Beyond Vector Search: Leveraging Hidden Structure in Enterprise Data

‍

Most organizations categorize their data as structured or unstructured, but Meibel CTO Kieran Devarakonda challenges this binary thinking. In this GAI Insights Learning Lab session, he introduces "structurable data," which refers to the extractable structure and rich patterns that exist within seemingly unstructured enterprise information.

What You'll Learn

The Structurable Data Concept Understand how most "unstructured" information contains patterns that can be leveraged for more effective retrieval than focusing solely on semantic relationships. Kieran walks through concrete examples: novels have character relationships and timelines, product sheets contain embedded units and values, legal contracts have temporal information and entity relationships that can be represented in graph-like structures.

Why Semantic Similarity Isn't Enough See why embeddings alone fall short through the Maryland legal code example. When someone asks about statute of limitations for misdemeanors, the answer requires understanding not just the base rule but all the amendments and exceptions, which are relationships that semantic similarity cannot capture. As Kieran demonstrates, "otherwise we would have only matched everything that that is to say about the statute of limitations but we wouldn't have had any depth."

Multiple Storage Strategies Learn why different data types require different approaches. Vector stores are one tool, but relational databases might be most effective for product information while graph databases work better for legal document relationships. Sometimes you need all tools working simultaneously to provide the most robust context.

Implementation Realities Hear about practical challenges through real customer examples. A managed IT services firm's efficiency metrics were hidden in columns labeled "BH" (billing hours), something no semantic approach could discover without domain knowledge. Kieran emphasizes that while machine learning models can identify some implicit structure, customer guidance remains essential: Kieran emphasizes that there are going to be cases when your industry and domain knowledge is a critical input .

Technical Architecture Explore how the platform creates data catalogs during ingestion to identify structurable information within data sources. When queries come in, the system decomposes them and references the catalog to determine optimal execution strategies across different storage modalities.

Live Demonstration Watch the Maryland legal code demo showing how comprehensive answers require traversing multiple document relationships that standard RAG systems miss entirely. The visual representation reveals the passage relationships that were implicit in the legal code but necessary for complete understanding.

The Core Insight

This session fundamentally challenges how enterprises approach data preparation for AI. Rather than treating documents as plain text for vector databases, Kieran demonstrates why recognizing and extracting inherent structure, whether hierarchical, relational, or temporal, enables AI systems that provide comprehensive context instead of fragments.

The presentation emphasizes that this is collaborative work requiring domain expertise, not a fully automated solution. Success comes from combining machine learning models that detect patterns with human guidance that provides business context.

Originally presented at GAI Insights Learning Lab, hosted by Paul Baier.

‍