Article

AI with context: How product ontologies improve technical support chats

“How can I replace the cathode on the XY electron microscope?” could be a typical question that a service technician enters into a support chat. If the question-answering system (QA) only relies on generic AI without product-specific context, the answers will also be generic. Most likely, the response will simply be wrong or lack essential information – for example, the note that a special tool is required. In our field of technical communication, however, it is significant that answers are reliable, context-specific, and derived from verified sources. Only then can such a technical Q&A be a powerful tool.

To integrate company-specific information into a language model, the method of Retrieval Augmented Generation (RAG) is used. In short, documents are divided into pages or sections and written into a vector database. These vectors are then matched against the user’s query. While this method works fairly well, it has significant weaknesses when the question does not semantically match the possible answers.

Questions that RAG-based chats in technical environments often struggle to answer include the following:

Definition questions (“What is …”) → Problem: ensuring an exact definition
Hierarchy questions (“What is … composed of?”) → Problem: potential incompleteness of information
Questions about related content (e.g., related docs such as additional info from recent reports) → Problem: lack of connection between these types of information
Questions about specific characteristics → Problem: information may be too similar
Aggregation questions (e.g., number, sum, minimum, maximum) → Problem: missing context, no knowledge of the complete dataset
Comparison questions (e.g., comparing the same attribute across two products) → Problem: difficulty in identifying relationships

“Understanding of Question” – analyzing the query and recognizing the underlying concepts and relationships – therefore remains a challenge even for a question-answering system that leverages RAG for product-specific information. One solution could be to use AI not to generate a new answer, but instead to classify the question or refine the search query. The actual response would then simply come directly from the database. However, there is also another option.

How can a support chat with RAG be further improved?

To further improve a QA system that uses RAG, we need a semantic structure that represents both the connections between the data and their meaning. Knowledge graphs are well-suited for this purpose, as they represent concrete pieces of information as nodes and their relationships as edges. The foundation for this is an ontology – for example, a product ontology – that defines the relevant terms, categories, and possible relations.

In addition, we need metadata that not only describe the data but also help to understand their meaning. This includes contextual information such as the age of the information or the source from which it originates, as well as semantic relations within the ontology. Such semantic metadata link information to layers of meaning and make it machine-interpretable.

Digression: The iiRDS standard as a semantic structure provider for technical RAG chats

Technical writing offers a good starting point for supplementing technical Q&As with a semantic structure: iiRDS, a standard that specifies metadata for tagging technical information. It defines information types (e.g. concept or task), provides administrative and functional metadata, and includes a small amount of product metadata (components and product features).

Simple processes can be easily mapped using iiRDS classes alone: a topic references a component that is part of a product. This allows technical documentation to refer to a component, for example, without mentioning the (variable) item number. To do this, the procedure and component must be provided with a corresponding metadata that makes it possible to find the specific article.

If the structure to be mapped for technical Q&A is more complex than this, the information world of this standard can also be expanded. New classes or subclasses of the specified iiRDS classes can be created. In addition, new classes can be linked to those of the standard in order to build a new product ontology or expand an existing one.

What does an expanded iiRDS product ontology look like in practice?

“How do I replace the pump in the drive of my L50 machine? But it’s the US model. What is the rated speed of the drive?”
A question like this is so complex for a support chat, even with RAG, that it requires a product ontology to be answered. This is because the question covers: Which product? Which component? Which market? Which properties?

The pure iiRDS standard alone cannot provide all these answers. For this, the product ontology and its metadata model need to be extended by four classes: a Market Class, a Device Class (the relationship between device and product), a Property Class, and a Property Value Class—the latter two forming a framework for feature definitions.

A question such as the one about the rated speed can now be answered because the information object (in this case, a table with technical data containing these values) is referenced both with a product and with a device. This way, the QA system can distinguish which is the correct data source.

Advantages of semantics in technical Q&A

When a question-answering system that uses RAG is enhanced with semantic product data, the quality of the answers improves.

By categorizing knowledge, the search for information becomes more efficient.
Structured data lead to more relevant answers, as the language model better understands the context and the relationships between concepts.
In complex domains, users can find the information they need more easily because navigation is improved.
By comparing user questions with known concepts, the system can better recognize the intent and interests of the users.

Outlook: Semantics as a compass for AI systems

Modern question-answering systems, such as support chats, are already important tools for efficiently responding to user inquiries. With the advancement of generative AI and its growing autonomy, the importance of semantic information will increase in the future to guide AI-generated answers and their decisions. One way to achieve this is through the use of semantic structures such as product ontologies enriched with meaning-creating metadata.

Whether in direct information delivery – e.g., within a portal or a chat – this approach improves the findability, manageability, relevance, and analysis or extraction of information. For the increasing autonomy of AI – AI agents are already on the horizon—this creates a safeguard. For us in technical communication, this should be motivation enough to systematically advance the maintenance of metadata and product ontologies and to strategically apply technologies such as knowledge graphs.