LatticeDB: An Embedded Graph Database with Native Vector and Full-Text Indexing

Introduction to LatticeDB

LatticeDB is an embedded property-graph database presented as a single-file solution for managing connected, semantic, and textual data. It allows users to perform graph traversals, vector similarity searches, and BM25 full-text searches on the same dataset using a unified query language and engine. The database is built for relationship-heavy workloads on a single machine, emphasizing zero-configuration setup and an embedded single-writer model.

Key Features and Capabilities

The database operates as a single, portable file, eliminating the need for a server or complex configuration. Its core innovation lies in its single query layer, which integrates graph traversal, HNSW vector similarity, and BM25 full-text search. LatticeDB also includes a single event log, providing durable named streams and a built-in graph changefeed that shares the same transaction and write-ahead log (WAL) path as graph writes. It is designed for local-first applications, supporting one owning process per machine with WAL-backed durability.

Performance and Use Cases

LatticeDB demonstrates performance metrics such as 0.13 μs for node lookups and 0.83 ms for vector search with 1 million vectors at 100% recall. Example use cases include Graph RAG, agent memory, and local knowledge tools, which benefit from its ability to query data by relationship, semantics, and text. The database provides a unified approach to querying diverse data types within a local application context.

Developer Integration

LatticeDB offers bindings and installation methods for multiple programming languages. Developers can integrate it into Python projects via pip, TypeScript/Node.js applications using npm, and Go projects with specific cgo workflows. Installation scripts are available for quick setup, and the project supports bundling native libraries for various platforms.