5.1.0

What's Changed

• Improve storage and BM25 search performance by @rudrankriyam in #74
• Use bounded top-K selection to avoid full result sorting by @rudrankriyam in #75
• Add realistic workload benchmark profile and optional CI stage by @rudrankriyam in #76

Full Changelog: 5.0.0...5.1.0

5.0.0

Highlights

• Add core SwiftEmbedder model-family resolution refactor and static embedding dimension handling improvements.
• Add ModernBERT support in SwiftEmbedder.
• Add RoBERTa and XLM-RoBERTa support in SwiftEmbedder.
• Add embedder validation and BERT fallback robustness fixes.
• Align package platform minimums with runtime requirements and clean up redundant availability guardrails.

Included pull requests

• #69 Refactor SwiftEmbedder core model resolution
• #70 Add ModernBERT support to SwiftEmbedder
• #71 Add RoBERTa and XLM-RoBERTa support to SwiftEmbedder
• #72 Embedder validation and errors
• #73 Raise platform minimums and remove availability guards

4.0.0

Breaking Changes

• VecturaMLXKit extracted to separate repository: VecturaMLXKit, VecturaMLXCLI, and MLX-related tests have been moved to rryam/VecturaMLXKit. This removes the need for Package Traits (SE-0450).
• Package Traits removed: The MLX trait no longer exists. MLX users should add VecturaMLXKit as a separate dependency.
• swift-tools-version reverted to 6.0: Swift 6.1 is no longer required for this package.
• mlx-swift-lm dependency removed: This package no longer pulls in any MLX-related dependencies.

Migration

Before (3.0.0 with traits):

.package(url: "https://github.com/rryam/VecturaKit.git", from: "3.0.0", traits: ["MLX"])

After (4.0.0):

.package(url: "https://github.com/rryam/VecturaKit.git", from: "4.0.0"),
.package(url: "https://github.com/rryam/VecturaMLXKit.git", from: "1.0.0"),

Both packages can be added directly via Xcode's "Add Package Dependencies" UI -- no workarounds needed.

Fixes #66

Full Changelog

3.0.0...4.0.0

3.0.0

Highlights

• Package Traits (SE-0450): MLX dependencies are now optional. Consumers using only VecturaKit or VecturaNLKit no longer download the heavy mlx-swift-lm package, significantly improving build times.
• Swift 6.1+ Required: Bumped swift-tools-version to 6.1 for Package Traits support.
• Smaller Default Model: Switched from potion-retrieval-32M to potion-base-4M for faster downloads and CI.

Breaking Changes

• Swift 6.1+ / Xcode 16.3+ required
• MLX users must enable the trait explicitly:

  .package(url: "https://github.com/rryam/VecturaKit.git", from: "3.0.0", traits: ["MLX"])

Usage

Without MLX (lightweight):

.package(url: "https://github.com/rryam/VecturaKit.git", from: "3.0.0")

With MLX:

.package(url: "https://github.com/rryam/VecturaKit.git", from: "3.0.0", traits: ["MLX"])

Changes

• Added #if canImport(MLXEmbedders) guards to all MLX-dependent source files
• Updated CI to run MLX tests separately with --traits MLX
• Default model changed to minishlab/potion-base-4M (users can still specify the 32M model explicitly)

Full Changelog

2.5.3...3.0.0

2.5.3

Highlights

• Hardened add/index rollback to restore overwritten documents and keep storage/search consistent.
• Stricter config/input validation (DB name, search options, embedder inputs); BM25 handles empty queries safely.
• Performance suites now default to a deterministic embedder for faster runs (opt-in real embeddings via VECTURA_PERF_USE_SWIFT_EMBEDDER=1).
• Updated MLX dependencies to latest mlx-swift-lm (2.30.3 / mlx-swift 0.30.3).
• CLI vector dimension is auto-detected by default; indexed storage guide defaults aligned.

Changes

• Storage I/O and candidate loading now fail fast on partial batch failures.
• Added tests for BM25 index and config validation; stabilized perf/memory expectations.

Tests

• swift test -c release --filter PerformanceTests
• swift test

2.5.2

• Fix SwiftLint trailing whitespace failures in MLXEmbedder.\n- Update the README installation snippet to 2.5.2.

2.5.1

Release 2.5.1: Throttled concurrency for file operations

Fixes potential file descriptor exhaustion and memory pressure when working with large document collections.

Changes:

• Add ConcurrencyExtensions.swift with concurrentMap, concurrentForEach, orderedConcurrentMap, and chunked utilities
• Replace unbounded parallel file loading in FileStorageProvider with throttled concurrentMap (max 50 concurrent operations)
• Add saveDocuments batch method to VecturaStorage protocol
• Implement saveDocuments in FileStorageProvider with nonisolated file I/O for true parallel writes while maintaining actor cache safety

2.5.0

Release 2.5.0: NaturalLanguage framework support via VecturaNLKit

Adds support for Apple's NaturalLanguage framework through a new VecturaNLKit module, providing contextual sentence embeddings with zero external dependencies.

2.4.0

Release 2.4.0: Package dependency updates and StaticEmbeddings model support

2.3.1

Release 2.3.1: Input validation, error handling improvements, and code quality fixes