Womblex

Document extraction pipeline for converting files 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.

Runs on your laptop, scales to a cluster

Local is the default and always works. pip install womblex gives you the entire pipeline — extraction, OCR, chunking, PII — running CPU-only against the local filesystem. No cloud account, no object store, no database, no API key, no network access at runtime (models are bundled or resolved from models/). A Chromebook is a supported deployment, not a degraded one.

Cloud is additive, not a different product. When one machine stops being enough, the same wheel runs behind object storage and a shared job queue, and you buy throughput by adding workers — --scale worker=8 is the whole operation. What does not change when you scale out:

• the extraction logic (womblex run and the cloud worker call a byte-identical process_batch body)
• the OCR engine (the bundled CPU one by default; the hosted Bedrock VLM engine ships in the base install too and is selected by config, not a separate install)
• the output layout (distributed runs land the ordinary shard layout, so every local --shards command consumes them unchanged)

Scaling out is not a lock-in: a distributed run's shards sync down and every local per-stage command (womblex manifest, chunk --shards, …) consumes them unchanged, or you run those stages in place with run-stage and never sync at all. See Environment-Agnostic Execution.

Design disclosure

This project is designed for everyone with a focus on inexpensive processing. 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

Files are in a mix of 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

Pick the row that matches where you are running it. The pipeline logic is identical in every row — the extras add reach (object storage, a shared queue, hosted APIs), never a different extraction path.

Enrichment/embeddings (Isaacus SDK — hosted API via ISAACUS_API_KEY or a private SageMaker deployment via ISAACUS_SAGEMAKER_ENDPOINTS), hosted VLM OCR (Mistral Pixtral Large via AWS Bedrock), and the AWS SDK (boto3) are core dependencies — they need no extra. Every real deployment uses them, and they are tiny next to the vision/ML stack already in the base install, so gating them behind extras only produced misconfiguration (a missing SDK read as "no API key"). They stay dormant until you configure them: no key, no endpoints, and no mistral-ocr engine selected means nothing calls out.

pip install womblex[local] is accepted and resolves to the plain base install — it exists so a deployment can state which mode it is, and so [local] and [cloud] read as a pair.

Two things worth being explicit about:

• Configured, not installed. Isaacus enrichment, the Bedrock VLM OCR engine, and boto3 are always present but never active until you turn them on — enrichment needs ISAACUS_API_KEY or ISAACUS_SAGEMAKER_ENDPOINTS, and the hosted VLM OCR engine only fires when extraction.ocr.engine: mistral-ocr is set. A default run stays fully local and CPU-only.
• Local vs cloud is a runtime choice, not a build-time one. The same wheel does both; see Environment-Agnostic Execution. Installing [cloud] does not commit you to running in the cloud.

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/embedding stages. Text extraction works without it, and the Isaacus SDK is already installed (a core dependency).

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

Or export directly:

export ISAACUS_API_KEY="your-key-here"

Isaacus on Amazon SageMaker (private deployment)

Isaacus models can also run inside your own AWS account, fully air-gapped — no API key, no egress. Deploy the Marketplace package(s), then set ISAACUS_SAGEMAKER_ENDPOINTS instead of ISAACUS_API_KEY; every stage that calls Kanon-2 (chunk with AI chunking, enrich, embed) routes through the endpoints with no other change.

Subscriptions are per model plus a universal one, so declare what you actually deployed — comma-separated name[@region][=model|model|...], where an entry with no =models part serves every model:

export ISAACUS_SAGEMAKER_ENDPOINTS="kanon-2-universal-001"                        # one endpoint, all models
export ISAACUS_SAGEMAKER_ENDPOINTS="embed-001=kanon-2-embedder,enrich-001=kanon-2-enricher"  # per-feature
export ISAACUS_SAGEMAKER_ENDPOINTS="embed-001=kanon-2-embedder,universal-001"     # mixed: plus a catch-all

export ISAACUS_SAGEMAKER_REGION="ap-southeast-2"   # optional; else the AWS SDK default
export ISAACUS_SAGEMAKER_PROFILE="my-aws-profile"  # optional; else the AWS SDK default

A stage whose model no endpoint serves fails before its first request, naming the model and listing what the endpoints do serve. Chunk-size token counting is unaffected: the Kanon-2 tokeniser is vendored and stays local.

