February 2026 focused on security hardening and compatibility updates for DS4SD/docling. The primary deliverable mitigated a known Pillow vulnerability (CVE-2026-25990) by relaxing version constraints and augmented the asr optional dependencies to support Python 3.11+ with Numba. These changes improve security posture, reliability across environments, and align with the project’s stability goals.

June 2025: Delivered targeted reliability improvements in the docling-eval cloud table processing module. Fixed text duplication in table extraction across Azure and Google, refined how table and paragraph data are extracted to prevent overlapping content, and improved handling of provenance items. Also resolved a divide-by-zero error in Google's prediction provider, stabilizing predictions for cloud-based workloads. These changes reduce data quality issues, prevent runtime errors, and enhance cross-cloud compatibility for downstream analytics and evaluation pipelines.

May 2025 performance summary for docling-eval: Delivered cross-provider layout-aware data extraction enhancements and strengthened reliability across AWS Textract, Azure Document Intelligence, and Google Document AI integrations. Key improvements include layout extraction, SegmentedPage support, and word-level OCR, backed by expanded test coverage. These efforts deliver richer, layout-aware predictions, improved data extraction robustness, and higher downstream value for customers relying on Docling's structured outputs.