AJ Bell Bare Trust — Document Extraction Pipeline

Extracts structured data from a handwritten AJ Bell Bare Trust Dealing Account application form (PDF) and outputs a validated JSON file.

Setup

pip install -r requirements.txt

Create a .env file in the project folder:

GEMINI_API_KEY=your_key_here

Run

python main.py --input DataExtraction.pdf --output output.json

Add --verbose to see per-page field counts, token usage, and cost breakdown.

Output

output.json — all form fields mapped to the schema defined in schemas/JSONSchema.txt.

How it works

1. PDF is converted to images (300 DPI, one per page)
2. Each page image is sent to Gemini vision — fields extracted as XML with chain-of-thought reasoning
3. All extracted fields are mapped to the target JSON schema using Gemini structured output
4. Result is validated against the schema; errors trigger an automatic correction pass
5. Post-processing applies deterministic fixes (NI numbers, sort code formatting, country casing)

Assumptions

• The form is always the standard 8-page AJ Bell Bare Trust Dealing Account form. Other page counts are handled with a generic fallback but page-specific accuracy will be lower.
• Titles (Dr/Mr/Mrs/Miss/Ms/Other) are marked by striking through all options except the chosen one — not by circling or ticking.
• Whatever is physically written in the Surname box is treated as the surname, even if it looks like a first name. The applicant's input is preserved as-is.
• Unticked checkboxes are extracted as false, not null. A null checkbox means that field is absent from the page entirely.
• NI numbers starting with N1 followed by a digit are corrected to NI — the digit 1 is a common OCR misread of capital I.
• Sort codes are formatted as XX-XX-XX. If the model returns 6 bare digits, post-processing adds the dashes.
• Country fields matching "uk" in any casing are normalised to "UK".
• If Section D (Second Trustee) is left completely blank, it is not included in TrusteeDetails.
• Signature rows with no name written are not included in trustee_signatures.
• The proof_of_registration_attached field is always null — the form does not have a checkbox for this; it is a separate document attached externally.
• identity_verification_consent and fraud_prevention_checks_consent are null on the standard form — these checkboxes do not appear on the version of the form used.