Models and data protection

Genie uses AI models to write answers, find passages in your knowledge bases, describe images and turn speech into text. This page explains how the company's models are set up, how a model's price feeds the cost reports, and how Genie keeps sensitive knowledge away from models that run outside your company network. Models are configured once for the whole company: system administrators add them, and every workspace uses them. Workspace admins and members only pick among them, in agents, knowledge bases and the chat's model menu. System administrators also set data protection once for the whole company: the model data policy, the fallback models and personal data masking every workspace uses. Your Genie provider usually sets up the first models for you, so most companies only need to check them.

Model management

System administrators open Settings → Model management, the first item of the System administration group (Model catalog comes right after it). Workspace admins and members do not see this page. The models are sorted into tabs by type: All, Chat, Embedding, ReRank, Vision and Speech. Each card shows the provider, where the model runs (Internal or External, marked "detected" when Genie worked it out itself), the context window or vector dimension, and the price when one is set.

Type	What it does for people	Where it is chosen
Chat	Writes answers, summaries and the agent's reasoning	In each agent and in the default assistant; knowledge bases also use one for summaries and wiki pages; titles, memory and other background work use the model for background tasks
Embedding	Lets Genie find passages by meaning when it searches a knowledge base	In the knowledge base's settings, when it is created
ReRank	Puts the most relevant passages first before Genie answers	In each agent's settings
Vision	Reads images in documents and pictures people paste into chat	In knowledge base and agent settings
Speech	Turns speech into text: voice input in chat, audio and video documents	In knowledge base settings, and for the chat microphone under Voice input in chat on the Speech tab

To add a model:

1. Open Settings → Model management and click Add Model.
2. Choose the Model Type.
3. Choose the Provider, for example OpenAI, Azure OpenAI, Anthropic or Google Gemini. For a model server your company runs, choose Custom (OpenAI-compatible). Your Genie provider decides which providers the list offers.
4. Pick the Model Name from the list, or type it. Enter the Base URL and the API Key (optional) you received from the model's provider.
5. Under Advanced Options, check Where it runs (see Where a model runs).
6. Under Default, turn on Make it the default of its type if agents that have not picked a model should use this one (see Shared models).
7. Click Test Connection. When the test succeeds, click Save and close.

Every model you add is shared with every workspace at once. Workspaces use it without ever seeing its API key.

To change a model, click its card, or open its ⋯ menu and choose Edit. The same menu has Copy (a quick way to add a second model with the same connection), Set price (see Model pricing) and, on a shared model, Make default or Remove as default. The bin icon deletes a model once nothing uses it (see Shared models).

Shared models

Every workspace on the platform uses the models on this page, so no workspace has to connect models itself. Workspace admins and members cannot see a model's key, edit, price, test or delete it: they pick a model in agents, knowledge bases and the chat's model menu. When a type has no model yet, the model pickers say No model of this type yet. Ask your system administrator., and only system administrators see the shortcuts that add a model.

• The default. Each type has one default model; its card says Default. Set it when you add the model, or later from the card's ⋯ menu with Make default; Remove as default takes the mark off. Making a model the default takes the mark off the previous one. Agents that have not picked a model use the default: the default assistant and the built-in agents of every workspace answer with the default chat model until an Admin picks another for them, so a new workspace can chat right away.
• A workspace's own model from before. A model that a workspace added itself, before models were managed centrally, says This workspace only (added before models were managed centrally), and only that workspace can use it. You see it only while you work in that workspace. Its ⋯ menu has Let every workspace use it, which shares it as it is: its key comes along, so there is nothing to enter again.
• Deleting. The bin icon deletes a model only when no workspace's knowledge bases, agents or long-term memory use it. For a shared model, Genie otherwise names the workspaces that still use it; for a workspace's own model, it shows Model cannot be deleted with the places that use it, each with Open settings. Switch those to another model first. A shared model is never handed back to a workspace: to retire one, move its users to another model and delete it.
• Models from the deployment. Your Genie provider can also declare models in the deployment (builtin_models.yaml). Their cards carry a lock. You can edit them, price them and make one the default, but only the deployment can delete them.

The Model catalog button at the top of the page opens the catalog, which decides which model names the Model Name list offers when you add a model, with defaults such as the context window. It does not add models by itself. See System administration.

Model Test sends a real request to a saved model and shows its reply, for example to check a model after its provider changed something. Each test is a real call and may cost money.

Screenshot coming soonModel management: the company's models by type, with where each one runswebsite-docs/public/screenshots/en/models.png

Model management: the company's models by type, with where each one runs

Where a model runs

Every model is either Internal (it runs on your company network, self-hosted) or External (a third-party cloud service). This decides which knowledge bases the model may read under the model data policy.

