Build a semantic ontology to power AI assistants on AWS – Part 1

This article explains how to build a semantic ontology that helps AI assistants navigate enterprise data efficiently using a bottom-up approach grounded in actual data structures.

• Dumping raw schemas into LLMs degrades performance; semantic layers provide targeted context retrieval instead
• Bottom-up ontology strategy automates foundational 80% using LLM-guided extraction rather than manual top-down design
• Three-layer architecture: property graph for structure, vector index for semantic search, facts layer for learned assertions
• Four-step pipeline: populate base graph hierarchy, profile columns to find hidden relationships, generate deduplicated business terms, index for semantic search
• Column profiling with LLM identifies semantic relationships like "describes" or "decodes" between related columns
• Semantic deduplication achieves 30-50% term reuse by matching business concepts across technical columns
• Hybrid search combines keyword matching with vector similarity to find entities by meaning, not just exact names

This approach enables agentic systems to efficiently reason through complex data tasks by providing the right information architecture rather than relying on larger context windows.