What's Changed

• refactor(ai): Extract letterbox bbox correction into shared utility by @deven96 in #313

Full Changelog: bin/ai/0.2.1...bin/ai/0.2.2

What's New

Bounding Box Metadata for Face Detection

Buffalo-L and SFace face detection models now return bounding box metadata alongside embeddings:

• Normalized coordinates (0-1 range): bbox_x1, bbox_y1, bbox_x2, bbox_y2
• Detection confidence scores
• Metadata automatically stored in database and returned with queries

This enables applications to:

• Extract face regions from original images
• Draw bounding boxes for visualization
• Filter results by confidence score
• Perform spatial queries on face locations

Breaking Changes

• ConvertToEmbeddings API response structure changed from StoreKey to EmbeddingWithMetadata
• Client libraries updated to v0.2.2 (Rust), v0.2.1 (Python, Node)

Documentation

See the documentation for usage examples.

Highlights

• Introduced non-linear HNSW index for approximate O(log n) similarity search
• Improved memory footprint and insertion performance
• Optimized similarity search execution
• Added GetStore command for retrieving a single store with explicit StoreNotFound error

Highlights

• Added support for HNSW index configuration with default configurations
• Added support for GetStore
• Added support for new AI models (buffalo_l, sface, clipvit audio)

Highlights

• Added buffalo_l face embedding model
• Added sface + yunet detection/embedding pipeline
• Introduced runtime model arguments (e.g., adjustable confidence threshold)
• Added clipvit audio model (max 10s audio input)
• Added GetStore command support