Genie detects the location when a model is added: models at a private network address are internal; cloud providers such as OpenAI, Azure OpenAI, Anthropic and Google Gemini are external. A model that your company hosts behind a public web address can be detected as external. To correct it:

1. Open the model in Model management.
2. Under Advanced Options, set Where it runs to Internal or External. Detect automatically goes back to Genie's own detection.
3. Click Save and close.

When Genie cannot tell where a model runs, it treats it as external.

Model pricing

A price lets the Usage & cost reports and cost budgets show money, not only tokens.

1. In Model management, open the model's ⋯ menu and choose Set price.
2. Choose the Currency and enter the Input and Output prices per one million tokens. Cached input is optional; empty means the input price. For a speech model, enter the Audio price per minute instead.
3. Click Save price. Remove price takes the price off again.

• A new price applies to calls made from then on. Earlier calls keep the cost they were recorded with.
• A model without a price is still counted, but its cost shows as unknown, and the totals say "at least" the amount shown.
• For some well-known models Genie shows a built-in list price in grey on the card until you set your own. Set your own when your contract or Azure deployment charges differently.
• Only system administrators set prices, on Model management.
• A model your company hosts itself costs nothing per call: leave it without a price or enter 0.
• Prices also decide which model does the background work: see Model for background tasks.

Model for background tasks

Part of Genie's work happens in the background, where nobody is waiting for an answer:

• conversation titles,
• document summaries, generated questions, tags, knowledge base and spreadsheet descriptions, the knowledge graph and wiki pages,
• long-term memory (distilling what members said),
• route cards, which help the default assistant pick an agent.

This work uses the Model for background tasks. Out of the box it is automatic: the cheapest active chat model the workspace can use. Genie adds the Input and Output price per one million tokens: the price you set with Set price, or the built-in list price shown in grey on the card. Models with a price come before models without one, and with equal prices the model added first wins. So adding an expensive model for answers does not make every title and summary expensive.

A system administrator can choose the model instead:

1. Open Settings → System administration → Model management, on the All or Chat tab.
2. Under Model for background tasks, pick a chat model. Each model shows its price per 1M tokens. The first option, Automatic: the cheapest chat model (currently: …), names the model the automatic choice uses now.
3. Click Save.

The choice applies to every workspace. A workspace that picked its own model before models were managed centrally keeps it only while Automatic is selected here.

• A model chosen on the feature itself still comes first: a knowledge base's summary model (and its tagging, description and wiki models) and the Distillation model of long-term memory (see Privacy and retention). The model for background tasks is used when those are empty.
• This setting never changes the model that answers: agents and the default assistant keep their own chat model.
• Titles follow the model data policy. When the policy keeps a conversation away from the model for background tasks, Genie writes the title with the conversation's own model, and when that is not allowed either, uses the start of the question.
• If the chosen model is deleted, stops being active or is not a chat model, background work goes back to the automatic choice and the page says so until you pick another.
• Prices are compared as plain numbers, so price all your models in one currency.
• Changes are recorded in the system audit log as Model for background tasks changed.

For example, Contoso answers with an expensive reasoning model and also has a small, cheap chat model. With the automatic choice, titles, summaries and memory run on the small model; Minh Le picks the reasoning model under Model for background tasks only for a period when summary quality matters more than cost.

Voice input in chat

People can dictate a question with the microphone in the message box (see Voice input). The microphone shows only while a speech model is available. A system administrator sets voice input up on the Speech tab of Model management, under Voice input in chat, and clicks Save:

Setting	What it does
Show the microphone in chat	Off hides voice input for everyone
Speech model	The model that transcribes. Empty (Automatic: the default or oldest speech model) uses the default speech model, else the one added first. An agent's own speech model is used instead for that agent
Spoken language	The language hint sent with each recording. People can pick another one for their own recordings. Detect automatically sends no hint
Longest recording (seconds)	Between 10 and 600; 120 by default
Vocabulary	Names the model should spell right, such as products, branches and abbreviations

For example, Contoso adds OpenAI's GPT-4o Transcribe (gpt-4o-transcribe) on the Speech tab, makes it the default speech model, and lists Contoso, Fabrikam, Da Nang branch, SOP under Vocabulary.

While no speech model is available, system administrators see the microphone with a short note on how to set it up; everyone else sees no microphone. Like the model for background tasks, these settings apply to every workspace, and only shared speech models can be picked.

Model data policy

Every knowledge base has a sensitivity level: Public, Internal, Confidential or Restricted. Its creator or an admin sets it on the knowledge base (see Document permissions). The model data policy says, for each level, which models may receive that knowledge base's content. It applies to every workspace, and only system administrators change it.

