paperlm

Scientific PDFs → Markdown, built for LLMs. A MarkItDown plugin that replaces the built-in PDF converter with Docling-powered layout analysis and PaddleOCR for scanned documents.

Status: ✅ v0.1.0 released on PyPI. main is preparing v0.1.1 with operational tooling such as paperlm-tools warmup.

中文版说明见 README_zh.md。

Why

MarkItDown's built-in PDF path is optimized for invoices and form-style documents; its own source code acknowledges it is "not [designed] for multi-column text layouts in scientific documents". This plugin fixes that for research use-cases.

Before / After, case 1 — arXiv table-heavy paper

On arXiv:2505.11545 (the TARGET benchmark paper), MarkItDown's built-in heuristics mistake the title block for a GitHub-flavored table and produce this garbage:

|     | TARGET: |     | Benchmarking |     |     | Table | Retrieval | for Generative |     | Tasks |
| --- | ------- | --- | ------------ | --- | --- | ----- | --------- | -------------- | --- | ----- |
XingyuJi*1,ParkerGlenn2,AdityaG.Parameswaran1,MadelonHulsebos3
|     |     |     |          | 1UCBerkeley |     |     | 2CapitalOne          | 3CWI |         |
| --- | --- | --- | -------- | ----------- | --- | --- | -------------------- | ---- | ------- |
5202 yaM 41  ]RI.sc[  1v54511.5052:viXra Thedatalandscapeisrichwithstructureddata,

With paperlm enabled, the same page renders as:

## TARGET: Benchmarking Table Retrieval for Generative Tasks

> * Correspondence to madelon.hulsebos@cwi.nl and jixy2012@berkeley.edu

Large Language Models (LLMs) have become an indispensable tool in the
knowledge worker's arsenal, providing a treasure trove of information at
one's fingertips. Retrieval-Augmented Generation (RAG) (Lewis et al., 2020)
further extends the capabilities of these LLMs by grounding generic dialog...

## 1 Introduction

The data landscape is rich with structured data, often of high value to
organizations, driving important applications in data analysis and
machine learning. Recent progress in representation learning and
generative models for such data has led to the development of natural
language interfaces to structured data, including those leveraging
text-to-SQL...

The garbage tables are gone, the vertical arXiv watermark (5202 yaM 41 ]RI.sc[ = 51 May 2025 [cs.IR] read bottom-up) is stripped, paragraphs reflow, and the title picks up a proper ## heading.

Before / After, case 2 — 27-page bioRxiv benchmark

	MarkItDown baseline	paperlm
First output	bioRxiv preprint doi:... watermark	## A systematic benchmark... (H2)
Authors	AliHamraoui1,2,AudreyOnfroy3,... glued	Ali Hamraoui 1,2 , Audrey Onfroy 3 , ...
Special chars	(cid:0) garbage	/a0 readable
Paragraph structure	hard-wrapped soup	clean paragraphs

Reproduce both side-by-sides with:

python tests/fixtures/fetch.py          # build the corpus
python benchmarks/w4d4_showcase.py      # write outputs to benchmarks/outputs/
diff benchmarks/outputs/sample_arxiv_table_heavy__{baseline,paperlm}.md

Known limit — bilingual documents

Papers that put a Chinese title/abstract right next to the English title/abstract on the same page (as the 《生命科学》 fixture does) still surface with interleaved ordering; Docling's layout model sometimes reorders these blocks. We will revisit in a later pass.

Install

# From PyPI
pip install paperlm[docling]            # default — Docling engine
pip install paperlm[docling,ocr]        # + PaddleOCR for scanned
pip install paperlm[all]                # all safe-license extras

# From source (contributor flow)
git clone https://github.com/youngfly93/paperlm
cd paperlm
pip install -e ".[docling,dev]"

Usage

from markitdown import MarkItDown

md = MarkItDown(enable_plugins=True)
result = md.convert("paper.pdf")

# Stable MarkItDown API
print(result.markdown)

# paperlm extensions (non-stable API — Python only, not exposed to CLI/MCP)
print(result.engine_used)       # "docling" / "paddleocr" / "pdfminer" / "failed"
print(len(result.ir.blocks))    # structured IR for downstream RAG/Agent use
print(result.ir.warnings)       # degradation trail
print(result.ir.metadata)       # OCR confidence, adapter metadata, etc.
print(result.paperlm_json)      # full IR JSON sidecar
print(result.paperlm_chunks_jsonl)  # one text-bearing block per JSONL line

Force a specific engine

from markitdown import MarkItDown

md = MarkItDown(enable_plugins=True, paperlm_engine="ocr")        # force OCR
md = MarkItDown(enable_plugins=True, paperlm_engine="docling")    # skip scanned-check
md = MarkItDown(enable_plugins=True, paperlm_engine="fallback")   # pdfminer only
md = MarkItDown(enable_plugins=True, paperlm_enable_ocr=False)    # disable OCR in auto mode
md = MarkItDown(enable_plugins=True, paperlm_enable_formula=True) # opt-in formula LaTeX extraction

Memory notes

The OCR path uses PP-OCRv5 mobile models by default: ~2.7 GB peak RSS, ~15 s / page on CPU, zero measurable quality loss vs server models on our test corpus (see benchmarks/phase3_perf.md). OCR conversions also populate page-level confidence under result.ir.metadata["ocr"] and add warnings for empty or low-confidence pages.

