The landscape of graph databases is shifting. For years, the industry was dominated by massive, server-centric architectures designed for enterprise-scale silos. However, the rise of local-first software, edge computing, and AI applications running on developer machines has created a demand for speed, portability, and simplicity.
Better support for CALL {} subqueries, allowing developers to execute complex conditional logic or localized aggregation within a broader matching clause.
Improved string, mathematical, and list-handling functions that mirror modern Cypher specifications, simplifying the migration of legacy queries to Kùzu. 3. Faster Data Ingestion (Copy Layer Improvements)
For developers (Python/C++/Rust/Node.js/Go), this version focused on API cleanup.
: Kùzu utilizes "worst-case optimal joins" (WCOJ) to handle dense, cyclic graph structures with high efficiency. Enhancements in Recent Versions
Graph creation requires ingestion from external formats like CSV, Parquet, or Arrow. In v0.1.3.6, the COPY FROM command features improved parallelization. The database engine splits larger files into smaller chunks more efficiently, ensuring that multi-threaded ingestion saturates available CPU cores without introducing thread contention. Seamless Integration with Arrow and DuckDB
To facilitate fast forward and backward traversals, Kùzu maintains dual-indexed adjacency lists for relationships. Whether you traverse from source-to-target or target-to-source, Kùzu locates the corresponding column blocks with
Since Kùzu is an embedded database, it runs directly in your application process without needing a separate server. Python : pip install kuzu Node.js : npm install kuzu Rust : cargo add kuzu Basic Usage Example (Python)
: Improvements to full-text search (FTS) and vector indices, which are crucial for AI-driven applications and Retrieval-Augmented Generation (RAG).
Variable-length path queries (e.g., MATCH (a)-[:FOLLOWS*1..5]->(b) ) are the ultimate test for any graph database. In v0.13.6, the execution planners and join indices have been fine-tuned. The engine now better predicts traversal densities, reducing the overhead of evaluating recursive boundaries. This results in measurable speedups for fraud detection patterns and lineage tracing queries. 3. Storage Layer Consolidation