acatome-extract

PDF extraction and enrichment pipeline for scientific papers. Converts PDFs into structured, searchable bundles with block-level summaries and embeddings.

Features

• Marker PDF extraction — structured block extraction with headings, tables, figures
• Fitz fallback — recursive character chunking when Marker is unavailable
• LLM enrichment — block and paper summaries via Ollama or litellm
• Embeddings — sentence-transformer embeddings for semantic search
• File watcher — acatome-extract watch monitors an inbox folder
• Bundle format — .acatome companion files for sharing pre-built extractions
• CLI — acatome-extract command for extract, enrich, and watch workflows

Installation

uv pip install -e .

On macOS/Linux this includes Marker for structured PDF extraction. On Windows it installs with the lighter pymupdf (fitz) backend by default. To add Marker on Windows (requires C build tools):

uv pip install -e ".[marker]"

With GPU acceleration (embeddings + torch):

uv pip install -e ".[gpu]"

Everything at once:

uv pip install -e ".[full]"

Usage

from acatome_extract.pipeline import extract

bundle = extract("/path/to/paper.pdf")

CLI

# Extract (RAKE summaries included automatically, no LLM needed)
acatome-extract extract paper.pdf
acatome-extract extract --type datasheet TI_LM317.pdf   # non-article types

# Enrich — embeddings only by default; add --summarize for LLM summaries
acatome-extract enrich /path/to/bundle
acatome-extract enrich --summarize /path/to/bundle       # enable LLM summaries
acatome-extract enrich --summarize --skip-existing dir/   # incremental LLM pass

# Watch — extract + embed + ingest; LLM summaries off by default
acatome-extract watch ~/papers/inbox
acatome-extract watch ~/papers/inbox --summarize          # enable LLM summaries

# Migrate old bundles to new summaries dict format + add RAKE
acatome-extract migrate ~/.acatome/papers
acatome-extract migrate ~/.acatome/papers --dry-run       # preview changes

# Supplements
acatome-extract attach parent-slug supplement.pdf --name s1

Summaries

Extraction always generates RAKE (extractive keyword) summaries — instant, no LLM required. LLM-based summaries are opt-in via --summarize and require an Ollama or litellm-compatible model.

RAKE summaries are used as the default for search and display. To add LLM summaries later:

acatome-extract enrich --summarize --skip-existing ~/.acatome/papers

Sidecar metadata

Place a <stem>.meta.json alongside any PDF to override metadata:

{"type": "datasheet", "title": "LM317 Regulator", "author": "Texas Instruments", "year": 2022}

Supported fields: type, title, author (string or list), year, doi, abstract, journal, s2_id, arxiv_id, verified.

Use explicit null to clear a field that the upstream Crossref/S2 lookup got wrong — empty strings are ignored for backward compatibility:

{"doi": "10.1002/9781118519301.ch5", "s2_id": null, "verified": true}

When "verified": true is set, the fuzzy-title verification gate is bypassed — useful for PDFs where the real title doesn't appear on page 1 (Elsevier header strips, ACS internal production PDFs, abstract collections).

Dependencies

• acatome-meta — metadata lookup and verification
• marker-pdf — structured PDF extraction
• litellm / Ollama — LLM-based enrichment

Testing

uv run python -m pytest tests/ -v

License

GPL-3.0-or-later — see LICENSE.