To change the policy:

1. Open Settings → Data protection (in the System administration group). The policy is the first section, Model data policy.
2. For each level, under May be sent to, choose Internal models only or Internal and external models.
3. Tick Always mask personal data for a level whose content may go to external models but must always have personal data taken out first, whatever personal data masking says.
4. Untick Allow web search for a level whose conversations must never search the web, so its content cannot leak into search queries.
5. Click Save policy. Restore defaults brings back the defaults below.

Level	Meant for	May be sent to (default)	Allow web search (default)
Public	Content anyone may see, such as published price lists and brochures	Internal and external models	Yes
Internal	Everyday company knowledge. Every new knowledge base starts here	Internal and external models	Yes
Confidential	Personnel records, contracts, salaries and customer data	Internal models only	Yes
Restricted	Formulas, board papers, legal matters	Internal models only	No

A more sensitive level cannot be looser than a less sensitive one: the page shows an error and does not save. A knowledge base's own Data protection setting shows what its level allows. Changes are recorded in the system audit log as Model data policy changed.

Fallback models and masking for every workspace

Below the policy, the same Data protection page (in the System administration group) sets the Fallback models and masking that every workspace uses. Workspace Owners and Admins do not see it; they just use it.

1. Choose the fallback models and set up personal data masking, described below.
2. Paste a sample under Try it to check the masking before you save.
3. Click Save for every workspace. Genie confirms how many workspaces received the settings, for example Saved and applied to 12 workspaces. If some could not be updated, it says how many; click the button again to retry.

• Saving writes the settings into every workspace, and workspaces created later start with them.
• Until the company's settings are saved for the first time, each workspace keeps what it had and the page shows the defaults: no fallback, masking on for every kind except addresses. A workspace that never had settings of its own masks in the same way. Set everything before the first save: it replaces what each workspace had.
• System administrators can give one workspace its own settings as an exception, and saving then skips it. Workspaces with their own setting, under the fallback models and masking, lists them: Add a workspace gives one its own settings, Edit changes them and Use the company's value puts it back. A workspace's own fallback models can be any of its models.
• Every save is recorded in the platform Audit log as System setting changed.

To see what one workspace uses, look under Workspaces with their own setting: a workspace listed there uses its own settings (Edit shows them), and every other workspace uses the company's. To see what the policy blocks in one workspace, pick it in the filter of Models the policy blocks.

Fallback models

When the policy does not allow the model an agent or knowledge base normally uses, Genie can switch to a fallback model instead of refusing. Under Fallback models, choose a Chat model, a Rerank model and a Vision model, normally internal ones. They are used only in that situation. Only models shared with every workspace can be chosen, since every workspace receives the same fallbacks.

The model the policy blocks	With a fallback	Without a fallback
The model that answers	The fallback chat model answers	The question is refused
The knowledge base's embedding model	No fallback is possible	The question is refused
The rerank model	The fallback reranks the passages	Passages are used without reranking
The vision model	The fallback reads the images	Image analysis is skipped

Document processing follows the same rule: summaries, wiki pages, tags, image descriptions and transcription switch to the fallback, and without one the step fails and the document shows an error naming the knowledge base and the model. A fallback that runs externally cannot replace a model for Confidential or Restricted knowledge bases; the page warns you when you pick one.

Personal data masking

Masking replaces names, phone numbers, ID numbers and similar values with placeholders such as [PERSON_1] before a prompt goes to an external model. The model is told that each placeholder stands for a real value, so it writes the placeholder where the value belongs. Genie puts the real values back in everything people get: the answer, the conversation title, and the knowledge searches, database queries and files the assistant makes with its tools. Internal models receive the text unchanged. Masking is on unless a system administrator turns it off.

1. Keep Mask personal data for external models ticked; it is on by default. Untick it only if external models may receive personal data as written.
2. Under What to mask, choose the kinds: Names, Phone numbers, Email addresses, ID numbers (CCCD, CMND, passport, driving licence), Tax codes, Bank accounts and IBANs, Payment card numbers and CVV, Crypto wallet addresses, Social and health insurance numbers, Dates of birth, Ethnicity, religion, blood type, criminal record, Vehicle plates, Passwords, API keys, tokens, IP addresses, Device identifiers (MAC, IMEI), GPS coordinates and Addresses. Every kind except addresses is on by default; addresses are the hardest to detect precisely. Point at a kind to see what it catches. ID numbers include passports and ID numbers from Singapore, Korea, Japan and China, and the machine-readable lines of a scanned passport. Ethnicity, religion, blood type and criminal record are masked only where a form states them, such as “Dân tộc: Kinh” or “Tôn giáo: Không”. Kinds added in an update are ticked automatically, in settings saved earlier too; a kind you untick stays off. Crypto wallet addresses, Ethnicity, religion, blood type, criminal record and Device identifiers (MAC, IMEI) came this way in November 2026.
3. In Never mask, list company, brand and product names that look like personal names, one per line, for example Contoso.
4. Under Custom patterns, click Add a pattern for your own identifiers: a Label such as Employee number and a Pattern (a regular expression) such as CT-\d{5}. You can add up to 20.
5. Paste a sample under Try it and click Mask the sample to see What the external model would see. Nothing is saved or logged.
6. Click Save for every workspace.

