Graph Shapes Over Queries for AI Agent Context

Deterministic graph outlines tackle a limitation of vector search: they answer 'what are we missing?' and navigate document structure. By building a table of contents from URIs and links—not embeddings—agents get a navigable, hierarchical view. This reveals gaps and supports cross-document reasoning without LLM-based entity extraction.

For structured data, a metadata graph (NeoCarta) acts as a semantic layer. It maps tables, columns, and join paths, enabling agents to generate correct SQL across many tables without moving actual data. This avoids data migration, security risks, and the cost of full ETL, while improving text-to-SQL accuracy.

The approach shifts context building from heavy LLM pipelines to lightweight, deterministic loading. It works best when documents have inherent structure; messy, unstructured text still needs extra parsing. The trade-off is that metadata graphs limit graph queries to schema-level insight, not raw data analytics, unless you later import data for algorithms.

For engineering leaders, these patterns offer incremental adoption: start with a metadata layer for fast text-to-SQL wins, then add document outlines for navigation, and later bring full data into a graph for deeper algorithms. The core insight: context can be assembled from pre-existing structure, not just generated.