Multimodal Semantic Extraction System

Act as a Senior AI Architect. Build a Python workflow using Gemini 1.5 Pro to extract semantic key-value details from unstructured documents.

INPUT: Accept PDF documents, scanned JPEGs, or text invoices.
OUTPUT: Extract raw vendor name, registered tax ID, detailed line-items array, total tax, and invoice total. Output formatting must be strictly valid JSON.
EXTRACTION PROTOCOL:
- If a value (like tax ID) is missing, do not guess; output 'UNKNOWN'.
- Run a validator check comparing line-item calculations to the total invoice amount before returning the JSON output. If they do not match, flag in a metadata block.