SageMaker credentials and the MinIO conflict. The /invocations calls are SigV4-signed by boto3's standard credential chain — on EC2 that resolves the instance role (e.g. one with a SageMakerInvokeEndpoint policy scoped to your endpoint), which is what you want. The trap: if the object store is MinIO, you may have set AWS_ACCESS_KEY_ID=minioadmin for s3fs. That env var is process-global, and boto3 checks env vars before the instance role — so it disables the role for the SageMaker signer too, which then signs minioadmin against real SageMaker and 403s (security token … is invalid). Symmetrically, setting the real AWS keys makes s3fs fail against MinIO (InvalidAccessKeyId). Give the object store its own credentials via WOMBLEX_S3_ACCESS_KEY_ID / WOMBLEX_S3_SECRET_ACCESS_KEY and leave AWS_ACCESS_KEY_ID unset — then s3fs authenticates to MinIO explicitly while SageMaker keeps the instance role. (ISAACUS_SAGEMAKER_PROFILE selects a non-default AWS profile if you need one for SageMaker specifically.)

Rotating the object-store key from the console. The WOMBLEX_S3_* keys above are defaults — often baked into the image. When one is rotated, an operator can save the new pair through the Resources Console (Run store card → S3 credentials) rather than rebuilding the container: the console persists it to its settings volume (--settings-dir / $WOMBLEX_UI_SETTINGS_DIR) and uses it over the env from the next request, and Test connection confirms it reaches the store. The pair overrides only the console's own store reads (and its enqueue/preflight against the ingest store); pipeline workers still read their keys from the env at their own start-up, so a rotated key that must reach the fleet is still an env/redeploy change there. The saved secret never appears in a response body or log — the card shows only its masked last-four and whether it came from saved or env. Because that settings volume now holds a credential, treat it as sensitive (mount it as you would any secret store); saving nothing keeps the env defaults, and Clear reverts to them.

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 money     --shards output/<run_id>/documents/               # *.money_spans.parquet + *.money_columns.parquet (offline amount 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>/

Environment-Agnostic Execution

The system scales from minimum hardware (e.g., a Chromebook) to distributed cloud clusters without altering extraction behavior. Configurable "knobs," such as parallel thread limits, allow you to optimize resource usage for your specific infrastructure.

You do not need any of this to use Womblex. Everything below is the scale-out path for when a single machine is the bottleneck; womblex run on a local directory remains fully supported and produces the same shards.

pip install womblex           # fsspec + s3fs + psycopg3 are already in the base install
pip install womblex[cloud]    # accepted, but empty — a marker that says "cloud deployment"

Selecting the backend

There is no STORAGE_TYPE / QUEUE_TYPE switch to set, because there is no branch for one to select. Both choices fall out of what you already pass:

RemoteStore.from_uri hands the URI to fsspec.core.url_to_fs, which returns a LocalFileSystem for a bare path or file:// and an S3FileSystem for s3:// (likewise gs://, az://). The staging code above it is one code path — a local --store runs the whole stage-in → process_batch → stage-out cycle with s3fs never imported. Credentials follow the same rule: the S3 store creds (WOMBLEX_S3_ACCESS_KEY_ID / WOMBLEX_S3_SECRET_ACCESS_KEY, falling back to AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY) and WOMBLEX_S3_ENDPOINT (for MinIO, falling back to AWS_ENDPOINT_URL) are read only for s3://, so a local store needs no configuration at all. Prefer the store-specific WOMBLEX_S3_* vars in the cloud: setting the ambient AWS_ACCESS_KEY_ID is process-global and would clobber the instance role boto3 uses to sign Isaacus-on-SageMaker /invocations calls (see the SageMaker section above).

Execution mode is the command, not a setting. womblex run processes batches in-process and checkpoints to CheckpointManager; enqueue/worker put the same process_batch body behind the Postgres queue, where the job row's status is the checkpoint. Both call byte-identical pipeline code, which is why a distributed run's output is the ordinary shard layout.

