Beru Learning

When you complete a trade, your broker issues a trade confirmation. These slips document essential tax and accounting information: settlement dates, commission schedules, exchange fees, and net cash amounts. Because brokers deliver them across incompatible PDF formats, importing them into standard spreadsheets usually requires tedious manual entry.

Sending personal financial PDFs to hosted AI APIs exposes sensitive account numbers and trading activity. Setting up a local LLM trade confirmation parser lets you process transaction documents on your own computer, outputting clean JSON records while keeping private account identifiers on your local drive.

The Structural Failure of Standard Text Parsers

Deterministic scrapers and regular expressions work well on straightforward tables, but trade confirmations frequently break them. Files often contain multi-column text wraps, unusual font encodings, and fragmented fee summaries.

A regex rule fails when an exchange assessment appears on its own line or when an options spread splits commission figures across multiple sections. Quantized language models running locally handle these layouts better because they parse natural language in context, recognizing that an isolated numeric value is an SEC transaction fee rather than a share quantity.

Setting Up the Local Prompt and Schema Pipeline

To extract data locally, run a quantized open-weights model using an open-source inference engine on standard consumer hardware. Pass the raw text extracted from the PDF alongside a strict output schema.

Define Exact JSON Keys: Require the model to populate specific keys, including trade_date, settlement_date, symbol, action, quantity, unit_price, fees, and net_amount.
Enforce Null Handling: Instruct the model to return null when a field is missing instead of guessing or inserting default text.
Strip PII Upstream: Run a basic script to redact your name, street address, and account numbers before passing the text buffer to the model.

Restricting the model to a defined JSON structure turns inconsistent PDF text into clean records ready for accounting databases.

Building Deterministic Post-Processing Guardrails

Language models generate tokens probabilistically, so their outputs should never be accepted as accurate accounting records without independent checks. Even a swapped digit throws off cost-basis calculations.

Pass the model output directly into a deterministic validation script. The script checks a standard arithmetic balance:

(Quantity × Unit Price) ± Fees = Net Settlement Amount

If the calculation does not balance to the penny, the script flags the confirmation for manual review. This setup combines the document-parsing capability of a language model with the precision of standard math assertions.

Frequently Asked Questions

What hardware is required to parse PDFs using a local model?

A computer with 16 gigabytes of unified memory or a dedicated graphics card with 8 gigabytes of VRAM can comfortably run quantized 7-billion or 8-billion parameter models at practical speeds.

Why not just use an online document converter?

Online file converters upload your documents to third-party servers, exposing account numbers, balances, full names, and broker relationships to potential data breaches.

Can a local model handle scanned physical receipts as well as native digital PDFs?

Yes, but scanned files require a local OCR step first to extract text characters before passing that text into the model's context window.

Key Takeaways

Local LLMs extract structured financial data without exposing private brokerage records to cloud APIs.
Open-weights models recognize semantic relationships across fragmented and multi-column PDF layouts.
Stripping personal identifiers upstream adds an additional privacy barrier during local document ingestion.
Deterministic validation code must programmatically verify that trade mathematics balance before saving to a ledger.
Structured JSON outputs make historical trade confirmation data easily searchable for tax basis calculations.

Using Local LLMs to Extract Clean Data from Trade Confirmations