its answer.

Ask for a report and send it in one request#

You can say “Send me an email with the list of non-compliant devices.” Nova retrieves the relevant report, then displays its content and destination for review. It does not send anything until you click Confirm send. You do not need to ask for the list separately first.

“My WhatsApp” uses your linked account; “my email” uses your active tenant account. A configured Teams channel is shared with its members. It is not a private chat with you. Nova explains this when you ask for a personal Teams message; say “Send this via Teams” to review the configured channel. Slack, Discord and Signal defaults likewise do not establish which destination is personally yours. Always check the destination shown in the preview.

Nova checks connector setup and gives the relevant recovery steps. Long reports may need several numbered messages; the preview states how many before you confirm. If delivery stops partway through, Nova reports how many messages were confirmed accepted and stops the rest. It does not retry automatically. Check the destination before repeating an uncertain send.

Read formatted answers#

The Conversation panel displays lists, bold and italic text, headings, tables, quotes, links and code blocks. Expand View result to read the full retrieved answer while Nova speaks a shorter summary. Message previews use the same formatting. The full visual answer stays in the app rather than being added to the speech model's context automatically. Raw HTML and executable links are not rendered.