# 1. Plan: list source docs in object storage, split into batches, enqueue.
#    Idempotent on (run_id, batch_num) — re-run to resume. --ingest names a
#    location distinct from --store's runs/ output (assert_disjoint_locations
#    enforces this); the whole ingest root is the run's input, no prefix
#    needed. --input-prefix still works for a sub-folder of --ingest.
womblex enqueue --store s3://womblex --ingest s3://womblex/inbox \
    --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 --ingest s3://womblex/inbox \
    --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>

# 5. Run downstream stages in the store, without syncing the run down.
#    `run-stage` generalises finalize's shape to the per-batch sidecar stages:
#    one batch staged in at a time, all declared outputs published or none.
#    Idempotent — re-run as more batches land. Ordering is yours to pick.
womblex run-stage --stage normalise --store s3://womblex --run-id <run_id>
womblex run-stage --stage chunk --store s3://womblex --run-id <run_id> \
    --config configs/example.yaml
womblex run-stage --stage embed --store s3://womblex --run-id <run_id>

# 5b. …or hand the whole sequence to the workers instead of typing it.
#     Same stages, same contracts, ordering supplied from PIPELINE_ORDER and
#     gated by the config (a run that does not embed enqueues no embed job).
#     Deliberately a separate press from step 1: a worker will not start a
#     stage until the run's batches have settled, but a *bad* extraction
#     should not spend Isaacus budget on enrich and embed either.
womblex enqueue-stages --run-id <run_id> --config configs/example.yaml
womblex enqueue-stages --run-id <run_id> --config configs/example.yaml --dry-run

run-stage covers normalise, spellfix, chunk, money, enrich, embed, link, pii, graph-refresh and quality. manifest is deliberately absent — finalize already does it. Two stages are special: graph-refresh rewrites *.enrichment_entities.parquet / *.graph_edges.parquet in place, so it is never skipped by output existence and relies on its own idempotency; quality is run-scoped, staging every batch's chunks in one pass because its duplicate-cluster ids are corpus-wide. Pass --shards <dir> instead of --store/--run-id to run the same contract locally.

enqueue-stages writes one queue row per stage instead of running it here, so execution stays on the fleet with its retry and crash recovery, and a dispatcher (the CLI, or the console) only ever writes rows. Rows are idempotent per (run_id, stage), so pressing it twice does not re-run what finished. A stage row is claimable only once nothing earlier in its run is pending or running — all of extraction, then each stage ahead of it — so the fleet self-sequences. It covers normalise, spellfix, enrich, chunk, graph-refresh, embed, money and link; pii and quality are never dispatched (PII masking is irreversible and must not run because a flag was left on in a copied config; quality is run-scoped), and both stay reachable through run-stage.

Connection details come from --store/WOMBLEX_STORE_URI, --ingest/WOMBLEX_INGEST_URI (defaults to --store when unset, for back-compatibility), --dsn/WOMBLEX_DB_DSN (or DATABASE_URL), and the S3 store env vars (WOMBLEX_S3_ACCESS_KEY_ID / WOMBLEX_S3_SECRET_ACCESS_KEY, or the AWS_* fallback, plus WOMBLEX_S3_ENDPOINT / AWS_ENDPOINT_URL — 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 — or run them in place with run-stage, above.

Scaling out

Throughput is workers. Each one claims batches with FOR UPDATE SKIP LOCKED, so they cooperate without a broker, without double-processing, and without coordinating with each other — which means you can add and remove workers mid-run, on the same host or across hosts, with no reconfiguration and no restart of the ones already going. A worker that dies mid-batch is not a lost batch: --stale-timeout returns its claim to pending and another worker picks it up. --idle-timeout exits a worker that finds no work, so a fleet can scale to zero on its own once the run drains.

A ready-to-run stack (Postgres + MinIO + scalable workers) lives in docker-compose.yml. It is self-contained by default and points at external Postgres + S3 the moment you set the connection env vars — one file, no code change. The bundled Postgres/MinIO sit behind a local profile, so bring the local stack up explicitly:

docker compose --profile local up -d postgres minio createbuckets init
# upload source docs to the 'womblex' bucket under inbox/, then:
docker compose run --rm womblex enqueue --config configs/example.yaml --create-schema
docker compose up --scale worker=4 worker     # raise or lower at any time

Cloud deployment (external Postgres + external S3)

The same compose file runs against externally-provided Postgres and an externally-provided S3 bucket with no code change — you set three connection env vars (plus S3 credentials) and skip the bundled backends. What makes this work: every connection value in the file is ${VAR:-<local default>}, so unset env is the local stack byte-for-byte and set env is the external service; the bundled postgres/minio sit behind the local profile (so a plain up never starts them); and init/worker/ui declare their dependency on those backends as required: false, so an absent bundled backend is a warning, not a missing-dependency error. (Requires Docker Compose ≥ 2.20.0 for required: false; the compose header documents the one mechanical edit for older engines.)

