A search-and-inference database from scratch in pure Zig
3 days ago
- Antfly rewrote its Go engine in Zig for first-principles reasoning, focusing on the model rather than embeddings, and adopting TigerBeetle-style simulation testing.
- Zig was chosen over Rust for portability, C interop, speed (no GC, no scheduler), and testability (std.Io enables VOPR-style fault simulation).
- The rewrite allowed full control over dependencies, enabling zero-dependency policy, holistic resource management, and embedding-friendly C/WASM interfaces.
- New capabilities include multiple data engines (document, relational, lake), storage modes (single-file to distributed), and indexing engines (full-text, dense/sparse vectors, graph).
- Enrichment engines process text, images, audio, and PDFs, with artifact-based indexing and support for fine-tuning and verification via VOPR and TLA+.