Claude Data Extraction: Unstructured to Structured

Extract structured data from the following [document type: invoice/resume/contract/etc.].

OUTPUT FORMAT: Valid JSON object with this exact structure:

{
  "field_name": "value or null",
  "field_name": "value or null"
}

FIELD DEFINITIONS:
- field_name: [What it represents. Where to find it in the document.
  If multiple values: which one to pick. If missing: use null.]

RULES:
- Return ONLY the JSON object. No markdown fences, no explanatory text.
- If a field is mentioned multiple times with different values, use [rule].
- If a field is ambiguous, use the most [specific/recent/authoritative] value.
- If a field is completely absent, use null (not "N/A" or empty string).
- For date fields, use ISO 8601 format (YYYY-MM-DD).
- For currency fields, use numbers without currency symbols.

Extract structured data from this invoice:

{
  "invoice_number": "string — the invoice ID/number. Usually labeled 'Invoice #' or 'Inv No.'",
  "invoice_date": "string — date the invoice was issued, ISO 8601 format",
  "due_date": "string — payment due date, ISO 8601. If not specified, use null",
  "vendor": {
    "name": "string — company or individual name sending the invoice",
    "tax_id": "string or null — VAT/GST/EIN number if present",
    "email": "string or null",
    "address": "string or null — full address as it appears"
  },
  "client": {
    "name": "string",
    "address": "string or null"
  },
  "line_items": [
    {
      "description": "string",
      "quantity": "number",
      "unit_price": "number",
      "total": "number"
    }
  ],
  "subtotal": "number — before tax",
  "tax_rate": "number or null — as percentage (e.g., 20 not 0.20)",
  "tax_amount": "number or null",
  "total": "number — final amount due",
  "currency": "string — ISO 4217 code (USD, EUR, GBP). Default: USD if not specified"
}

RULES:
- If the invoice has a 'paid' stamp or 'PAID' watermark, add "status": "paid"
- If line item totals don't sum to the stated total, flag with "discrepancy": true
- If tax amount is calculable from subtotal × tax_rate but differs from stated,
  use the STATED amount and flag the discrepancy

From this legal document, extract all mentioned entities and their relationships:

{
  "entities": [
    {
      "name": "string — official legal name",
      "type": "corporation|individual|government_agency|trust|other",
      "role": "buyer|seller|lender|borrower|guarantor|witness|other",
      "jurisdiction": "string — state/country of incorporation or residence",
      "aliases": ["string — other names used for this entity in the document"]
    }
  ],
  "relationships": [
    {
      "from": "string — entity name (must match an entity.name exactly)",
      "to": "string — entity name",
      "type": "owns|owes|guarantees|licenses|indemnifies|other",
      "details": "string — brief description of the relationship",
      "clause_reference": "string — section/paragraph where defined"
    }
  ],
  "key_terms": [
    {
      "term": "string",
      "definition": "string — how the document defines it, or null if undefined",
      "section": "string — where it appears"
    }
  ]
}

AMBIGUITY RULES:
- If an entity is referred to by multiple names, use the first official name as 'name'
  and list others in 'aliases'
- If a relationship is implied but not explicit, include it but set a flag: "explicit": false

Extract all events with their temporal information:

{
  "events": [
    {
      "description": "string — what happened",
      "date": "string or null — ISO 8601. If only year: YYYY. If year-month: YYYY-MM",
      "date_type": "exact|approximate|range_start|range_end|before|after|unknown",
      "date_context": "string — how the date was expressed ('Q3 2024', 'last Tuesday', 'sometime in spring')",
      "participants": ["string — who was involved"],
      "location": "string or null",
      "confidence": "high|medium|low — how confident are you in the extracted date?"
    }
  ]
}

TEMPORAL REASONING:
- 'Last Tuesday' relative to document date of 2024-03-15 → 2024-03-12
- 'The following quarter' after Q3 2024 → Q4 2024 (range: 2024-10-01 to 2024-12-31)
- If the document date is unknown and a relative date is used, set date to null and
  preserve the relative expression in date_context

Process the following [N] documents. For EACH document, return a JSON object
with the extraction schema. Return ALL results as a JSON array:

[
  { "doc_id": 1, ...extracted fields... },
  { "doc_id": 2, ...extracted fields... }
]

If extraction fails for a document (e.g., wrong format, unreadable):
{ "doc_id": N, "error": "reason for failure" }

After extraction, verify your output:

1. COUNT CHECK: Number of extracted items should match what you observed
   in the document. State: "Extracted [N] items. Document appears to contain [M]."

2. REQUIRED FIELD CHECK: For each required field, confirm it has a non-null value
   or explain why it's null: "Field [X] is null because [reason]."

3. SUSPICIOUS VALUE FLAG: If any value seems unusual (e.g., $1,000,000 for a
   coffee invoice), flag it: "SUSPICIOUS: [field] = [value]. Possible error in
   extraction or source document."

4. CONFIDENCE SCORE: Overall confidence in the extraction: [0.0 - 1.0]
   Below 0.8: list the uncertain fields and why.

This document may contain:
- OCR errors (scanned document)
- Handwritten annotations
- Conflicting information (e.g., two different totals)
- Missing pages or sections

When you encounter issues:
- OCR UNCERTAINTY: "The value appears to be [best guess] but could be [alternative].
  Context: [surrounding text]."
- CONFLICT: "Field [X] appears twice: value A (section 3) and value B (section 7).
  Using [choice] because [reasoning]."
- MISSING: "Section [X] appears to be missing. Fields [Y, Z] cannot be extracted."