ZeusDB Vector Database Files

Changed

• Unfiltered search is faster. The live record set the traversal checks each candidate against is a bitmap rather than a hash lookup, which removes a HashMap probe from every candidate the graph visits. Measured paired in one interpreter at 50,000 records, this is 22 percent on SIFT-128 under l2, 18 percent on GloVe-100 under cosine and 6.5 percent on DBpedia-OpenAI-1536 under cosine. Every result page is identical to 0.8.0's, id for id and score bit for score bit.

• rand moves to 0.10 and rand_chacha to 0.10. The generated stream is unchanged, verified against every seed the engine records over a million draws each, so every graph, codebook and saved directory is exactly what 0.8.0 produced.

Internal

• The engine is a Cargo workspace of three crates, zeusdb-vector-core, zeusdb-vector-hnsw and the Python binding, none published to crates.io. The source distribution carries all three under crates/ and bindings/. Neither engine crate depends on PyO3, so the engine builds and tests without a Python interpreter.

• Every failure the engine raises is one Error type, mapped to a Python exception at the binding boundary. Every exception class and every message is unchanged.

• HNSWIndex is a Python wrapper over the engine's collection type. Its methods, properties, return values and docstrings are unchanged.

• The duration_ms field on the add completion log record now measures the insertion alone, where it previously included parsing the arguments. Log record targets are unchanged.

• save() now flushes the graph file to disk before the directory is renamed into place, closing a window where a power loss could leave a complete manifest naming an incomplete graph.

• RUSTSEC-2025-0141, bincode being unmaintained, is accepted with a review date. No version of bincode clears the advisory, and the artefacts stay on 2.0.1 until they move to another format.

Unchanged, and verified

Across the restructure, 148 error probes are identical in class and message, 64,800 result pages across 24 index configurations are identical, every saved artefact is identical by content, all 24 recorded directories reload identically, and the saved directory format is unchanged at 1.1.0. Every directory saved by 0.8.0 opens as it did.