Womblex

Document extraction pipeline for converting Australian government documents into ML-friendly corpus or collections. Extracts text from PDFs and Word documents (native, scanned, forms, hybrid). Spreadsheets are ingested as cell-grained element streams with automatic header/preamble detection, ready for per-record semantic analysis. Reference registers (G-NAF, ABN bulk extract, geospatial) have standalone Parquet ingests that bypass the NLP pipeline.

Design disclosure

This project is designed for everyone. All design decisions favour air-gapped edge deployment, running on limited resources. This means Womblex doesn't include many of the more robust 'all in one' OCR models.

Mature OCR models are used to compete with Womblex for evaluations and guide development.

Add-ons/integrations

Optionally outputs are prepared for semantic analysis via Isaacus.

The Problem

Government document releases arrive as a mix of file formats:

• PDFs — native (selectable text), scanned (narrative, forms, tables), hybrid, or redacted
• Word documents (.docx) — paragraphs and embedded tables
• Spreadsheets (.csv, .xlsx, .xls) — row-level data, glossaries, key-value lookups, and narrative sheets

One-size-fits-all OCR fails because each format and sub-type needs a different extraction strategy. Womblex detects the document type first, then routes to the right extractor.

Installation

pip install womblex

With Isaacus enrichment:

pip install womblex[isaacus]

For development:

git clone https://github.com/DeepCivic/womblex.git
cd womblex
uv sync --extra dev

A minimal test-fixture set is vendored in this repo (fixtures/fixtures/), so a fresh clone runs most of the suite with no extra setup. The full benchmark set lives in a separate repository — see THIRD_PARTY_DATA.md for how to obtain it.

System Dependencies

No system-level dependencies beyond Python. All extraction backends are pure Python packages:

• PyMuPDF (fitz) — native PDF text and structure
• PaddleOCR (rapidocr-onnxruntime) — scanned-page OCR with layout analysis (no Tesseract or PaddlePaddle required)
• python-docx — Word document extraction
• pandas + openpyxl — spreadsheet ingestion (CSV/Excel)

Once you have extraction working, semantic analysis via Isaacus (embeddings, classification, extractive QA) is straightforward.

Isaacus API Key (optional)

Required only for the enrichment stage (pip install womblex[isaacus]). Text extraction works without it.

cp .env.example .env
# Edit .env and add your key from https://isaacus.com/

Or export directly:

export ISAACUS_API_KEY="your-key-here"

Quick Start

# Process a document set using a config (E2E composition)
womblex run --config configs/example.yaml

# Resume from checkpoint after interruption
womblex run --config configs/example.yaml --resume

# Process individual files (PDF, DOCX, CSV, Excel)
womblex extract document.pdf -o output/
womblex extract report.docx -o output/
womblex extract dataset.xlsx -o output/

# Per-stage commands (primary workflow for staged corpora): each consumes the
# prior stage's shard directory and writes its own sidecar in place, with an
# independent resumable CheckpointManager.
womblex normalise --shards output/<run_id>/documents/               # *.normalised_text.parquet (offline text cleanup)
womblex spellfix  --shards output/<run_id>/documents/               # *.spellfix_text.parquet + *.spellfix_corrections.parquet (offline OCR repair)
womblex chunk     --shards output/<run_id>/documents/               # *.chunks.parquet
womblex quality   --shards output/<run_id>/documents/               # *.chunk_quality.parquet (offline chunk annotation)
womblex redact    --shards output/<run_id>/documents/ --pdfs <dir>  # *.redactions.parquet
womblex enrich    --shards output/<run_id>/documents/               # *.enrichment_entities.parquet (Kanon-2; needs ISAACUS_API_KEY)
womblex link      --shards output/<run_id>/documents/ --config <yaml> # *.entity_links.parquet (register match)
womblex embed     --shards output/<run_id>/documents/               # *.embeddings.parquet (Kanon-2 chunk embeddings)
womblex pii       --shards output/<run_id>/documents/               # *.pii_spans.parquet (audit) + *.clean_text.parquet (masked, terminal)

