Industries are engaging with generative artificial intelligence (genAI) more than ever. Seemingly every day, genAI is evolving in ways that promise new opportunities and value. However, genAI can often seem overwhelming or too risky for organizations to implement. While there are some limitations, technological advances and complimentary tools like knowledge graphs are making genAI’s data analytics capabilities more robust. Knowledge graphs included in users’ AI environments can make model outputs more accurate, consistent, traceable, secure, and private. 

Gartner analysts recognize knowledge graphs as an essential infrastructure for organizations building more advanced genAI solutions. Knowledge graphs can include a semantic layer called an ontology, which provides understandable, comprehensive domain-specific meanings to an enterprise’s unique data. This can be used as a grounding context to provide the mapping between end-user questions and data for analytics processing and data-driven answers. Additionally, large language models (LLMs) can be combined with knowledge graph technology to deliver trusted, verified outputs using conceptual models as context.

Limitations of GenAI

However, like all technology, genAI is no silver bullet, and some limitations and challenges remain. These include the following:

Lack of Precision

For example, when genAI is asked for extreme precision, it doesn’t usually perform well. Many models struggle with simple math and logic problems. For example, LLMs asked to return pi to a few hundred decimal places are very unlikely to return the correct number. I general, it's usually easy to create a prompt that will generate a picture that gets some of the details right – but changing a couple minor details on a picture will not yield the best results.

Prone to Hallucinations

LLMs are deceptive in that they appear to be stores of knowledge when they really aren’t. Knowledge is a side effect of training data, probabilities of clusters of letters, the response being generated, and numeric representations (tokens) following earlier sets of numbers or tokens in the input. Responses are generated from learned paths and clusters of similar concepts from the model’s training, rather than a structured knowledge base or database.

Additionally, LLMs can’t provide current information until they’re trained with newer data. Information encoded into them is limited to the training cutoff date, and LLMs have no ability to update on their own. This means LLMs are prone to creating convincing hallucinations, particularly when responses to questions are unknown and the answer is based on similar information. LLMs produce information that appears extremely credible even when it’s not. Users are easily misguided, especially if they don't know the question’s domain well enough to distinguish a correct answer from an incorrectly hallucinated one.

Bias

As explained earlier, LLMs generate responses to inputs based on data patterns with no conscious thought. In other words, they lack real comprehension. This means LLMs are also prone to biases and limitations of their training data. They can provide responses influenced by those same biases and limitations with no regard to the user's sensibilities.

There are, however, ways around these limitations. For example, genAI performs well when a grounding context is part of the prompt. A grounding context limits the response generated by an LLM to mostly the information in that context. Provided the information in the context is accurate, the generated response is also likely to be accurate too. This genAI technique is called retrieval-augmented generation (RAG). The major search engines have started using RAG techniques with their existing search indices to generate answers to search questions quite effectively, often providing the reference URLs to the webpages used to ground their LLMs. Knowledge graphs are an important new form of RAG that supports a few recently developed techniques for both producing or utilizing grounding context. 

Less Specific Grounding Contexts

GenAI can still be useful in cases where there is less specific grounding context and more reliance on the internal real-world knowledge gained from training data. For example, genAI systems can do a good job with things like generating marketing content, writing poems and stories, or brainstorming creative ideas because they are broad questions with no “right” answer. Most genAI solutions are now built using combinations of techniques that rely on both these approaches.

How Knowledge Graphs Enhance Enterprise Data and Support GenAI Solutions

A knowledge graph allows for the visual representation of data, but its benefits extend far beyond this ability. Its primary value lies in its semantic modeling and descriptions, which provide meaningful context to enterprise data. Ontologies are conceptual models that describe data in terms familiar to business users and domain experts, streamlining data integration. Ontologies facilitate a semantic layer, which enables the easy addition and connection of new data through shared concepts. This makes it easier to discover, understand, and reuse data in ways that drive business value.

