Context Mole v0.2.7

Context Mole v0.2.7

Context Mole v0.2.7 improves indexing efficiency, makes project progress easier to understand, and strengthens pause and restart recovery.

OCR and text extraction

• PDF OCR reuses renderers and sends page pixels directly to the OCR engine, reducing repeated work and memory copies while preserving page resolution.
• Page images are allocated only when OCR can process them, keeping queued files from holding large rendered images in memory.
• Long text lines keep their proportions during recognition. A historical scan used in testing now yields 1,455 characters instead of 112.
• Spreadsheet rows are read incrementally, email resources are released promptly, and image/TIFF OCR timeouts appear as current issues.

Clearer project progress

• Separate Up to date, Processing, Queued, and Needs attention counts account for all discovered files.
• Scheduled retries and files waiting for processor capacity are included in the queue. Searchable files are counted separately because previous indexed versions remain available during updates.
• Current issues can be collapsed and browsed in pages of 25. Resolved failures clear, and large projects retain every outstanding issue.
• Unavailable or unreadable folders have a clear status that clears after a successful rescan.

Recovery and performance

• Folder discovery runs alongside indexing, and large bursts of file changes use bounded buffers with a rescan when needed.
• Ordinary change notifications verify file content before repeating extraction. Reindex still forces a complete rebuild.
• Pause, shutdown, and restart retain unfinished jobs and the last searchable revision. Shutdown also waits for discovery cleanup when cancelled during model startup.
• Similar-length passages are grouped for both FP32 and quantized embeddings. Local benchmarks measured 2.2–2.4× faster FP32 processing and about 1.5× faster quantized processing. Quantized retrieval relevance stayed stable or improved overall on the labeled test workload; individual ranking differences and workload limits are documented.
• Large embedding requests use smaller transfers, and outdated vector caches are released when searches switch to streaming.

Validation and benchmarks

The Release build and all 282 automated tests passed locally. Five checked-in public fixtures cover native PDF, Word, scanned PDF, JPEG, and multilingual TIFF extraction, with source attribution, licenses, hashes, and quality checks. All five fixture checks passed with the pinned OCR models.

See benchmark results and commands to reproduce the measurements. Performance varies with the documents and machine.

Use Reindex on existing projects to apply the improved OCR to files already indexed.

Install

Download JPCodeCraft.ContextMole-win-x64-Setup.exe for Windows 10+ x64. Releases are currently unsigned, so Windows SmartScreen may show an unrecognized-app warning.

---

Full Changelog: v0.2.6...v0.2.7