# 1. Point at the external services. WOMBLEX_STORE_URI takes a path prefix, so
#    Womblex keeps its output in its own folder of a shared bucket — verified:
#    fsspec splits s3://shared/womblex into bucket `shared` + root `womblex/`,
#    and every run lands under runs/<run_id>/documents/ beneath that prefix.
#    WOMBLEX_INGEST_URI must be disjoint from that runs/ output — its own
#    prefix of the same bucket, or a different bucket entirely.
export WOMBLEX_DB_DSN=postgresql://user:pass@db.example:5432/shared  # pragma: allowlist secret -- placeholder, not a real DSN
export WOMBLEX_STORE_URI=s3://shared/womblex        # own prefix of a shared bucket
export WOMBLEX_INGEST_URI=s3://shared/womblex/inbox # disjoint from the store's runs/
export WOMBLEX_S3_ENDPOINT=                          # leave empty for real AWS S3
export AWS_REGION=ap-southeast-2
# Store credentials go on the store-specific vars, NOT AWS_ACCESS_KEY_ID —
# see the SageMaker section: an ambient AWS_ACCESS_KEY_ID is process-global
# and clobbers the instance role the isaacus-sagemaker signer needs. For real
# AWS S3 reachable by the same instance role, leave these unset too.
export WOMBLEX_S3_ACCESS_KEY_ID=... WOMBLEX_S3_SECRET_ACCESS_KEY=...  # pragma: allowlist secret -- placeholders
export ISAACUS_API_KEY=...                           # if enrichment/embeddings run (omit on SageMaker)

# 2. Create the one table Womblex owns (womblex_jobs) in the external DSN.
#    Either run the init one-shot (it resolves WOMBLEX_DB_DSN)...
docker compose run --rm init
#    ...or apply the checked-in schema directly, for DBA review / grants:
#    psql "$WOMBLEX_DB_DSN" -f sql/womblex_jobs.sql

# 3. Enqueue and scale workers exactly as local — no bundled backend started.
docker compose run --rm womblex enqueue --config configs/example.yaml
docker compose up --scale worker=4 worker
docker compose up -d ui                              # optional console, :8080

Womblex owns exactly one table (womblex_jobs) and writes no vectors. It is a well-behaved tenant of a shared database: every statement is scoped to that one table (no DROP/TRUNCATE, no CREATE DATABASE/SCHEMA, no search_path), so it coexists with another system's tables in the same database provided the name does not collide. Embeddings are published as *.embeddings.parquet in the object store (under your WOMBLEX_STORE_URI prefix), not written to Postgres — so pgvector is a property of the shared database for other consumers of those embeddings, never a Womblex requirement. Womblex does not read or write a vector column.

Console (optional)

womblex ui serves a web console over artefacts a run has already written — a run selector and documents table, a Dashboard (queue state and per-stage checkpoint progress), Corpus and Chunk inspectors (a document's chunks with their entity / PII / money overlays), a Pipeline Composer (build a WomblexConfig against the live JSON Schema, with named presets — built-in and operator-saved — and the stage DAG rendered from the stage contracts; it can save the composed config as a preset and enqueue a run over the deployment's configured ingest location — the composer is the console's one dispatch surface), and a Resources console (store / ingest / queue / Isaacus connection checks, with editable ingest and output locations). It is a sidecar, never in-process with the pipeline, and reads either a local run root or the object store a distributed run published to:

pip install womblex[ui]                        # reads a local root or an s3:// store out of the box
womblex ui --output-root output/               # local runs, at :8080
womblex ui --output-root output/ --presets-dir presets/  # + save composer presets
womblex ui --store <uri> --dsn <dsn>           # + dispatch a run into the queue
docker compose up -d ui                        # or beside the stack above

It adds no pipeline logic. Wire a --store, a --dsn and an --ingest location and the Pipeline Composer can plan a run into the queue (workers do the work; the console runs no scheduler); a console with no store or queue configured still reads and inspects runs. Its writable surfaces are deliberately narrow. The report action (POST /api/runs/{run_id}/feedback) files a reviewer's note about a record as a single JSON file under a feedback/ location that is always a sibling of the runs, never inside one — so re-running a stage or purging a run neither disturbs accumulated feedback nor is disturbed by it. The Pipeline Composer's saved presets work the same way: locally they need a writable --presets-dir (or $WOMBLEX_UI_PRESETS_DIR; without one the built-in presets still serve but saving is disabled), and in store-backed mode they land under the object store's own presets/ prefix, alongside feedback/ — so the compose service writes both feedback and presets to the object store, needing no writable mount for them. There is no authentication, so it binds to loopback unless --host says otherwise; put your own control in front of anything wider.

