# How to Add a Private AI Layer to Rockwell FactoryTalk On-Prem > Source: https://netray.co/erp-help/private-ai-layer-rockwell-factorytalk-on-prem-architecture > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: ERP Help ## Short answer FactoryTalk is a family, so pick the read surface per product: Historian SE through its PI-based interfaces, View SE alarm and event logs through the SQL database, ProductionCentre through its documented interfaces, and VantagePoint where you already consolidate data. Put a read-only tool server in front, serve an open-weight model locally, and respect FactoryTalk Security. ## Kind advanced ## System mfg-systems ## Erp Rockwell FactoryTalk ## Module FactoryTalk View SE / Historian SE / ProductionCentre / VantagePoint ## Query how to add AI to Rockwell FactoryTalk on-prem ## Alternate Queries - FactoryTalk Historian AI integration - FactoryTalk ProductionCentre AI assistant without cloud - local LLM for FactoryTalk alarm and event data - private AI Rockwell Automation plant data ## Applies To FactoryTalk View SE, FactoryTalk Historian SE, FactoryTalk ProductionCentre and FactoryTalk VantagePoint on-prem. Product versions differ; confirm interfaces against your installed release. ## Steps ### Reference build, in order **Items:** - Inventory which FactoryTalk products you actually run, and for each one name the supported read interface (Historian interfaces, SQL log database, product SDK or web services, VantagePoint models). - Create read-only identities in each system and, where possible, use a reporting copy or replica rather than the production server. - Build a tag and asset catalog from Historian and the HMI project (descriptions, units, areas) for retrieval. - Index plant documents: SOPs, batch and lot records, maintenance notes, alarm philosophy, with source permissions attached. - Write a tool server with named read operations per source and strict time and row caps. - Serve an open-weight model on local GPUs through vLLM or Ollama, connected only to the tool server. - Mirror FactoryTalk Security roles and areas into the tool server's allow lists, and log every call. - Roll out read-only first, then drafted outputs, then approved write-backs in business systems only. ## Sections ### The FactoryTalk read surfaces, product by product **Body:** FactoryTalk Historian SE is built on a PI Data Archive, so the same read patterns as the PI System apply: PI Web API where it is available in your release, or the PI-family SQL and SDK interfaces. Check what your installed version ships and whether PI Web API is deployed; do not assume it. FactoryTalk View SE keeps its runtime in the HMI server and its configuration under the FactoryTalk Directory. Alarm and event history can be logged to a SQL Server database when that logging is configured, which gives you an ordinary, queryable source. The exact table and view names depend on how logging was set up, so inspect your own database rather than trusting a generic script. FactoryTalk ProductionCentre is Rockwell's MES. Its business objects are best reached through its supported SDK or web interfaces for your version, with direct SQL reserved for read-only reporting if your integrator has documented the views. FactoryTalk VantagePoint is a useful single read surface if you already model plant data there. **Code:** # Historian SE (PI-based), if PI Web API is deployed GET https://ft-hist.plant.local/piwebapi/streams/{webId}/summary?startTime=*-24h&endTime=*&summaryType=Average&summaryDuration=1h -- Alarm and event log (SQL Server). Object names depend on your logging setup. SELECT TOP (200) * FROM WHERE >= DATEADD(hh, -24, GETDATE()) ORDER BY DESC; ### Reference architecture **Body:** Keep the AI stack on the business side of the network, behind the same firewall conventions you use between the plant and enterprise zones. The tool server is the only component allowed to reach the Historian, the log database and the MES interface, and it uses read-only identities. Retrieval combines metadata (tags, areas, equipment names) with text (SOPs, deviations, lot notes). Questions like 'which lots ran on press 4 during the pressure alarm' need MES lot data, alarm history and Historian values, and the tool server should expose a named operation for each of them. For an air-gapped site, load models and containers from approved media, run embeddings locally and keep the audit log in the isolated zone. ### Permissions through FactoryTalk Security **Body:** FactoryTalk Security centralises users, groups and actions through the FactoryTalk Directory, and authenticates against Windows accounts. It controls what a person may do in View SE, not what a SQL login can read, so you must bring the idea of areas and roles into the tool server yourself. Translate FactoryTalk groups and areas into allow lists of tags, equipment and lots. Apply them when building answers and when filtering the retrieval index, so a document or tag from a restricted area never reaches the model's context. Test by role, including a contractor identity, and keep the mapping in a reviewed file with an owner. ### Write-back, sizing and phases **Body:** Controller tags and HMI write paths are out of scope for an assistant. Useful write-backs are in business systems: a maintenance request, a quality hold, a draft downtime reason, each submitted through the target system's API after a person approves the exact payload. If ProductionCentre transactions are involved, use its supported transaction interfaces, never table updates. As a range, one 24 to 48 GB GPU serves a department of tens of occasional users; a few hundred active users usually need two to four GPUs plus headroom. A document index for tens of thousands of files is a few GB to low tens of GB. Phase 1: read-only alarm, trend and lot questions for one line. Phase 2: drafted shift and deviation reports. Phase 3: approved requests and holds in downstream systems. ### Limits and when the vendor tools are enough **Body:** If your goal is dashboards and reports across FactoryTalk data, VantagePoint and the standard reporting tools may already do it, and an AI layer is optional. The assistant earns its place when users must combine alarms, history, lots and documents in plain language, or when people do not know where to look. Expect integration effort to vary a lot by product and version. Do not promise a unified answer across View SE, Historian and ProductionCentre until each read path is tested individually. Rockwell and its partners also offer their own analytics offerings; they may be sufficient if cloud is acceptable. ## Pitfalls - Assuming one interface covers all FactoryTalk products. Each has its own supported read path. - Copying generic SQL for alarm logs without inspecting the real object names in your logging database. - Reading ProductionCentre tables directly when a supported interface exists for the same data. - Assuming FactoryTalk Security automatically limits what a SQL login can read. - Letting an AI service near controller or HMI write paths. - Querying the live HMI server for history instead of a log database or Historian replica. ## Ai Angle Netray builds the tool server and retrieval layer against whichever FactoryTalk interfaces you have approved, and links them to ERP data through ERPray so a question can span alarms, lots and purchase orders. Tool calls and sources are shown to the user. ## Faq ### Is FactoryTalk Historian the same as the PI System? **Answer:** FactoryTalk Historian SE is built on PI technology, so many PI patterns apply. Capabilities and which interfaces ship depend on the release, so confirm against your install before designing around PI Web API or other PI tools. ### Can the assistant acknowledge alarms or change setpoints? **Answer:** It should not. Controller and HMI write paths are outside the safe scope for chat-driven automation. Limit write-backs to business records such as maintenance requests, each approved by a person. ### Do we need VantagePoint? **Answer:** No. It is helpful if you already consolidate FactoryTalk and other data there, because it gives one read surface. Without it, the tool server can read each source directly. ### Can all of this run air-gapped? **Answer:** Yes. Models, embeddings, index, tool server and logs all run on your own servers, loaded from approved media. No outbound connection is required. ## Search terms AI for FactoryTalk, FactoryTalk Historian AI, FactoryTalk ProductionCentre AI, Rockwell Automation local LLM, FactoryTalk View SE alarm history SQL, private AI Rockwell plant data, air-gapped AI manufacturing MES, FactoryTalk VantagePoint AI, MCP server FactoryTalk, on-prem AI for MES --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json