The full power of TDengine, now free forever for up to 5,000 tags.

Explore More

On-Premises Industrial AI: Connecting a Local LLM to Plant Data

Juno Qiu

November 10, 2026 /

Running a large language model on local infrastructure can be part of an on-premises industrial AI architecture. It does not, by itself, show where every prompt, data record, response or log travels. An application may still connect to external services, use an outside identity provider or retain information in a separate logging system.

For an OT architect, the task is to map the entire request path and verify it against the site’s operating and security requirements. That means identifying where industrial data is accessed, where inference runs, which systems communicate across network boundaries and who can inspect or change the configuration.

Start with one permitted engineering task

Imagine a water-treatment plant considering an engineering assistant that summarizes a bounded equipment trend. An engineer asks for a summary of a selected measurement over an agreed time window. The intended result is a review aid, not an automatic command to equipment.

The architect identifies three distinct components:

  • Application or client: receives the engineer’s request, coordinates access to data and presents the response.
  • Industrial data platform: stores or provides access to the permitted measurements and their context.
  • LLM endpoint: receives model input and runs inference to generate a response. A locally hosted endpoint runs that inference on infrastructure within the chosen local boundary.

These components may be hosted on different machines or in different network zones. Their ownership and connections must be identified separately. For background on organizing industrial data and context, see the TDengine industrial AI architecture overview.

Follow the information through the system

A conceptual request may follow this sequence:

  1. The engineer submits a question through the application.
  2. The application accesses authorized industrial data needed for the selected equipment and period.
  3. The application prepares model input, which may include the question and selected retrieved context, and sends it to the configured LLM endpoint.
  4. The model generates a response, which the application presents for engineering review.

This sequence is a planning example, not a claim about the exact behavior of every product or deployment. The order, components and content exchanged depend on the application. Verify whether raw readings, aggregates, metadata or other context are actually sent to the model. Also determine what is recorded in application, platform, model and infrastructure logs.

Include the paths that are easy to overlook

A data-flow review should cover more than the main application request. Depending on the deployment, logs, telemetry, backups, identity services, support tools, model downloads and software updates may involve other systems or network destinations. The existence of these connections must be checked; they should not be assumed to be present or absent.

For each path, record the source and destination, the information that may be transmitted, the service identity, the network route, retention rules and the responsible owner. Review both the configured design and the observed behavior with authorized security personnel.

On-premises, private and air-gapped are not interchangeable

On-premises generally describes components hosted at an organization’s own site. It does not mean that those components have no external network access.

Private deployment describes a deployment intended for a restricted organization or environment. The term does not, by itself, specify whether infrastructure is on-site, in a private cloud or connected to external services. Ask for the actual hosting and access boundaries.

Air-gapped describes an environment isolated from external networks. Whether a complete solution meets that requirement depends on every required connection and operational dependency, including how updates and models are introduced. A local model endpoint alone is not proof of air-gapped operation.

With a locally hosted endpoint, inference occurs on the infrastructure hosting that model. The architect still needs to inspect the application’s data sources, network routes and supporting services. With an externally hosted endpoint, model input must reach that service; review what is sent, the provider’s handling of it, applicable permissions and network policy. Neither placement guarantees a secure or compliant system without a deployment-specific assessment.

Define access and operational responsibilities

Map the application, data platform and model endpoint to network zones. Identify which service account or user is authorized at each boundary and what industrial assets that identity may access. Check credential storage, transport protection, certificate trust, access revocation and the operational owner for each component.

Also consider what happens if the model endpoint is unavailable, a credential expires or a data source returns incomplete history. The application should not make an unsupported claim simply because a response is expected. Define how users see failures and when an engineer must return to the original source.

These checks are not a substitute for the site’s OT security process. A private network and a successful connection test do not automatically prove least privilege, compliance or safe operational use.

How can an architect validate the intended data flow?

Work with the data owner, application team and security staff to build evidence for one representative request.

  1. Inventory components: Record the application, data platform, model service, identities, endpoints and responsible owners.
  2. Trace information: Identify the intended request, retrieved data, model input, response and logging destinations. Check what the deployed system actually transmits and retains.
  3. Inspect dependencies: Review allowed outbound routes and the services used for identity, updates, telemetry, backups, support and model acquisition.
  4. Test boundaries: Under an approved test plan, verify that an identity cannot access an out-of-scope asset or endpoint and that unexpected connections are handled according to policy.
  5. Test failure and review: Check how unavailable data or model services are reported. Have the engineer compare the summary with source records and document remaining uncertainties.

For the hypothetical water-treatment assistant, this review might show that the model is local while a separate logging or update service uses another network path. The team would then decide whether that path is acceptable or needs a change. The example does not assert that any particular TDengine deployment has such a connection.

Distinguish local LLM access from TDgpt analytics

Some industrial AI functions use an LLM endpoint; others may use a different analytical component. TDengine IDMP’s LLM connection documentation describes an OpenAI-compatible interface for connecting to an external provider or a self-hosted model service. The actual model, version, endpoint configuration and task suitability need validation in the intended environment.

The same documentation distinguishes TDgpt-based anomaly detection, forecasting and missing-data imputation from the LLM connection, and states that those capabilities require TDgpt alongside IDMP. A local LLM and TDgpt therefore should not be treated as the same component or assumed to have identical deployment dependencies.

A connection to a local endpoint is a useful architectural option. It does not establish that every local model is compatible, that all data remains on-site, or that the complete system operates offline. Verify the model and the full information path before making those claims.

Related Articles