PERLA Extract

PERLA Extract turns a photovoltaic paper and its supplementary information into device-centered, source-linked data. It keeps device families, individual cells, performance observations, population summaries, and stability experiments distinct.

pip install perla-extract
export OPENAI_API_KEY="your-openai-key"

perla-extract \
  --pdf paper.pdf \
  --supplement paper_si.pdf \
  --model openai/gpt-5.2 \
  --max-cost-usd 2.00 \
  --output-dir results/paper

Model calls use LiteLLM. The provider-prefixed model name selects the backend and its standard credential—for example, openai/... with OPENAI_API_KEY, openrouter/... with OPENROUTER_API_KEY, or anthropic/... with ANTHROPIC_API_KEY. The default calls OpenAI directly; the extraction workflow itself remains provider-neutral.

The extraction directory contains the refined rich result, its retained first draft, a source-grounded ledger of experimental objects and atomic claims, a claim-coverage audit, a bounded repair pass for audit-visible gaps, local evidence checks, and audited conservative finalization of unsupported optional claims, and audited composition/processing enrichment. Every finalization preserves the complete candidate and exact removed content. Accepted interpretations feed atomic archives for the pinned NOMAD schema without rewriting reported facts in extraction.json. Use --dry-run to parse the documents and estimate call size without calling a model. Every model request contains parser-produced text and tables only; the workflow does not send rendered pages to a vision model.

Documentation

The documentation site explains:

• the study and evidence model;
• single-call and long-supplement extraction;
• human ground-truth review;
• deterministic rich-schema evaluation;
• audited composition and processing interpretation;
• direct NOMAD export;
• paper discovery and Zotero journal-club intake; and
• the optional reduced-schema compatibility boundary.

Build it locally with:

pip install -e '.[docs]'
mkdocs serve

See CONTRIBUTING.md for development and test commands.

Discover new papers

PapersBot combines packaged journal feeds, OpenAlex topic searches, and an optional Zotero group; applies one replaceable JSON selection policy; retrieves available PDFs; and records incremental state as readable JSON. It is deliberately separate from scientific data extraction:

pip install 'perla-extract[papersbot]'
perla-papersbot downloaded_papers --state-dir .papersbot-state

Use --feeds-file or repeated --feed options to replace the journal list and --selection-file to replace the relevance policy. --zotero-group-id also ingests a Zotero group, including stored PDF attachments. A configured collection can be a journal-club-curated intake queue. Zotero access is read-only and works with a key scoped to that group. The supported unattended deployment is an internal cron job, with ordinary settings available through PAPERSBOT_ environment variables. Each PDF acquisition records its source and access basis; authorized local retrieval mechanisms can implement the small PdfSource interface without adding publisher cases to the bot.

Authors

• Sherjeel Shabih — sherjeel.shabih@hu-berlin.de
• Pepe Marquez — jose.marquez@physik.hu-berlin.de
• Kevin Jablonka — mail@kjablonka.com
• Sharat Patil — sharat.patil@physik.hu-berlin.de

Citation information will be added with the first public release of this workflow.