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 workflow — womblex run is extraction only, so chunking is always dispatched as its own stage over the extraction shards.
• E2E composition: womblex chunk --config <yaml> (no --shards) extracts and chunks in one process (a back-compat convenience for simple corpora; womblex run itself no longer chunks).

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.

Preset pipelines

Ready-to-run configs for common end-to-end shapes live in configs/. The console's Pipeline Composer offers the same ones by name (e.g. DEFAULT-Isaacus); the config file is the CLI source of truth.

configs/default-isaacus.yaml — the reference Isaacus pipeline (extract → normalise → spellfix → enrich → chunk → build_graph → embed → money → link → done, for PDF/DOCX). The text is cleaned first (normalise, then spellfix, selected via processing.text_source), then the entity graph, chunk embeddings, monetary amounts and entity links are produced over the one run. Note that womblex run alone runs only extract → redaction detection (extraction only); normalise, spellfix, chunk, enrich, build_graph (graph-refresh), embed, money, link and pii are per-stage commands, normalise/spellfix run before enrich, and enrich must precede chunk so AI chunking reuses the enrichment (no double cost):

RUN=out/$(date -u +run-%Y%m%dT%H%M%SZ); SHARDS=$RUN/documents
CFG=configs/default-isaacus.yaml

womblex run           --config $CFG --run-id "$(basename "$RUN")"  # 1. extract
womblex normalise     --shards "$SHARDS" --config $CFG             # 2. clean text
womblex spellfix      --shards "$SHARDS" --config $CFG             # 3. OCR repair (chains on 2)
womblex enrich        --shards "$SHARDS" --config $CFG             # 4. enrich BEFORE chunk
womblex chunk         --shards "$SHARDS" --config $CFG             # 5. AI chunk (reuses enrich)
womblex graph-refresh --shards "$SHARDS"                           # 6. build_graph edges
womblex embed         --shards "$SHARDS" --config $CFG             # 7. chunk embeddings
womblex money         --shards "$SHARDS" --config $CFG             # 8. amounts (offline)
womblex link          --shards "$SHARDS" --config $CFG             # 9. entity links (needs a register)
womblex manifest      --shards "$SHARDS"                           # 10. consolidate manifest

On object storage, steps 2–9 are the same via womblex run-stage --stage <normalise|spellfix|enrich|chunk|graph-refresh|embed|money|link> --store <uri> --run-id <id> --config $CFG. Linking needs a corpus reference register (linking.reference) you supply. The full command sequence and per-setting rationale are commented inside the config file itself.

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.

The per-stage womblex enrich --shards writes sidecars alongside each batch (enrichment is a core capability — no extra install, just a key or SageMaker endpoints):

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/annotation stages (normalise/spellfix/quality/money)
│   ├── 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.

Commit hook

Two checks run on the files you stage: a secret scan (detect-secrets) and SAST over the local semgrep rulesets in .semgrep/rules/. Install it once per clone:

pip install pre-commit==4.6.1 && pre-commit install

# Run both over the whole tree rather than just staged files
pre-commit run --all-files

pre-commit is deliberately not in the [dev] extra — adding it would change pyproject.toml and rewrite uv.lock, which is a dependency-scoped decision of its own.

The hook is the only thing here that stops an action, and git commit --no-verify walks straight past it. CI re-runs both over the whole tree and adds a scan of reachable history, which is what catches a credential committed behind --no-verify and removed later. Nothing blocks a merge — that is branch protection, a GitHub setting.

If a scan flags something you have checked and know to be safe, record the reason rather than switching the check off: # pragma: allowlist secret for detect-secrets, or # nosemgrep: <rule-id> -- <reason> for semgrep. Both rulesets document their known false-positive classes in their file headers. When a new benign finding is real drift rather than a one-off, regenerate the baseline with the exclusions recorded in .pre-commit-config.yaml and review every new entry before committing it.

Environment check

bash .github/scripts/doctor.sh

Compares what .env.example declares against what is actually set, and any declared runtime pins (mise.toml, .tool-versions) against the interpreters on PATH. It reads variable names only and never prints a value. Variables under an # Optional: comment are reported but never failed, so an unset ISAACUS_API_KEY on an extraction-only clone is a note rather than an error. Not wired into CI, where none of these are set.

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