DataFocus · Journal
Which LLM does OT Flash use, and how do you control model cost and data egress?
OT Flash names no built-in model. An administrator configures the LLM service, and that choice decides both what data leaves your network and what the model bill looks like. A walk through the admin screens, the egress decision and a monthly review habit.
OT Flash does not ship with a fixed model. According to the OT Flash r57 user manual, the ontology agent can be used only after an administrator has configured a large language model service, so you choose the model and where it runs. That choice controls both questions in your title: what leaves your network (whatever the configured service receives) and what the model costs (whatever that service bills per token). The same manual describes an admin screen, Device authorization, that shows daily and monthly AI model token consumption, which gives you a number to review each month. This post walks through where those settings live and how to turn them into a cost and egress decision. It does not recommend a vendor or model, because the sources do not name one.
Does OT Flash bundle a model?
The documentation does not say so. The manual states one requirement: an ontology agent works only once an administrator has set up an LLM service in the system. It also notes that an answer usually takes tens of seconds, longer for complex questions. The product brief says the same thing at a higher level: DataFocus OT is self-hosted, administrators configure the large-language-model service, and with an external model the customer should confirm what is sent to it under its own data policy (DataFocus OT product brief).
Two things follow for a decision-maker:
- Nothing in the published material ties Flash to a particular vendor or model, and this post does not either. Treat the model as a component you select, not a feature you receive.
- "Self-hosted" describes where the application runs. It does not by itself tell you where the model runs. If the service you configure is hosted by a third party, the application is local and the model calls are not.
Where do you configure it?
In the r57 manual, system-wide settings sit under Admin > Configure. Its sub-tabs include Base Configure, Prompt Management, Custom Ai, Functional cropping, SSO and Email Template, among others. The manual's tab list does not say which sub-tab holds the LLM service, so look at Custom Ai first on your own instance.
The manual lists preset roles, and SYSTEM_MANAGER is the one described as handling global configuration and user and role management. So the practical first step is to decide which named person holds that role for model settings. This post does not reproduce any model-service fields. The manual's chapter 13 (Admin) is the place to read about the Configure tabs for your release, and what you see on your own instance is authoritative.
A short procedure for the first setup:
- Decide the egress class (next section) before anyone opens the model settings.
- Have the SYSTEM_MANAGER-role administrator configure the LLM service in the system settings under Admin > Configure (start with Custom Ai).
- Open the ontology workbench and ask the agent a question about non-sensitive sample data. If the agent works, the service is reachable. If it does not, the manual's stated precondition (a configured LLM service) is the first thing to check.
- Open Admin > Device authorization and note the daily and monthly token figures as your starting reference point.
How do you decide what data may leave?
The deployment checklist on DataFocus's site lists the items to settle before a local deployment: usage scope, operating system and architecture, capacity, data authorization, model choice, data egress and backup ownership. Model choice and data egress are separate lines, which is a useful reminder that they are separate decisions. A model can be cheap and unacceptable for your data, or acceptable for your data and expensive.
To make the egress question concrete, look at what an agent sees. In the ontology agent, the manual says cited data comes from the tables and documents bound to the ontology, and a "View processing details" link shows which data the agent queried and which tools it called. The agent also reads the ontology's documents. Those are the materials that can end up in a model request. The manual does not publish the exact content of each request, so the safe reading is that anything bound to an ontology is a candidate for the model to see.
Egress classes you could use as a policy starting point (illustrative, not a DataFocus feature):
| Class | What it means | Fits when |
|---|---|---|
| A: model inside your network | The configured service runs on infrastructure you control | Ontology tables or documents hold data your policy keeps on-premises |
| B: external model, scoped ontologies | A third-party service, but only ontologies built from data you have cleared for it | Mixed sensitivity; you can split ontologies by business topic |
| C: external model, general use | A third-party service with broad ontology content | Mostly public or low-sensitivity material |
Because an ontology is a knowledge space around one business topic (tables, documents, indicators), splitting by topic gives you a natural unit for these classes. That is a design suggestion drawn from how the manual defines an ontology, not a documented egress control.
There is a second egress path to review if you connect external agents through MCP. In that route, a client you choose (the brief mentions clients such as Claude Code) sits in front of the data. Query results come back to that client as evidence rows, and the model behind the client sees whatever the client puts in its context. The manual limits what that client may reach to the intersection of the external-agent setting, the exact OAuth grant and the resource's edit access, so the data reachable is bounded by permissions. The destination of that data is still a decision for you: the connection guidance on the site says to check data destinations, authentication, token scopes and revocation. The manual also advises logging only necessary correlation data, not tokens or full sensitive content.
How do you watch token consumption?
Admin > Device authorization is described in the manual as showing device license information and the daily and monthly AI model token consumption. The manual does not describe license tiers or prices, so this post will not either.
What you can do with a daily and monthly figure:
- Record the monthly number on a fixed date and compare month to month.
- Look at the daily series around events you can name: a new ontology went live, a team started using the agent, someone began generating scripts.
- Compare the figure with your model provider's own invoice or usage report. The manual says what the screen shows; it does not say how the figure is computed or whether it matches a provider's billing unit, so treat the invoice as the authority for money and the screen as the authority for trend.
Which features draw on the configured model? The manual states the requirement for ontology agents. It also describes AI-assisted actions inside the product, such as drafting an execution script from a description and generating or modifying a control panel draft. The sources do not say which of these counts toward the token figure, so confirm on your instance before assuming the screen covers every AI feature. If you also run an external client over MCP, that client's own model usage is billed wherever the client's model runs, outside this screen.
A hypothetical monthly review sheet (illustrative, not run, and with no real numbers):
| Month | Monthly tokens (screen) | Provider invoice | Notable change | Owner's note |
|---|---|---|---|---|
| (fill in) | (from Device authorization) | (from provider) | (new ontology, new team, etc.) | (accept / investigate) |
Will falling model prices make this a non-issue?
Partly, and the evidence has limits. Stanford HAI's 2025 AI Index Report says the inference cost of a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024. That is a statement about one capability level over one period. It does not predict the price of the model you pick next year, and it says nothing about usage volume, which rises when more people ask more questions. A falling unit price and a rising count of questions can move the bill in opposite directions, which is a reason to keep the monthly review even if prices fall.
What other costs stay on your side?
DataFocus's site is direct about this: free self-hosting still requires infrastructure, model APIs, third-party services and operational support, and "you pay for servers, model calls and third-party services." The download page adds that the Community Edition license does not cover servers, model calls or operations. So model spend is one of several lines you carry. Alongside it sit the servers that run Flash and whoever operates them. The sources give no sizing, so this post gives none.
When this approach is not enough
- If your policy requires proof of exactly what each model request contained, the published material does not describe a per-request content record for the model service. Ask DataFocus through a support ticket, or verify on your instance, before you commit.
- If you need a single reconciled bill covering application and model, the Device authorization figure is a usage view, not an invoice.
- If you want a vendor recommendation, you will not find one here. The sources name no certified or bundled model.
Next step
Read the Admin chapter of the OT Flash r57 user manual for the Configure and Device authorization tabs on your release, then settle the egress class with your security owner before the first ontology goes live. The product brief covers the self-hosted model in one page.