Invoices From OCR

Coding prompt that turns raw OCR output from scanned or photographed invoices into clean, validated, structured invoice records — complete with line items, totals, and confidence scores.

<role>
You are a senior software engineer specializing in document intelligence. You build dependable, production-quality parsers that convert imperfect optical character recognition (OCR) output into reliable structured invoice data for accounting and ERP systems.
</role>

<task>
Write a complete, runnable solution in [programming language] that takes the raw OCR text of one or more invoices — noisy, misordered, and error-prone, as produced by [OCR engine such as Tesseract or Google Cloud Vision] — and returns a structured invoice record for each document.
</task>

<context>
Input: raw OCR text, typically in [source language], containing OCR artifacts such as misread characters, broken line breaks, shifted columns, duplicated headers, and inconsistent date and number formats.
The parser must be resilient to realistic messiness rather than assuming perfectly formatted text.
It is consumed downstream by an accounting system that expects stable field names, so the output shape must remain identical across all documents in a batch.
</context>

<constraints>
- Extract these fields: [invoice number], [invoice date] and [due date] (normalized to [target format, e.g. ISO 8601]), [vendor name] and [vendor tax ID], [buyer name], [currency code], subtotal, tax amount, total amount, and a line_items array where each entry contains [description], [quantity], [unit price], and [line total].
- Convert amounts using [target currency] and [number format, e.g. European decimal comma] and use exact decimal arithmetic, never floating point.
- Handle multiple date and currency formats, including [list of formats seen in your documents].
- Attach a confidence score from 0.0 to 1.0 to every extracted field, and clearly mark fields that require human review when confidence falls below [threshold, e.g. 0.8].
- Tolerate missing optional fields gracefully; never invent a value that is not supported by the source text.
- Include a [validation and normalization step] that resolves conflicts such as a total that does not equal subtotal plus tax.
- Structure the code into clear, single-responsibility functions or classes, include type hints, and add concise docstrings explaining the parsing decision for each field.
- Supply a small runnable example with sample OCR input, the resulting JSON output, and a set of unit tests covering messy real-world cases such as a rotated table and a merged header row.
- Keep dependencies minimal and list any required packages with installation instructions.
</constraints>

<format>
Return the answer in this order:
1. A brief explanation of the parsing strategy in no more than [number, e.g. 5] sentences.
2. The complete code in a single, runnable file or module.
3. Sample input and expected JSON output.
4. Test cases with their expected results.
5. A short list of parsing limitations and the tuning options a developer can adjust.
</format>

<tone>
Write in a clear, pragmatic engineering voice: precise, concise, and free of filler. Explain trade-offs directly and prefer robust, well-tested patterns over clever shortcuts.
</tone>

Now produce the full solution for parsing invoices from [OCR engine] output, using [programming language] and targeting [target system or framework].