• A level whose policy says Always mask personal data is masked even when this box is off; the page lists those levels.
• Web search queries have personal data removed, not replaced, since a placeholder means nothing to a search engine.
• IP addresses leave out numbers that only look like one: versions (Chrome/120.0.0.0), section numbers (Mục 2.1.3.4), amounts written with dots (2.150.200.100), netmasks and loopback addresses.
• Embeddings and images are never masked. For content where that matters, use a higher sensitivity level so it stays on internal models.
• Genie never stores the masked values. It records only how many values of each kind were masked.

Models the policy blocks

At the bottom of the Data protection page, Models the policy blocks, in every workspace lists every knowledge base and agent, in every workspace, whose configured models the policy does not allow at its sensitivity level, and what happens instead. The columns are Workspace, Where, Sensitivity, Setting, Model and What happens: Uses and the name of the fallback model, Skipped or Refused. The most sensitive come first. An agent counts at the level of its most sensitive knowledge base, and a fallback model that cannot serve a workspace's most sensitive knowledge base is listed too.

• Pick a workspace in the filter to see only its rows. Each workspace in the filter shows how many rows it has, so you can see where the problems are.
• The list shows 50 rows a page; the pages are below it.
• Click Check again after you change a model, a fallback or a sensitivity level. When nothing is blocked the section says Every configured model is allowed in every workspace.
• A workspace that could not be checked is named above the list; the server log has the reason.

Check this list right after a knowledge base is raised to Confidential or Restricted: anything marked Refused stops answering until it gets an internal model or a fallback. Whoever picks the level also sees a warning in the knowledge base's own sensitivity setting, listing which of its models will no longer receive its content.

Screenshot coming soonData protection: the model data policy, the fallback models and masking for every workspace, and the models the policy blockswebsite-docs/public/screenshots/en/data-protection.png

Data protection: the model data policy, the fallback models and masking for every workspace, and the models the policy blocks

What people see when a model is blocked

People never need to know the policy to stay within it. Genie tells them what happened on the answer itself:

• A fallback answered, web search was turned off or passages were not reranked. The answer's toolbar shows a Data protection note, for example Answered by … under the data policy, Web search was turned off for this answer or Results were not reranked.
• Personal data was masked. The same note shows Masked with the number of values, by kind.
• The question was refused. The answer says the question cannot be answered, names the knowledge base, its level and the model that runs externally, and says who can fix it. In Teams, Zalo and other channels the bot sends the same text as its reply.

Each refusal is recorded in the workspace audit log as Question refused by the data policy, with the knowledge base, its level and the model, never the question itself.

To fix a refusal, do one of these:

1. As an admin, give the agent an internal chat model in its settings, or
2. As a system administrator, set an internal Chat model under Fallback models on the Data protection page of the System administration group, or
3. If the refusal names an embedding model, move the knowledge base to an internal embedding model in its settings (see Knowledge bases). There is no fallback for embedding models.

For example, Contoso marks its Human resources knowledge base Confidential and answers with an Azure OpenAI model. Minh Le picks an internal chat model, which runs on Contoso's own server, as the fallback, so questions about salaries are answered by the internal model while everything else keeps using Azure OpenAI.

Good to know

• Only system administrators see Model management and add, change, price, test or delete models. Everyone else picks among the models in agents, knowledge bases and the chat's model menu.
• Saved API keys are never shown again on the page.
• A knowledge base keeps its embedding model. To move every knowledge base of every workspace from one embedding model to another, a system administrator uses Re-embed knowledge bases on the Embedding tab of Model management: each knowledge base keeps answering with the old model until its switch is done.
• Answers that draw on a knowledge base whose content may not go to external models are not copied into long-term memory or the conversation search index, where the policy could no longer follow them.
• Adding, changing, sharing, copying and deleting models, their credentials and prices are recorded in the audit log of the workspace the system administrator was working in. Changes to the model data policy, the fallback models and masking are recorded in the platform Audit log.
• The model data policy only checks where a model runs. Who may read a document is still decided by document permissions.

Related pages

• Document permissions: setting a knowledge base's sensitivity level
• Agents and the default assistant: choosing an agent's models
• Knowledge bases
• Usage, analytics and audit
• Privacy and retention
• System administration