DICIE Pipeline

Paper Fig. 1 — process → classify → extract → respond

DICIE

Implements the Document Image Classification and Information Extraction chain from Raj, Dickinson & Fung, Document Classification and Information Extraction framework for Insurance Applications.

NoteJump links

Usage · Pipeline hub · Notebook walkthrough · Commands

flowchart LR
  IN[PDF / images / text] --> S1[1 · Document Processing]
  S1 --> S2[2 · Classification]
  S2 --> S3[3 · Information Extraction]
  S3 --> OUT[4 · Aggregate + respond]

This module (src/docie/) is image-first and application-scoped. It complements the existing markdown → classify → extract → vision → summarize chain in src/pipeline/, which remains the general ACORD intake / memo path.

Applications (paper §III)

Application Classes Extracted fields Taxonomy
medical_bills hcfa, ub04, other claim_id, name, dob, patient_id, address taxonomy/medical_bills.yaml
salvage_claims log, sales, other claim_id, vin, year, make, model taxonomy/salvage_claims.yaml
acord ACORD intake taxonomy claim / policy / loss fields taxonomy/acord_form_categories.yaml

Each taxonomy may define:

  • categories — labels + aliases used by Stage 2 heuristics
  • extraction_fields — Stage 3 field set
  • business_rules.prefer_non_other — prefer non-other on classification ties
  • business_rules.review_confidence_threshold — human-review gate (default 0.55)

Stages

1. Document Processing (src/docie/processing.py)

  • PDF → page images via PyMuPDF at configurable DPI (default 300)
  • Image preprocessing: grayscale conversion, retain dimensions
  • OCR via PyTesseract when installed (pip install -e ".[ocr]"), with PDF text-layer fallback; produces word tokens + 0–1000 normalized boxes for LayoutLM-style extractors
  • Plain text inputs are rendered to a page image so the image-first chain stays intact
  • Cached page images under data/pipeline/cache/docie/<application>/pages/

2. Document Classification (src/docie/classify.py)

  • Per-page scoring (keyword / alias heuristics; optional ViT when weights match)
  • Document-level aggregation: confidence-weighted majority vote
  • Business rule: prefer non-other on ties (paper medical / salvage setups)

3. Information Extraction (src/docie/extract.py)

  • Field extraction conditioned on the predicted document type
  • LayoutLM / token-classifier weights when present under models/extractor*
  • Regex heuristics covering medical + salvage field patterns otherwise
  • Missing expected fields (e.g. LOG without VIN) lower confidence and flag review

4. Output aggregation (src/docie/aggregate.py)

  • Merge classification + extraction
  • Route low-confidence / OCR-empty / missing-field cases to human review
  • Compact response_payload() for REST / claim-center downstream updates
  • Optional downstream_sink callback for webhooks / queue publishers

Evaluation (paper Table I / Table II)

Classification reports accuracy + AUC (OVR/OVO); extraction reports per-field precision / recall / F1 (harmonic mean (2PR/(P+R))).

python -m src.docie.eval --all
python -m src.docie.eval --application salvage_claims
# Writes evaluation/reports/docie_{application}_{metrics.json,report.md}

Gold set: data/eval/docie_eval_set.jsonl.

For a larger regenerable sample house (claim bundles, richer skeletons, train/val/test splits), use the SQLite corpus in src/storage/ — see Sample Document Corpus. Export DICIE JSONL with python -m src.storage export --format docie.

Usage

# Salvage Letter of Guarantee (text demo)
python -m src.docie \
  --application salvage_claims \
  --text "LETTER OF GUARANTEE
Claim Number: CLM-2024-100200
VIN: 1HGCM82633A004352
Year: 2018
Make: Honda
Model: Accord" \
  --response-only

# Medical HCFA PDF / image
python -m src.docie --application medical_bills --pdf path/to/bill.pdf
python -m src.docie --application medical_bills --image path/to/page.png

# Batch JSONL
python -m src.docie \
  --application salvage_claims \
  --in tests/fixtures/sample_docie_documents.jsonl \
  --out data/pipeline/docie/salvage_demo.jsonl

Batch outputs:

Artifact Description
*.jsonl Full prediction rows (DociePrediction.to_dict())
*.summary.json Counts, label histogram, mean confidences
*.human_review.jsonl Compact payloads with needs_human_review=true
data/provenance_log.jsonl Provenance row (stage=docie_pipeline)

Optional FastAPI server (paper §VI ECS/FastAPI serving shape):

pip install -e ".[serve]"
python -m src.docie.serve --application salvage_claims --port 8080
# GET  /health
# POST /v1/predict       (multipart file)
# POST /v1/predict/text  (JSON {text, application, record_id})

Python entry points:

from src.docie import DociePipeline, process_document

process_document(application="salvage_claims", text="LETTER OF GUARANTEE\n...")
pipe = DociePipeline(application="medical_bills")
prediction = pipe.process(record_id="x", pdf_path="bill.pdf")

See src/docie/README.md for the full CLI flag list, prediction schema, and REST contract.

Relation to the chained analysis orchestrator

Concern src/docie/ (this module) src/pipeline/
Ordering process → classify → extract → respond markdown → classify → extract → vision → summarize
Primary signal page images + OCR structured markdown for LLM context
Applications medical bills, salvage claims ACORD intake + adjuster memo
Output classification + fields (+ review flag) classification + fields + memo
Front-ends CLI + optional FastAPI CLI + Discord (/analyze)

Use DICIE when matching the paper flowchart / insurance workflow apps; use the chained orchestrator when you need markdown-first LLM memos.

Testing & fixtures

pytest tests/test_docie_pipeline.py -q

Synthetic fixtures (no real insurer data): tests/fixtures/sample_docie_documents.jsonl.

Optional extras

Extra Purpose
.[ocr] PyTesseract OCR for scanned pages
.[serve] FastAPI + uvicorn REST server
.[dev] pytest + ruff for local development

Trained ViT / LayoutLM weights are optional; heuristic backends keep the chain runnable without them.

See also

Pipeline hub · Architecture · Sample corpus · DICIE notebook · Data Provenance

Back to top