Bankable

End-to-end CLI tool for extracting structured transaction data from bank and credit card statement PDFs into clean CSV files.

Pipeline: PDF → Marker → .md → Substitutions → LLM → .csv → Pandas cleanup

Features

• PDF to CSV in one command — single pdf_to_csv.py handles the entire pipeline
• Plug-and-play LLM — switch Ollama models via --model flag without code changes
• Batch processing — processes entire directories of PDFs, merges into one master CSV
• Smart table extraction — custom patches fix Marker's handling of collapsed financial tables
• Password support — handles encrypted PDFs via --password flag
• Text substitutions — user-configurable find/replace rules (SUBSTITUTIONS.md) clean up OCR artifacts before LLM processing
• Deterministic post-processing — Credit/Debit classification uses original Markdown markers, not LLM guesses
• Apple Silicon optimized — MPS GPU acceleration where supported

Installation

uv sync

Ollama Setup

brew install ollama
ollama serve
ollama pull llama3.1:8b

Quick Start

# Single file → produces statement.md + statement.csv
uv run pdf_to_csv.py statement.pdf

# Batch directory → produces per-file .md files + all_transactions.csv
uv run pdf_to_csv.py /path/to/statements/

# Encrypted PDFs
uv run pdf_to_csv.py /path/to/statements/ --password "your_password"

# Use a different model
uv run pdf_to_csv.py statement.pdf --model llama3.2:3b

# Overwrite existing outputs
uv run pdf_to_csv.py /path/to/statements/ --overwrite --password "your_password"

Output

Single File

• statement.md — Markdown with extracted tables
• statement.csv — cleaned transaction data

Directory

• <name>.md for each PDF — individual Markdown files
• all_transactions.csv — merged, deduplicated, sorted master CSV

CSV Format

Date,Time,Description,Amount,Type
13/03/2025,21:13:01,SWIGGY,937.00,D
02/04/2025,12:29:45,NETBANKING TRANSFER,42545.00,C

• Type: C (Credit) or D (Debit)
• Sorted by date and time
• Deduplicated across statements

Command Line Options

Option	Description
input	PDF file or directory of PDFs (required)
--model, -m	Ollama model for extraction (default: llama3.1:8b)
--password, -p	Password for encrypted PDFs
--overwrite	Overwrite existing output files
--verbose, -v	Verbose output
--ollama-url	Ollama base URL (default: http://localhost:11434)
--config	Marker config file (default: MARKER.md)
--substitutions	Substitutions config (default: SUBSTITUTIONS.md)
--use-llm	Enable Marker's built-in LLM mode
--no-llm	Disable Marker's built-in LLM mode
--html-tables	HTML tables in Markdown output
--workers, -w	Batch size for parallel processing

Configuration

MARKER.md

Marker conversion settings (output format, table processing, DPI, etc.). CLI flags override these.

SUBSTITUTIONS.md

User-maintained text replacements applied before LLM processing:

"<br>" → " "
"SMARTBUYBANGALORE" → "SMARTBUY BANGALORE"
"PayU*Swiggy Limited" → "SWIGGY"

Project Structure

bankable/
├── pdf_to_csv.py          # Main program — end-to-end PDF → CSV
├── marker_patched.py      # Marker subprocess wrapper (applies patches)
├── table_split_patch.py   # Fixes collapsed transaction rows in tables
├── pdfprovider_patch.py   # Adds password support to Marker's PDF provider
├── MARKER.md              # Marker configuration
├── SUBSTITUTIONS.md       # Text substitution rules
└── pyproject.toml         # Dependencies

Requirements

• Python 3.13+
• macOS with Apple Silicon recommended
• Ollama with a model pulled (e.g. llama3.1:8b)

Notes

• Existing files are skipped by default. Use --overwrite to reprocess.
• GPU: Apple Silicon MPS where supported; some models fall back to CPU.
• LLM accuracy: llama3.1:8b is recommended over llama3.2:3b for fewer digit transposition errors.
• Credit detection uses Cr markers from the original Markdown tables, not LLM output.

License

MIT License

Acknowledgments

• Marker — PDF to Markdown conversion
• Ollama — local LLM runtime