Generative AI in FME: From Code Automation to Advanced Data Virtualization

Generative artificial intelligence is having a growing impact on how we work with data. Spatial data is no exception. In GIS and ETL environments, where data volumes are expanding faster than the teams managing them, Generative AI (GenAI) is helping to automate tasks that previously required specialized knowledge and hours of manual work.

Safe Software has integrated GenAI directly into the FME platform. On one hand, it serves as a built-in assistant that simplifies coding; on the other, it acts as an “engine” for advanced data workflows. These include RAG (Retrieval-Augmented Generation) solutions that allow you to “chat” with your organization’s own data.
Read on to find a structured overview of the use cases for GEN AI in FME.

FME AI Assist is a built-in helper within FME Workbench, available starting from FME Form 2024.0. It understands natural language and uses it to generate the code or expressions needed within your workspace.

A user simply types a command, such as “Select cities with a population over 100,000,” and the assistant generates the appropriate query in code, ready for immediate use in FME.

The assistant is integrated directly into editing windows. It appears exactly where syntax knowledge is typically required, such as when building SQL queries, Python scripts, or Regular Expressions (Regex).

How do you launch the AI Assistant, and what data does it use?

You can enable the assistant in FME Options. It can also be disabled at any time, leaving full control in the hands of the user or the FME administrator.

For beginners, AI support means a lower barrier to entry. They don’t need to know the syntax of SQL, Python, or Regex by heart. They just need to describe the task in words.

For experienced users, it represents real time savings. They no longer have to manually write repetitive code snippets and can focus entirely on process logic.

Beyond the built-in assistant, FME can connect to external AI services via transformers such as the OpenAICompletionsConnector, OpenAIChatGPTConnector, OpenAIImageGenerator, Generative AI Reader, and LeapAIConnector (Safe Software is continuously working on new solutions). This allows generative artificial intelligence to become an integral part of the data workflow.

In this section, we will look at several examples of how to apply generative AI when working with FME.

2.1. Contextual prompting in OpenAIChatGPTConnector

In FME, using the OpenAIChatGPTConnector, you can consciously control the behavior of the GenAI model. Instead of asking a single, general question, you can first define the model’s role and task, and then request a specific result with a defined structure. This ensures the response is better suited to the needs of the FME process.

For example, you could create the following prompt:

”You are a bot (role) that generates datasets regarding company profiles.

Generate a JSON file containing the 50 largest companies in Poland; provide the company name, number of employees, and revenue. Name the columns: company, employees, revenue. Return only the raw JSON file in your response.”

The resulting JSON can be loaded into FME and processed just like standard tabular data—filtered, merged with other sources, or published as a service.

This method of using GenAI in FME has very practical applications where you need to quickly generate or supplement data, especially during the prototyping or preliminary analysis stages.

2.2. Generating spatial datasets with automatic context

The Generative AI Reader in FME allows you to generate entire datasets on demand. Instead of manually designing a table structure, the user simply describes the dataset they need in the reader’s parameters.

For example: “The largest employers in Poland, including the city, industry, company name, address, and approximate number of employees”. The Reader creates the appropriate fields and rows. The resulting set can be immediately viewed in Visual Preview and processed further like classic spatial data.

This workflow can be expanded by chaining queries. Sticking to the example above, a user could first generate a list of provinces along with their capitals, area, and population. Then, for each capital, they could request a list of the largest employers in that city, including industry, address, and employee count. As a result, within minutes, a multidimensional dataset ready for analysis in FME is created. All of this is achieved without manually designing schemas or writing auxiliary code.

2.3. Creating graphics

This function can be used, for instance, to speed up work on map portals. Instead of manually searching for or creating individual icons, you can use FME and the LeapAIConnector to automatically generate a set of consistent illustrations based on object attributes. A good example would be building a city service portal that presents planned investments: new bus stops, EV charging stations, or selective waste collection points.

2.4. Self-service document processing application

Generative AI in FME can also act as the engine for web applications. A prime example is a solution where a user uploads a PDF report via a simple interface, and the system returns a summary prepared by AI. In FME, this is handled by transformers like the OpenAIChatGPTSummarizer and OpenAIChatGPTTextSummarizer.

Such a solution can be particularly useful for city halls and design offices that regularly receive extensive reports. Instead of reading every document from cover to cover, an employee uploads the file and receives the most important information within seconds. This makes it easier to decide which document requires a full analysis and which can be archived immediately.

Public AI models are oblivious to the data stored within specific organizations. They operate based on general knowledge from the Internet. When asked about something specific to your business, they may simply make things up (a phenomenon known as “hallucinating”). This is a serious limitation for business applications of generative AI.

The solution to this problem is the RAG architecture—Retrieval-Augmented Generation. In this approach, FME acts as the layer that prepares and provides the AI model with reliable data from company systems. FME retrieves documents, reports, GIS data, and other sources, processes the text into vectors (embeddings), and loads them into a vector database. This creates a database containing the organization’s specific domain knowledge.

Skillful prompt formulation is also key.

Short, precise sentences, specifying table and column names, and defining the expected result format significantly improve the quality of responses. If the first version doesn’t meet expectations, you need to refine the command and ask for a correction.

You must also keep costs in mind. Model API calls are billed in tokens. With large data volumes, suboptimal AI usage can prove expensive. A good practice is to filter data before sending it to the model and to limit AI usage to tasks where it truly adds value.

The final element is security. Even if AI Assist doesn’t send record values externally, other integrations might. You must adhere to GDPR rules and internal policies, and clearly define what information is allowed to leave the organization.

Summary

FME is becoming a real bridge between traditional data engineering and modern generative AI. AI Assist speeds up coding. Integrations with LLM models open up new scenarios, from generating synthetic data and creating graphics or web apps, to acting as an assistant for SQL and Python.

Meanwhile, RAG architecture and data virtualization allow for building systems that answer questions based on real organizational knowledge rather than general web data.

Now is a good time to start experimenting with FME and GenAI. It is worth doing so in a controlled manner to verify where GenAI genuinely accelerates processes and decision-making.