Skip to content

Packets

Generate coordinated sets of inter-related documents — like a loan application with a credit report, pay stubs, W-2s, and bank statements, all sharing consistent context (same applicant, address, and dates) and merged into one multi-page PDF.

The output is a ready-to-use evaluation dataset for document splitting and classification models.

How It Works

1. Load packet config     → which document types, how many of each
2. Resolve shared context → LLM generates consistent names/addresses/dates
3. Generate sub-documents → each runs the full pipeline (data → critic → PDF → critic)
4. Merge into packet PDF  → individual PDFs concatenated, page indices tracked
5. Emit labels            → per-section ground truth (class, pages, extracted fields)

Each sub-document uses the same staged pipeline as single/batch generation; the packet layer adds coordination: shared-context injection, PDF merging, and label emission.

Usage

Output Structure

The output has two parts: the evaluation dataset (what downstream models consume) and the generation artifacts (intermediates for debugging).

output/<batch-name>/
├── classes.yaml                   # all document types in this packet
├── input/                         # merged multi-document PDFs
│   └── <packet_id>.pdf
├── baseline/                      # ground-truth labels per section
│   └── <packet_id>.pdf/sections/<n>/result.json
├── artifacts/                     # per-packet generation files
│   └── <packet_id>/
│       ├── shared_context.json
│       └── <doc-type>/{pdfs,data,generation_scripts}/
└── config/
    ├── packet.json                # copy of the packet config used
    └── packet_manifest.json       # full run metadata

Label format

Each result.json:

{
  "document_class": { "type": "Insurance Claim" },
  "split_document": { "page_indices": [0] },
  "inference_result": {
    "InsuredName": "Michael Andrew Thompson",
    "ClaimNumber": "CLM-2024-06789"
  }
}
Field Description
document_class.type Document type label (matches classes.yaml)
split_document.page_indices Zero-indexed pages in the merged PDF
inference_result All fields from the generated data — the KIE ground truth

Packet Configuration

Packet configs live in src/seed_data/packets/<name>/packet.json:

{
  "name": "insurance-claim-packet",
  "description": "Property insurance claim with supporting financial documents.",
  "documents": [
    { "document_class": "Insurance Claim", "schema_dir": "insurance-claim",
      "required": true, "min_instances": 1, "max_instances": 1 },
    { "document_class": "Bank Statement", "schema_dir": "bank-statement",
      "required": true, "min_instances": 1, "max_instances": 2 },
    { "document_class": "Invoice", "schema_dir": "invoice",
      "required": true, "min_instances": 1, "max_instances": 3 }
  ],
  "shared_context": {
    "claimant_name": "Full legal name of the person filing the claim",
    "claimant_address": "Street address of the claimant / insured property",
    "incident_date": "Date of the loss/incident (YYYY-MM-DD)"
  },
  "shared_context_hint": "All documents relate to the same insurance claim."
}
Field Required Description
name yes Packet type identifier
documents yes List of document types in the packet
documents[].document_class yes Label used in result.json and classes.yaml
documents[].schema_dir yes Schema directory name (resolved against the sibling schemas/)
documents[].required no Must this doc type appear? Default true (optional docs have a 30% skip chance)
documents[].min_instances / max_instances no Count range per packet (random within range)
shared_context no Explicit shared fields (name → description)
shared_context_hint no Natural-language hint about how documents relate

Shared context

Ensures the same person, address, and dates appear across every sub-document. Two modes:

  1. Explicit — define shared_context; the LLM generates values matching the descriptions.
  2. Inferred — omit it; the LLM reads all schemas, finds fields representing the same entities, and generates consistent values.