# Standalone register ingests (bypass the NLP pipeline, write Parquet directly)
womblex ingest-gnaf "G-NAF/G-NAF FEBRUARY 2026" -o output/gnaf   # G-NAF PSV → Parquet
womblex ingest-abn  extracts/ -o output/abn                      # ABN bulk extract XML → records + names Parquet
womblex ingest-geo  shapefiles/ -o output/geo                    # SHP → GeoParquet

# Audit shard integrity (extraction stage)
womblex verify-shards output/<run_id>/

Distributed / cloud execution (optional)

The local commands above run single-process and CPU-first — the default, and all that's needed for air-gapped edge use. For large corpora where you don't want to wait hours, the cloud extra adds horizontal scale-out without changing extraction behaviour: the same per-batch body runs, just across many workers that share a Postgres job queue and an object store.

pip install womblex[cloud]   # fsspec + s3fs + psycopg3

# 1. Plan: list source docs in object storage, split into batches, enqueue.
#    Idempotent on (run_id, batch_num) — re-run to resume.
womblex enqueue --store s3://womblex --input-prefix inputs/demo \
    --config configs/example.yaml --create-schema

# 2. Process: run as many workers as you like (separate hosts/containers).
#    Each claims batches via FOR UPDATE SKIP LOCKED — no double-processing.
womblex worker --store s3://womblex --config configs/example.yaml \
    --stale-timeout 900            # requeue batches orphaned by crashed workers

# 3. Watch progress.
womblex jobs --run-id <run_id>     # pending/running/done/failed counts

# 4. Finalise once the fleet drains: consolidate the per-batch shard manifests
#    into <store>/runs/<run_id>/manifest.parquet (the local `run` does this at
#    its end; a distributed run has no single end, so it's an explicit step).
womblex finalize --store s3://womblex --run-id <run_id>

Connection details come from --store/WOMBLEX_STORE_URI, --dsn/WOMBLEX_DB_DSN (or DATABASE_URL), and the standard AWS_* / WOMBLEX_S3_ENDPOINT env vars (MinIO works as an S3 endpoint). Shards land at <store>/runs/<run_id>/documents/ in the ordinary layout, so once synced down, womblex manifest / chunk --shards / every per-stage command consume a distributed run exactly like a local one.

A ready-to-run stack (Postgres + MinIO + scalable workers) lives in docker-compose.yml:

docker compose up -d postgres minio createbuckets init
docker compose run --rm womblex enqueue --input-prefix inputs/demo \
    --config configs/example.yaml --create-schema
docker compose up --scale worker=4 worker

How It Works

1. Per-page profiling + plan-driven orchestrator

PDFs are profiled per-page (PageProfile per page) rather than at document level. The orchestrator dispatches operations page-by-page based on the profiles, then merges into a single ExtractionResult. A doc-level summary type still surfaces in metadata.

Per-page operations the orchestrator can apply:

Doc-level shape detection (informs the orchestrator):

Other formats — routed by file extension:

2. Extraction

Each document type routes to an appropriate extractor. extract_text() always returns a list[ExtractionResult]:

• PDFs return a single-element list. The per-page orchestrator dispatches _apply_native_page or _apply_ocr_page based on each page's PageProfile. PaddleOCR returns per-region confidence scores stored in the document profile. YOLO layout analysis (DocLayNet yolo11n_doc_layout.pt, with COCO yolov8n.pt as fallback) is called on OCR pages by _layout_blocks_and_tables to populate Element.kind for the layout regions it detects; a full-page scan whose dominant region is a figure but which OCR's to substantial text is tagged paragraph rather than figure so its content reaches chunking.
• DOCX returns a single-element list with paragraphs and tables interleaved in OOXML body order.
• Spreadsheets return one ExtractionResult per workbook. Each sheet contributes a leading kind='sheet_meta' element followed by one kind='sheet_cell' element per non-empty cell. Export products that open with title rows or key: value metadata blocks above the real header (e.g. AusTender contract-notice exports) are handled: the header is detected by run-scoring (the candidate row starting the longest run of table-consistent rows below it), preamble rows land verbatim on sheet_meta.meta["preamble"], and row 0 of the cell grid is always the real header. Ragged CSVs (a one-field title row above a wide header) parse rather than fail.

Each result carries a document_id used as the primary key downstream.

