tomohiro-owada/devrag releases

Bug Fixes

• Fix search crash with zero-vector chunks (#22): vec_distance_cosine() returns NULL for zero vectors (stored as a fallback when chunk embedding fails). The scanner now handles NULL distances gracefully and skips those chunks instead of crashing.
• Fix CoreML GPU acceleration on Apple Silicon (#21): The MLComputeUnits option value was CPU_AND_GPU (underscore) but ONNX Runtime expects CPUAndGPU (CamelCase). M1/M2/M3 Macs will now correctly use CoreML GPU acceleration instead of falling back to CPU.

Full Changelog: https://github.com/tomohiro-owada/devrag/compare/v1.4.3...v1.4.4

Full Changelog: https://github.com/tomohiro-owada/devrag/compare/v1.4.2...v1.4.3

Fix: SanitizeForTokenizer space counting with problematic Unicode runes (#18)

Bug: SanitizeForTokenizer() replaced Box Drawing characters (U+2500–U+257F etc.) with spaces but reset the consecutive-space counter, allowing long space runs to survive and trigger the sentencepiece normalizer panic in sugarme/tokenizer v0.3.0.

Fix: Unified the space and problematic-rune branches so both increment spaceCount. This ensures that sequences like " ─── " are properly collapsed instead of producing 9 consecutive spaces.

Impact: Eliminates the remaining 2 chunk failures reported in v1.4.1 testing (ASCII flowcharts and wide Markdown tables with Box Drawing borders).

Credit: @hrkzogw for the thorough root cause analysis and suggested fix.

Bug Fixes

• Fix tokenizer panic root cause (#18): The real crash trigger was runs of 4+ consecutive spaces (deep indentation), not Unicode characters. The sentencepiece normalizer in sugarme/tokenizer v0.3.0 has an off-by-one bug in byte offset tracking for long whitespace runs. SanitizeForTokenizer() now collapses runs of 4+ spaces to 3 before tokenization, eliminating the remaining zero-vector fallbacks.

  Thanks to @hrkzogw for the detailed root cause analysis.

New Features

• Configurable global model cache directory (#16): The embedding model (~450MB) is now cached in an OS-appropriate location by default, avoiding redundant downloads per project.

  • macOS: ~/Library/Application Support/devrag/models
  • Linux: $XDG_DATA_HOME/devrag/models (fallback ~/.local/share/devrag/models)
  • Windows: %LOCALAPPDATA%\devrag\models

  Override with (highest to lowest precedence):

  1. --model-dir <path> CLI flag
  2. DEVRAG_MODEL_DIR environment variable
  3. "model_dir" field in config.json

Bug Fixes (from v1.3.2)

• Fix tokenizer panic on Unicode-heavy markdown files (#18)
• Fix startup hang on macOS Apple Silicon (#17)

Bug Fixes

• Fix tokenizer panic on Unicode-heavy markdown files (#18): Added Unicode sanitization for box drawing, block elements, geometric shapes, and braille characters before tokenization. Added panic recovery to all Tokenizer methods as defense-in-depth against sugarme/tokenizer v0.3.0 bugs.
• Fix startup hang on macOS Apple Silicon (#17): Changed CoreML MLComputeUnits from ALL to CPU_AND_GPU to avoid Neural Engine compilation hangs. Added timeout-based session creation (60s for CoreML, 30s for CPU) with automatic CPU fallback when CoreML fails or times out.

Full Changelog: https://github.com/tomohiro-owada/devrag/compare/v1.3.0...v1.3.1

Full Changelog: https://github.com/tomohiro-owada/devrag/compare/v1.2.2...v1.3.0

What's Changed

• Make physical file deletion optional in delete_document tool by @tomohiro-owada in https://github.com/tomohiro-owada/devrag/pull/11
• Add comprehensive unit tests for MCP tools by @tomohiro-owada in https://github.com/tomohiro-owada/devrag/pull/12
• Remove unused files and test binaries by @tomohiro-owada in https://github.com/tomohiro-owada/devrag/pull/14

Full Changelog: https://github.com/tomohiro-owada/devrag/compare/v1.2.1...v1.2.2