Queries against knowledge graphs offer tailored views of the data, providing customized services and experiences for different users. In essence, knowledge graphs serve both as a tool for integrating diverse data sources and as the result of applying ontology models to these datasets, creating a comprehensive and actionable knowledge resource.

Knowledge graphs support genAI solutions through the following:

Rich Contextual Foundation 

Knowledge graphs unify data sources in a multidimensional model, offering more depth than traditional databases. This information can be used in several different ways to enhance AI’s understanding of complex real world and domain specific subject matters. Knowledge graphs’ structured, high-quality data improves AI model accuracy and output relevancy.

Ontology-Driven Alignment

Ontology-driven alignment describes data in an abstract way from underlying sources and in human-readable context for people who consume the data. Knowledge graphs describe data in an abstract and human-readable language, aligning the LLM with how it’s trained on textual language. This alignment allows knowledge graphs to facilitate precise, context-aware AI solutions. It also facilitates accurate graph query generation when the ontology is used as grounding context to both guide and limit the generated graph queries.

Complex Relationships and Inference 

Knowledge graphs capture intricate data relationships and perform advanced inference. This interrelated information can be extracted and provide a richer context that empowers genAI to generate relevant, insightful answers to users’ ad hoc questions.

Scalability and Semantic Precision 

Knowledge graphs evolve and include data that’s necessary for real-time analytics or aggregation-style computation. This ensures AI outputs remain current and accurate.

Fine-Tuning or Model Training LLMs

Training is essential for getting internal data into solutions and making it accessible. By giving LLMs an understanding of the ontologies and reference data, LLMs can improve their outputs.

Supporting RAG with Vector Embeddings for Subgraphs 

Knowledge graphs can generate data by creating text fragments. A vector database vectorizes text fragments for integration. This method supports a RAG approach, where the LLM leverages vector embeddings to retrieve relevant information. This provides valuable context that enriches LLMs’ outputs. 

Just-In-Time Computed Grounding Context 

When data access needs to be dynamic, knowledge graphs provide just-in-time computed grounding context. LLM responses adapt in real-time to input specifics. Generated responses use optimized information and analytics for accuracy. 

Combining Vector Embeddings with Graph Attributes 

Vectors in a database store uniform resource identifiers (URIs) as metadata. The URI is represented in the knowledge base. Vector embeddings merging with graph attributes enriches data representation by including semantic relationships, detailed structural insights, and specific node attribute detail. This approach elevates AI’s grounding context and decision-making capabilities through a more sophisticated understanding of content, context, and connections. 

Prompt Orchestration  

Like Langchain, Llamaindex, and Haystack, gen-AI based solutions are delivered through open-source frameworks. These frameworks use knowledge graph technology. Depending on what the end user asks and what the solution does for an organization, managed workflows invoke different techniques with a series of prompts and tools. Knowledge graphs’ data is utilized through prompt orchestration.

Conversational Interface 

Users can have an interactive dialogue with their operational data within a knowledge graph. The conversational interface examines ontologies, allowing LLMs to answer an end user’s questions with text, tables, and charts. 

Conclusion

The combination of knowledge graphs and genAI is the backbone of modern data stacks. Knowledge graphs combined with genAI will unlock hidden insights in an enterprise’s data ecosystem. Knowledge graphs enhance data by making it more accessible and understandable, which benefits both general analytics and serves as a foundation for genAI. Users can apply knowledge graphs to increase the quality and accuracy of their data, merge their existing and new data to understand complex relationships, and enhance their current database into a multidimensional framework. 

Altair® Graph Studio™, an enterprise-level data discovery and integration toolset, hosts knowledge graphs. Users clean, harmonize, and interconnect data from multiple sources, simplifying access to structured and unstructured data. With one cohesive data layer capable of integrating many data sources, Graph Studio is built to answer users' ad hoc questions. Knowledge graphs merge siloed data into an interconnected data fabric unique to your enterprise.

To learn more about Graph Studio and knowledge graphs, visit https://altair.com/altair-graph-studio.