Stitching LLMs and data together with the power of semantic layers

Semantic layers act as the ontological glue between large language models (LLMs) and information, offering a structured framework for interpreting and organizing data. Through semantic layers, AI systems become more contextually grounded, yielding more meaningful interactions based on the complex relationships between concepts.

Experts from Vertesia and Graphwise joined KMWorld’s latest webinar, Empowering LLMs with a Semantic Layer, to dissect the ways semantic layers can unlock exponential value for both generative AI (GenAI) and agentic applications.

Grant Spradlin, VP of product, Vertesia, explained that with the explosion of various AI buzzwords—GenAI, RAG (retrieval-augmented generation), LLMs, etc.—it’s easy for meaning to be lost. That’s why Spradlin defined semantics for webinar viewers, where semantics often refers to the study of language meaning, or more broadly, the meaning of a word, phrase, sentence, or text.

Diving into its technological definition, a traditional semantic layer, Spradlin explained, is a business representation of data to help users find or understand it. These layers are critical for connecting users to unstructured content or complex documents; unstructured data is purpose-built for human consumption, but not for machines, which struggle to quantify and connect nuances in text, images, videos, graphs, and other symbols.

While many believe AI and LLMs are the end-all for machines being able to understand human language and concepts, this is a rushed conclusion. Spradlin explained that there are still many limitations; PDFs, for example, “were designed to look the same everywhere, but they were not meant to be understood by machines,” said Spradlin, and passing a PDF through an LLM loses critical information.

To combat this challenge, Vertesia’s Semantic Layer for Documents uses advanced AI solely for the analysis of documents while retaining the original text. By transforming PDFs to structured XML, Vertesia ensures that “not a single character [of the document] is altered, rewritten, or fabricated during the conversion process,” said Spradlin. “Our groundbreaking approach preserves the true meaning behind unstructured content.”

Vertesia’s solution also offers:

• Customizable semantic zone detection
• Identification and normalization of tables
• Intelligent processing and description of images
• Preservation of the full content hierarchy
• Zero hallucinations
• Preparation of LLMs for RAG

Andreas W. Blumauer, SVP growth, Graphwise, emphasized what’s at stake for businesses and AI. According to Gartner, 49% of leaders involved in AI report that their organizations struggle to demonstrate the value of AI. Ultimately, bad data management leads to AI project failure, noted Blumauer, and to be successful, enterprises must govern, integrate, connect, and contextualize their data and content.

Echoing Spradlin, the path toward this sort of robust data organization is paved with semantic layers, which Blumauer identified as three main types:

1. Business Intelligence (BI) Semantic Layer: Ensures consistent KPIs and business metrics; democratizes data analysis and querying for business users; works best with relational databases and well-defined schemas
2. Labeled Property Graph (LPG) Semantic Layer: Offers flexible and intuitive data modeling through nodes, edges, and properties; utilizes graph-based relationship analysis; optimized for complex, multi-hop relationship queries
3. Knowledge Graph RDF Semantic Layer: Uses ontologies and inference to connect and enrich diverse datasets; is standards-based and interoperable; supports logical inference, automated reasoning, and semantic search

Focusing on RDF semantic layers, Blumauer explained that its support for broad, multi-modal data enables enterprises to build AI with all their data while putting an end to hallucinations. Graphwise’s platform establishes this flavor of semantic layer—dubbed GraphRAG—simplifying the interaction between users and disparate data sources. With a consistent semantic metadata layer, an enterprise knowledge graph, and domain knowledge models, Graphwise enables organizations to successfully power their data for AI and GenAI use cases.

This is only a snippet of the Empowering LLMs with a Semantic Layer webinar. To view the full webinar, featuring detailed explanations, real-world use cases, a Q&A, and more, you can view an archived version of the webinar here.