solyanka 0.0.8

Transaction pattern utilities and dataset for statement generators

Solyanka

Toolkit + dataset for transaction-pattern driven synthetic statements and downstream LLM fine-tuning. This package ships the curated YAML files, their schema, and a tiny loader that apps or notebooks can use without worrying about file layout.

Install

pip install solyanka              # consumers
pip install -e ".[dev]"           # local hacking (tests + linters)

Runtime use

from solyanka import PatternsService

svc = PatternsService()                         # auto-discovers packaged data
general = svc.load_general_patterns()
eea = svc.load_eea_patterns()
thailand = svc.load_country_patterns("Thailand")

# Recommended helper: general + (EEA) + country
bundle = svc.get_country_patterns("Germany")

# Fine-grained slices (e.g., validation scripts)
custom = svc.get_patterns(country="Germany", include="general,eea")

# API-ready dictionaries
payload = svc.get_pattern_dicts(country="Spain")

Override the dataset path (e.g., while editing YAML) via PatternsService(base_dir=Path("./transaction_patterns")) or the TRANSACTION_PATTERNS_DIR environment variable.

Layout

• solyanka/transaction_patterns/data/*.yml — curated pattern files (general.yml, eea.yml, <country>.yml).
• solyanka/transaction_patterns/data/schema.json — JSON Schema enforced by tests/CI.
• solyanka/transaction_patterns/service.py — public loader API (keep backward compatible).
• tests/ — schema regression + loader behaviour.

Field spec

Required fields

Field	Meaning
title	Merchant label. Plain string or template object.
currency	Uppercase ISO 4217 (EUR, USD, GBP, ...).
amount_range	{min, max} floats describing the observed local-currency range.
amount_format	Rounding strategy: n>0 decimals, 0 whole units, n<0 powers of ten (e.g., -2 rounds to 100s).
types	Non-empty list of lowercase tags (shopping, restaurant, transportation, ...).

Optional fields

Field	Why / how
weight	Relative selection probability. 100 = baseline, 120–150 very common, 50 niche.
refund_probability	Chance (0–1) that the generator emits a CARD_REFUND for this pattern.
refundDelayMinHours/Max	Boundaries for automatic refund timing (defaults: 72 / 288 hours).
numberOfOccurrences	Global cap per statement (useful for rare, one-off merchants).
subscription_frequency_days	Frequency for recurring charges (e.g., 30 for monthly subscriptions).

Template titles

title:
  type: template
  template: "Revolut**{num}* DUBLIN"
  params:
    num:
      generator: random_digits
      length: 4
      zero_pad: true
      global_constant: true
      transform:
        case: upper

Generators & parameters

Generator	Required params	Optional params	Notes
random_digits	length	zero_pad (default true)	digits only; zero_pad keeps leading zeroes
random_alnum	length	charset	mix of letters/digits; charset restricts symbols. The default is abcdefghijklmnopqrstuvwxyz0123456789.
choice	options	weights (same length)	uniform when weights omitted

Extras:

• global_constant: true — reuse the same generated value across the statement (great for IDs).
• transform.case: upper, lower, or title.

Examples

Simple grocery merchant

- title: "Tesco Express"
  currency: "GBP"
  amount_range: {min: 5.0, max: 50.0}
  amount_format: 2
  types: ["groceries", "shopping"]
  weight: 120

Subscription service

- title: "Netflix.com"
  currency: "EUR"
  amount_range: {min: 13.49, max: 13.49}
  amount_format: 2
  subscription_frequency_days: 30
  numberOfOccurrences: 10
  types: ["entertainment", "subscription"]
  weight: 300

Template with refund metadata

- title:
    type: template
    template: "Airbnb * {code} 662-105-6167"
    params:
      code:
        generator: random_alnum
        length: 12
        charset: "abcdefghijklmnopqrstuvwxyz0123456789"
        transform:
          case: lower
        global_constant: true
  currency: "USD"
  amount_range: {min: 70.0, max: 900.0}
  amount_format: 2
  refund_probability: 0.4
  types: ["housing"]
  weight: 700

Pattern authoring workflow

1. Pick the right file (general.yml, eea.yml, or <country>.yml).
2. Study existing entries (Thailand’s file is a good reference for tone + “uglified” merchant names).
3. Choose realistic amount_range, amount_format, tags, and weights.
4. Use templates when merchants expose reference numbers.
5. Annotate generated blocks with comments (e.g., # Generated transaction pattern - online food).
6. Run pytest to validate against schema.json before committing/publishing.

Tests & release

pytest                # validates YAML + loader invariants
python -m build       # optional local artifact check

• CI: .github/workflows/ci.yml runs pytest on push/PR.
• Release automation: merge PRs into main with major release, minor release, or patch release labels to control how .github/workflows/release-tagger.yml bumps the version after CI finishes green. No label defaults to a build bump (v1.2.3 → v1.2.3.1, etc.). The workflow updates pyproject.toml and tags the commit as v<version>.
• PyPI publish: semantic tags (vMAJOR.MINOR.PATCH with optional .<build_or_label>) trigger .github/workflows/release.yml. The release tagger simply creates the tag, so publishing is entirely driven by tag pushes (manual or automated).
• Need to generate new country patterns for a task? See AGENTS.md for the full enrichment workflow.

Pattern preview workflow

Pull requests that touch solyanka/transaction_patterns/data/** automatically run .github/workflows/pattern-preview.yml. The workflow uses python -m solyanka.pattern_preview to diff the branch against the PR base, synthesize up to three example transactions from the touched patterns, and posts a Markdown table comment back onto the PR so reviewers can eyeball the new merchants. Preview-only fixtures live under tests/pattern_preview/ and are injected via the workflow using the --extra-patterns flag so they stay separate from the shipped data. If the rendered tables grow beyond GitHub’s comment limit, the workflow automatically splits the output into sequential comments while keeping each pattern block intact. Run the same command locally to preview the output before pushing changes:

python -m solyanka.pattern_preview \
  --base-ref origin/main \
  --head-ref HEAD \
  --samples-per-pattern 3 \
  --extra-patterns tests/pattern_preview

Conventions: keep YAML human-readable (sorted keys, helpful comments), avoid UUID-looking titles, and update schema/tests whenever the structure changes.

Purpose recap

Solyanka is the single source of truth for transaction-pattern assets used by the bank-statement generator and any LLM training pipelines. Treat it like a dataset project: tight validation, small focused API surface, deterministic releases.