Why I built it
AI and retrieval systems often end up stitched together from several separate products: one store for graph structure, another for vectors, another for full-text search, and then a layer of application logic to reconcile the results. LatticeDB is an attempt to collapse that stack into a single embedded system with a smaller operational footprint.
What it does
LatticeDB is an embedded knowledge graph database built from scratch in Zig with zero external dependencies. It stores everything in a single file, in the spirit of SQLite, but is designed around AI retrieval workloads rather than general-purpose OLTP.
A single query can traverse a property graph, run HNSW vector similarity search, and execute BM25 full-text matching without bouncing between separate services. The query language is Cypher-like for graph work, with extensions for vector and text retrieval.
Why it matters
The project is a concrete expression of a broader interest of mine: retrieval and context are systems problems, not just prompt problems. If the storage engine, query model, and operational footprint are wrong, the rest of the stack inherits that complexity.
Client libraries are available for Python and TypeScript/Node.js.