For worker machines under 4 GB RAM:

from paperlm.engines.ocr_adapter import OCRAdapter

adapter = OCRAdapter(low_memory=True)   # ~2.0 GB peak, slightly slower
# Or: OCRAdapter(variant="server") for ~10 GB peak — opt-in only

For now these flags are exposed only on the Python adapter class; hook them up via a custom plugin kwargs integration if you need to drive them from the CLI. Direct MarkItDown(...) kwarg support is planned in a later release.

Batch/server worker pool

For batch ingestion services, avoid loading Docling/PaddleOCR in the API process. DoclingWorkerPool keeps one MarkItDown instance alive per worker subprocess, applies per-task timeout/RSS limits, and restarts only the offending worker on failure.

from paperlm.workers import DoclingWorkerPool

with DoclingWorkerPool(num_workers=2, timeout_s=300, max_rss_mb_hard=6144) as pool:
    results = pool.convert_many(["paper_a.pdf", "paper_b.pdf"])

for result in results:
    print(result.status, result.engine_used, len(result.markdown))

This is a Python API for service deployments, not a replacement for the default markitdown --use-plugins paper.pdf CLI path.

CLI

markitdown --use-plugins paper.pdf -o paper.md
markitdown --list-plugins        # should show "paperlm"

For batch/server ingestion, use paperlm-batch so Docling is loaded in reusable guarded worker subprocesses instead of once per document:

paperlm-batch paper_a.pdf paper_b.pdf \
  --workers 1 \
  --timeout-s 300 \
  --max-rss-mb-hard 6144 \
  --output-dir outputs/ \
  --output-jsonl outputs/results.jsonl

--output-dir writes *.md, *.paperlm.json, and *.chunks.jsonl artifacts. The JSONL result stream stays compact by default; add --include-markdown only if you want full Markdown embedded in each row.

JSONL input is supported for queues and ETL jobs:

{"id": "doc-1", "pdf_path": "paper_a.pdf"}
{"id": "doc-2", "pdf_path": "paper_b.pdf"}

paperlm-batch --input-jsonl jobs.jsonl --output-jsonl results.jsonl --output-dir outputs/

Operational tools

Use paperlm-tools warmup during deployment to pre-download and initialize optional engines before production traffic hits the first conversion. Warmup runs in guarded worker subprocesses, so timeout/RSS limits kill only the warmup worker instead of the parent process.

paperlm-tools warmup --engine docling
paperlm-tools warmup --engine docling,ocr --timeout-s 600 --max-rss-mb-hard 6144

The command fails if the requested engine silently falls back to another engine. For example, --engine docling must finish with engine_used=docling; otherwise install paperlm[docling] or check the runtime environment.

How it works

Input PDF
  │
  ▼
EngineRouter (auto mode)
  │
  ├─ sample text layer (scanned_detector.py)
  │
  ├─ if scanned (no text)  →  OCRAdapter (PaddleOCR, Apache-2.0)
  │                            ↓ on empty/fail
  │                            DoclingAdapter (MIT)
  │                            ↓ on empty/fail
  │                            FallbackAdapter (pdfminer)
  │
  └─ if has text layer    →  DoclingAdapter
                              ↓ on empty/fail
                              OCRAdapter (if installed)
                              ↓ on empty/fail
                              FallbackAdapter (pdfminer — shipped as core dep)
  │
  ▼
IR (Block / BlockType / BBox / reading_order)
  │
  ▼
MarkdownSerializer → result.markdown

The router is fail-safe as long as the package is properly installed: a real pip install -e . pulls pdfminer.six and pdfplumber as core deps, so the FallbackAdapter is always attempted last. If you are running tests with PYTHONPATH=src but without actually installing the package (a common reviewer flow), pdfminer.six may be missing — in that case the router returns an IR with engine_used="failed" and result.ir.warnings explaining what happened. No exception propagates to the caller either way.

