PDF Extraction

Pull structured data out of PDFs using Google Gemini, with a schema you control.

Give it a document and a field list, and it returns clean JSON. No per-document
templates, no coordinate mapping, no retraining — the schema is the configuration,
so the same toolkit handles invoices, statements, forms, reports and contracts
without changes.

What it does

• Schema-driven extraction — describe the fields you want and their types; the model
  is instructed to return exactly that shape, and the result is validated against it
  before it reaches you.
• Native PDF understanding — the document goes to Gemini as a PDF, so layout, tables
  and multi-column text are read as a person would read them. No OCR step to configure.
• Confidence and provenance — every field comes back with a confidence score and the
  page it was found on, so low-certainty values can be routed for review instead of
  trusted silently.
• Missing means missing — a field that isn't in the document returns null rather
  than a plausible guess.

Typical uses

• Invoices, purchase orders and receipts into an approvals queue
• Bank or card statements into a reconciliation table
• Application forms and submitted questionnaires
• Certificates, licences and compliance documents with an expiry to track
• Any recurring PDF a person is currently retyping

Setup

1. Install the toolkit.
2. Create a Gemini API Key secret.
3. Call extract_pdf_fields with a storage path and your field list.

The API key is read from the secret store at call time and is never returned in a
response or written to the extraction record.

Defining fields

{
  "fields": [
    {"name": "invoice_number", "type": "string",  "description": "Invoice or reference number"},
    {"name": "issue_date",     "type": "date",    "description": "Date of issue, ISO 8601"},
    {"name": "total_amount",   "type": "number",  "description": "Total including tax"},
    {"name": "currency",       "type": "string",  "description": "ISO 4217 code"},
    {"name": "is_paid",        "type": "boolean", "description": "Whether marked paid"}
  ]
}

Supported types are string, number, integer, , and .