smallwords

smallwords is a tiny Python package for controlled-vocabulary prompting plus portable output resources. It keeps one wordlist at the center of the workflow so prompt text, GBNF, JSON Schema, and post-generation validation all stay in sync.

The package ships with a small set of bundled wordlists: direct source-backed lists such as moby_898, basic_850, and special_english_1475, plus a couple of intentionally themed remixes. By default, the built-ins also allow slight family variants such as go, goes, and going.

It supports Python 3.10 and newer.

The hosted API-and-examples docs live at cmccomb.github.io/smallwords.

Installation

pip install smallwords

For local development, create and activate a virtualenv first:

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev]"

Quick Start

from smallwords import OutputResources, OutputShape, allow_input_words
from smallwords.prompts import build_prompt

shape = OutputShape(max_words_per_line=24, max_lines=1)
spec = allow_input_words("basic_850", "bridge")
resources = OutputResources.from_wordlist(spec, shape=shape)
prompt = build_prompt("explain", "How does a bridge work?", wordlist=spec)
schema = resources.json_schema(key="answer", title="bridge_explanation")

text = "A bridge is a structure that helps people and things move across a river or a deep place."
result = resources.validate(text)
assert result.valid, result.issues

The contrast is the point. build_prompt(...) is the soft instruction layer. OutputResources gives you the matching hard constraints in both GBNF and JSON Schema form. resources.validate(...) checks a response against the same vocabulary and output shape. The older is_compliant(...) helper remains a vocabulary-only check; it does not check punctuation or output shape.

Run the complete offline example without a model server:

python examples/offline_quickstart.py

If you want the model to be able to repeat an essential topic term such as bridge, pass just that term to allow_input_words(...). Passing an entire technical passage will also allow every technical word in it.

Built-In Wordlists

• moby_898: the full normalized alpha-only Moby Words II frequency list
• basic_850: Charles Ogden's Basic English 850
• special_english_1475: Voice of America Special English
• caveman_898: a size-neutral surface-only moby_898 remix with caveman adjustments
• pirate_898: a size-neutral moby_898 remix with pirate adjustments

The bundled text files live in src/smallwords/data/. moby_898, basic_850, and special_english_1475 are direct source-backed lists. caveman_898 and pirate_898 are derived size-neutral remixes built on top of moby_898.

The themed remixes live in src/smallwords/themes/caveman.py and src/smallwords/themes/pirate.py. If you want to build your own, use remix_wordlist(...) with a base list plus curated additions and removals.

Contrastive Example

This is the clearest way to see what smallwords is trying to do. Both blocks below are genuine local Qwen outputs from April 5, 2026. The first uses a plain prompt. The second uses the same base prompt plus an explicit basic_850 vocabulary list, the topic word bridge, and the generated GBNF.

A plain prompt stays fairly natural:

A bridge connects two points, usually across a body of water or a gap, allowing people and vehicles to cross safely.

A constrained basic_850 + topic words run stays simpler while still sounding reasonably natural:

A bridge is a structure that helps people and things move across a river or a deep place.

These runs use llama-server from llama.cpp and Qwen/Qwen3-8B-GGUF via bartowski/Qwen_Qwen3-8B-GGUF.

Reproduce that comparison from a clone of the repository with an activated virtualenv:

llama-server -hf bartowski/Qwen_Qwen3-8B-GGUF:q4_k_m --host 127.0.0.1 --port 8080 --reasoning-budget 0 --log-disable
python examples/readme_bridge_contrast.py

Examples

Start with the offline quick start above. For live-model examples, see the repository's examples/README.md for the runnable examples. The bridge contrast, pirate greeting, and technical rewrite call a live llama-server model with a prompt plus generated grammar.

Development

Run these commands from an activated virtualenv:

python -m pip install -e ".[dev]"
python -m ruff check .
python -m ruff format --check .
python -m pytest
python scripts/check_documentation.py
python -m sphinx -W --keep-going -b html docs docs/_build/html
python -m build
python -m twine check --strict dist/*

CI runs linting, tests, the documentation policy check, a >=90% coverage gate, a Sphinx docs build, and a package build on GitHub Actions.

For release steps and Trusted Publishing setup, see RELEASING.md.