Skip to content

Releases: saibal-roy/docling-batch-extract

Release list

v0.1.0

v0.1.0 Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 08 Oct 20:11
v0.1.0
8c08563

First public release: a batch PDF → Markdown extractor for RAG on one CPU-only 2 vCPU / 8 GB server running Ubuntu 26.04 LTS. Plans: v1–v7.

Added

  • extract.py: converts inputs/*.pdf with docling-serve to outputs/<name>.json, then moves each PDF to completed/, or to errors/ on failure. Per-document logs in logs/<name>.log. Exit codes: 0 all succeeded, 1 some PDFs failed, 2 server down.
  • Parallel page-range conversion for text PDFs (--chunk-pages 5); scanned PDFs are detected automatically, converted whole and OCR'd (--ocr auto).
  • Memory-aware dispatch (--mem-high 0.75), container-restart recovery, and "busy is not down" polling.
  • pdf_backend=pypdfium2 by default. docling's default PDF reader ran out of memory on JBIG2 scans.
  • Every output JSON carries extractor_version and engine (docling-serve and docling versions, PDF reader). extract.py --version prints the version.
  • docker-compose.yml: CPU-only docling-serve-cpu:v1.36.0 (latest stable), published on 127.0.0.1 only. Sizing comes from .env (written by setup_ubuntu.sh from the machine's CPUs and RAM): 1 worker × 2 threads, ~6.5 GB limit on the 2 vCPU / 8 GB target. extract.py reads the worker count from the container.
  • Scripts: setup_ubuntu.sh (server setup), demo_run.sh (first run + benchmark, with a CC BY 4.0 demo PDF), cleanup.sh, preview_site.sh and check_github_pages.sh (documentation site).
  • Tests: 18 acceptance checks, benchmarks, tests/profile_benchmark.sh (size any machine), and the go-ahead gate tests/ubuntu_container_test.sh: Ubuntu 26.04 LTS on a simulated 2 vCPU / 8 GB machine.
  • CI (lint, data guard, setup + acceptance on Ubuntu 26.04 LTS), a Pages workflow (documentation site), and a Release workflow. All dependencies pinned to their latest stable releases (requests 2.34.2, pypdf 6.19.0, Actions checkout@v7 …).
  • Documentation: README, STORY.md, prompts.md, design history, and a documentation site built with MkDocs.

Known trade-off

  • docling-serve-cpu v1.36.0 converts scanned PDFs ~20–25 % faster but text PDFs ~45–80 % slower than v1.30.0 at 2 vCPU (tests/results/2026-10-08_profiles/). Chosen for scan-heavy legal and claims batches. Re-check when upgrading docling.

Security

  • docling-serve has no authentication. It's bound to loopback, and acceptance check A25 and demo_run.sh fail if it's published beyond loopback. The README's AWS guidance: never open port 5001 in a security group.
  • SECURITY.md: report vulnerabilities privately through GitHub (private vulnerability reporting is enabled); never attach real documents.
Assets 2
Loading