44 lines (31 loc) · 2.08 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.0.0] - 2025-12-02

Summary

First official production release. Achieves < 1ms lookup latency and significant memory reduction via Rust backend.

Known Issues

Full Cache Load Time: Loading a 1M-entry cache takes ~300-350ms total. While the binary index is search-ready in <10ms, Python pickle deserialization of response objects adds ~300ms overhead. This is a known limitation of the hybrid Python/Rust architecture and will be addressed in a future release.

Added

Rust Storage Backend: Core storage moved to Rust (RustCacheStorage), reducing memory usage to 44 bytes per entry (codes + metadata).
Persistence Format v3: New binary format with SHA-256 integrity checks. Loading 1M entries now takes < 10ms (index ready).
OpenAI Integration: Native OpenAIEmbeddingBackend with automatic batching, token bucket rate limiting, and cost tracking.
List-Based Response Storage: Python-side response storage optimized from dict to list, reducing overhead by ~90%.
Migration Tool: CLI tool to migrate legacy v2 caches to v3 format (binary-semantic-cache migrate).

Changed

Performance: Lookup latency for 100k entries reduced from 1.14ms to 0.41ms (2.8x speedup).
Memory: Total memory footprint per entry reduced from ~119 bytes to ~52 bytes.
Dependency: Added maturin build system for Rust extensions.

Fixed

Parity: Resolved bit-exactness issue between Rust and Python encoders (verified via OAI-08 regression test).
Windows Compatibility: Fixed Unicode encoding issues in benchmark reporting.

[0.2.1] - 2025-12-02 (Beta)

(Superseded by 1.0.0)

[0.1.0] - 2025-11-25

Added

Initial release.
Python-only implementation (NumPy + Numba).
Binary quantization (256-bit Hamming distance).
Ollama support (OllamaEmbedder).
Basic save/load persistence (pickle).