Text at the extraction boundary is verbatim — _normalise_text no longer runs in the extraction hot path. Whatever the producing extractor (native text layer, PaddleOCR, DOCX, spreadsheet-print, …) emits is what lands on the element's text field. Downstream stages (PII, redaction, chunking) may rewrite pages[i].text, but the parquet writer serialises elements, so on-disk content stays extraction-time verbatim. Cleanup (font-encoding artefacts, running OCR footers, OCR character-confusions) belongs to downstream offline stages — womblex normalise writes a *.normalised_text.parquet overlay and womblex spellfix writes a *.spellfix_text.parquet overlay, both leaving the verbatim elements untouched. See docs/extraction.md.

3. Redaction

Redaction runs as a post-extraction stage, separate from extraction. This avoids false positives that occur when running redaction detection inside OCR (form fields, chart regions, and diagram fills trigger the detector).

Redacted regions can be replaced with <REDACTED> markers (preserving sentence structure) or deleted entirely. The stage is configurable: apply after chunking, after enrichment, or both.

4. Chunking

Extracted text is split into semantically meaningful chunks using semchunk with the Kanon tokeniser (default 480 tokens, leaving 32-token headroom for Isaacus 512-token context windows). Tables are converted to markdown and chunked separately, with each chunk tagged as "narrative" or "table". <REDACTED> markers are preserved across chunk boundaries.

Chunking has two invocation modes that share one engine (chunk_batch):

• Per-stage: womblex chunk --shards <run_dir>/documents/ consumes the extraction-stage shards directly and writes *.chunks.parquet siblings. Independent CheckpointManager so the chunk stage resumes without re-extracting. This is the primary workflow for staged corpus runs.
• E2E composition: womblex run --config <yaml> extracts and chunks in one process (kept for users with simpler corpora).

Both modes reassemble narrative + tables from each source's element stream, then feed every doc's narratives into a single semchunk call (with overlap) and every doc's table markdowns into another (no overlap), so processes parallelises across the whole batch. Chunks carry (start_char, end_char, page_start, page_end, has_redaction, content_type); they join back to elements via source_hash plus offset-range overlap.

AI chunking (optional). Setting chunking.chunking_model (e.g. kanon-2-enricher) switches narrative chunking to semchunk 4's AI chunking — boundaries follow the Isaacus enricher's document structure instead of the offline token split. Off by default, so non-Kanon setups are unaffected. When the enrich stage also runs, enrich it once: run womblex enrich before womblex chunk, and enrich persists the graph (*.enrichment_doc.parquet) for chunk to reuse instead of enriching twice. A byte-identity guard ensures reuse only happens when the persisted text matches the chunk source; otherwise it self-enriches.

5. PII Cleaning

An optional PII stage masks personal identifiers in chunk text. It is graph-driven: the primary candidates are PII-typed entities from the Kanon-2 enrichment graph (natural→PERSON, address→ADDRESS), mapped onto chunks via mention offsets — so PII runs after enrichment, not before. Recall is flexed by enrichment granularity, not by a separate detector.

A local regex + cosine-context backstop (PERSON via all-MiniLM-L6-v2, ADDRESS via street-type regex) exists but is opt-in and off by default (pii.use_regex_backstop = false): on this corpus it is low-precision (~15% — orgs and headings get tagged PERSON), so it is reserved for recall experiments. The all-MiniLM-L6-v2 model is pre-bundled in models/ and loaded from disk — no network access at runtime.

Masking is terminal. The stage writes two siblings and never rewrites the raw chunks that feed Isaacus:

• *.pii_spans.parquet — one row per detected span (audit/reversible), carrying the graph entity_id and its <PERSON_n> replacement.
• *.clean_text.parquet — the masked, publishable text layer (<PERSON_1>, <ADDRESS_1>, … — typed and numbered off the graph entity), written by default (pii.write_clean_text = true).

See docs/accuracy/PII_CLEANING.md for the measured baseline and docs/decisions.md for why masking is terminal.

6. Embeddings and Enrichment

Clean chunks feed into Isaacus models:

• kanon-2-embedder: Semantic embeddings for search/retrieval
• kanon-universal-classifier: Zero-shot document classification
• kanon-answer-extractor: Structured field extraction (dates, names, references)

Graph construction

Using Isaacus outputs an entity graph can be created for further analysis.