License matrix for optional engines

Extra	Engine	License	Commercial use
(core)	pdfminer.six / pdfplumber	MIT / MIT	✅ safe
[docling] (recommended default)	Docling 2.90	MIT	✅ safe
[ocr]	PaddleOCR + paddlepaddle	Apache-2.0	✅ safe
[formula] (opt-in)	pix2tex	MIT	✅ safe
[marker] (optional)	Marker	GPL-3 + OpenRAIL-M	⚠️ copyleft; requires Datalab license above $2 M revenue
[mineru] (optional)	MinerU	AGPL	⚠️ strong copyleft; triggers on network distribution

Main package is Apache-2.0 and never transitively pulls GPL/AGPL dependencies unless you opt in.

Development

uv venv --python 3.12 ~/.venvs/paperlm
source ~/.venvs/paperlm/bin/activate
export UV_LINK_MODE=copy          # needed when src is on exFAT
uv pip install -e ".[docling,ocr,dev]"

# Fast tests (no ML model loads)
make test-fast

# Slow integration tests (require Docling / PaddleOCR model downloads
# on first run).
pytest tests/test_docling_adapter.py       # needs pip install -e '.[docling]'
pytest tests/test_ocr_adapter.py           # needs pip install -e '.[ocr]'
pytest tests/test_pdf_converter_e2e.py     # needs pdfminer.six + docling

Memory note: do not run Docling + PaddleOCR tests in the same pytest process — together they hold ~3 GB of models. Run them sequentially.

The competitor benchmark is also guarded by default. make benchmark-compare runs only the smoke fixture profile and the core triad (MarkItDown, paperlm, Docling), with per-run timeouts, RSS warnings, and hard RSS limits that kill only the offending subprocess. Use the full/heavy matrix only when you intentionally want to load every installed parser:

make benchmark-compare       # safe smoke profile
make benchmark-compare-full  # opt-in full corpus + Marker/MinerU runners
make benchmark-long-perf     # one long math-heavy PDF, with paperlm timing breakdown
make benchmark-long-profile  # cProfile top 30 for paperlm long-PDF breakdown
make benchmark-worker-pool   # compare fresh subprocesses vs one pooled worker

Reproducing integration evidence without installing every extra

The fast test suite deliberately skips anything that needs an ML model or a missing core dep. That is expected behaviour — a skipped test's green state is not a claim that the underlying optional engine works.

Three progressively stronger ways to verify the integration paths:

1. Trust pre-recorded evidence — benchmarks/phase4_integration.md records an 8-fixture sweep on macOS-CPU, one subprocess per fixture. All eight converge to non-empty IRs, zero errors, every peak RSS ≤ 4 GB.

2. Run the GitHub Actions integration job on your fork:

   gh workflow run test.yml --field ref=<branch>
   

   It fetches fixtures, installs .[docling,ocr], and runs test_docling_adapter.py / test_ocr_adapter.py / test_pdf_converter_e2e.py on a fresh Linux runner.

3. Reproduce locally:

   python tests/fixtures/fetch.py                        # build the corpus
   pip install -e '.[docling,ocr,dev]'
   make test-all                                         # ~3-4 minutes
   python benchmarks/w4d5_integration_sweep.py           # per-fixture peak RSS
   

Test fixtures (not in git — regenerate locally)

The corpus is 8 PDFs covering different layouts, languages, and dense content (tables, formulas). Rebuild it with one command:

python tests/fixtures/fetch.py          # download + build everything
python tests/fixtures/fetch.py --check  # verify all present, exit 1 if missing

Fixture	Pages	Character
sample_en_two_col.pdf	27	EN double-column bioRxiv bioinfo benchmark
sample_zh_mixed.pdf	10	ZH double-column 《生命科学》2024 review
sample_arxiv_llm_survey.pdf	14	EN survey, moderate tables + figures
sample_arxiv_table_heavy.pdf	12	EN benchmark paper, 5+ tables in first 5p
sample_arxiv_math.pdf	86	EN long, math-heavy (DeepSeek-R1 report)
sample_arxiv_long_ir.pdf	60	EN long single-column survey
sample_scanned.pdf	5	Synthetic — ZH paper rasterized to images
sample_scanned_1p.pdf	1	Synthetic — single page for fast OCR tests

Status & roadmap

• ✅ Week 1 — skeleton + Docling integration + IR + Markdown serializer
• ✅ Week 2 — pdfminer fallback + EngineRouter + scanned detection + PaddleOCR
• ✅ Week 3 — formula inline/block detection, table polish, caption linking, reading-order repair
• ✅ Week 4 — coverage, CI, benchmark docs, performance/RSS guardrails
• 🚧 Week 5 — release hardening, TestPyPI/PyPI publication, launch docs

See ../PRD.md for the full spec.

License

Apache-2.0. See LICENSE.