Why AI agents need a graph layer next to vectors and SQL
Vectors retrieve similarity. SQL retrieves facts. A graph retrieves the connective tissue agents need when the next step depends on how entities, events, permissions, and history relate.
Blog
Notes on graph systems, AI agents, and connected context.
Field notes from building an embedded graph layer for modern data platforms, lakehouses, and retrieval-heavy AI products.
Latest theme
GraphRAG
How real relationships change retrieval, planning, personalization, and fraud analysis.
Embedded first, cloud when shared context matters
Start with a dependency. Move to managed graph capacity only when multiple services need the same traversal layer.
Zero-ETL graph projections over lakehouse tables
Use table formats and SQL DDL as the foundation for graph labels, edge types, and derived traversals.
Popular topics
Practical writing for teams connecting operational data, analytics, and AI systems.
Multi-hop retrieval without pipeline sprawl
Patterns for grounding LLM calls in relationship paths instead of isolated chunks.
Traversal patterns for application engineers
Neighbourhood expansion, shortest path, deduping, projection, and result shaping.
Turning SQL schemas into graph APIs
How tables, foreign keys, and views map cleanly to vertices, edges, and derived relationships.
Recent posts
Short, implementation-minded reads from the CrabGraph team.
CrabGraph 1.0: embedded graph server in one dependency
SQL schema definitions, persistent storage, SDKs, and TinkerPop-compatible traversals.
Planning graph capacity for shared AI workloads
How to think about hot traversals, fanout, materialization, and cost controls.
From relational data to first graph query
A quick schema walkthrough using vertices, edges, and views with standard SQL DDL.