Configuration

Configs are YAML files defining paths, detection thresholds, and analysis settings:

dataset:
  name: my_dataset

paths:
  input_root: ./data/raw/my_dataset
  output_root: ./data/processed/my_dataset
  checkpoint_dir: ./data/checkpoints/my_dataset

detection:
  min_text_coverage: 0.3
  form_signal_threshold: 0.5
  table_signal_threshold: 0.4

extraction:
  ocr:
    engine: paddleocr
    dpi: 200

chunking:
  tokenizer: "isaacus/kanon-2-tokenizer"
  chunk_size: 480
  enabled: true
  chunk_tables: true

processing:
  batch_size: 25
  checkpoint_every: 25

See configs/example.yaml for a complete example.

Output

Each batch writes four sibling Parquet shards. The shard base name is the caller's choice (e.g. batch-0001):

batch-NNNN.elements.parquet — one row per structural element (paragraph, heading, table, form, image, sheet cell, …). Canonical output.

batch-NNNN.table_cells.parquet — children of kind='table' elements, one row per cell. Joins back via (source_hash, parent_elem_order).

batch-NNNN.form_fields.parquet — children of kind='form' elements, one row per field. Same join key.

batch-NNNN._manifest.parquet — one row per source file with provenance, status, and element / cell / field counts.

See docs/extraction.md for the canonical schema reference, element kinds, the reassembly query, and the verbatim-text policy.

With womblex[isaacus] enrichment enabled, the per-stage womblex enrich --shards writes sidecars alongside each batch:

batch-NNNN.enrichment_entities.parquet — flat entity mentions for filtering / PII candidates

batch-NNNN.enrichment_meta.parquet — document-level enrichment metadata

batch-NNNN.enrichment_doc.parquet — (only with enrichment.persist_document, auto-enabled when AI chunking is on) the raw ILGS Document per doc, reused by the chunk stage for AI chunking

The E2E graph path (womblex run) additionally emits entities.parquet and graph_edges.parquet for graph queries.

Project Structure

A file-level map of the source tree lives in docs/project-structure.md. At a glance:

womblex/
├── configs/           # Dataset-specific configurations
├── docs/              # Architecture docs, ADRs, accuracy reports
├── fixtures/          # Test fixtures (separate repo, see THIRD_PARTY_DATA.md)
├── src/womblex/
│   ├── cli/           # CLI subpackage — per-topic command modules
│   ├── operations/    # Independent operations (extract/redact/chunk/pii/enrich)
│   ├── ingest/        # Detection, per-page profiling, PDF/non-PDF extraction
│   ├── redact/        # Redaction detection + post-extraction stage
│   ├── pii/           # Graph-driven PII detection + terminal masking
│   ├── process/       # Chunking + offline text stages (normalise/spellfix/quality)
│   ├── link/          # Record linkage to reference registers
│   ├── analyse/       # Isaacus enrichment + embeddings + entity graph
│   ├── store/         # Parquet schemas, sidecar IO, checkpoints, retention
│   ├── utils/         # Metrics + local model path resolution
│   └── verify/        # Two-pass extraction quality verification
└── tests/

See docs/project-structure.md for the full per-module breakdown.

Development

# Install with dev dependencies (pytest, ruff, mypy live in the extras)
uv sync --all-extras

# A minimal fixture set is vendored; the full benchmark set is optional —
# see THIRD_PARTY_DATA.md.

# Run the suite (no addopts filter — runs everything; heavy tests skip on a
# bare checkout). Use -m "not slow and not benchmark" for the fast subset.
uv run python -m pytest

# Run OCR and accuracy benchmarks (need the full fixtures; minutes-long)
uv run python -m pytest tests/test_fixture_accuracy.py tests/test_womblex_collection_accuracy.py -v

# Type checking
uv run mypy src/

# Lint
uv run ruff check src/

Accuracy docs (docs/accuracy/*.md) are regenerated automatically at the end of each test run — no manual editing needed.

License

Apache 2.0

Acknowledgements

• Isaacus for legal AI models
• semchunk for semantic chunking
• PyMuPDF for PDF handling
• RapidOCR for OCR (bundles PaddleOCR v4 ONNX models, no PaddlePaddle required)
• Ultralytics for YOLOv8 layout analysis
• python-docx for Word document extraction
• pandas + openpyxl for spreadsheet ingestion