adjudicate

A small library for one pattern: a detector proposes a span in a document and a short list of candidate replacements; a local language model picks one of them, or KEEP, from the surrounding sentence and a rule of thumb; the applier splices the choice in with case preserved. Every change is a bounded, logged, reasoned edit. The model never rewrites prose.

It is the toolkit behind stylefix (US/UK spelling and style) and proofix (typos, punctuation, commas). It is published because they depend on it; the API may change between 0.x releases.

pip install adjudicate

Modules: llm (OpenAI-compatible client and endpoint resolution), core (Item, Decision, Cache, adjudicate), apply (Edit, apply_edits), context (sentence context, Markdown prose regions), dictionary (Hunspell lookups over the bundled en_US and en_GB dictionaries), setup (run_setup and check_endpoint, behind the tools' --setup and --check-endpoint).

Choosing the model endpoint

resolve_endpoint(app, preset="auto", url=None, model=None, key=None, think=False, timeout=None) merges settings from, highest first:

1. its arguments (the calling tool's --url, --model, --timeout)
2. <APP>_LLM_URL, <APP>_LLM_MODEL, <APP>_LLM_KEY, e.g. STYLEFIX_LLM_URL
3. ~/.config/<app>/config.toml
4. ADJUDICATE_LLM_URL, ADJUDICATE_LLM_MODEL, ADJUDICATE_LLM_KEY
5. ~/.config/adjudicate/config.toml, shared by every tool

($XDG_CONFIG_HOME replaces ~/.config when set.) Layers merge field by field, except that a layer naming a url drops the model, key and thinking_switch of the layers below it, since those described another server. With no url anywhere, a local Ollama is used if it is running with a model pulled; otherwise there is no model and callers fall back to listing items for review.

[llm]
url = "https://api.example.com/v1"   # any OpenAI-compatible base URL
model = "some-model"                  # optional: the server's first listed model
api_key_env = "EXAMPLE_API_KEY"       # or api_key = "..."; not both
timeout = 120                         # seconds per request
thinking_switch = false               # true for vLLM/LiteLLM serving a reasoning model

thinking_switch = true sends chat_template_kwargs: {enable_thinking: false}, which vLLM honours and strict APIs reject; if a server answers 400 or 422 to it, it is dropped for the rest of the run. Unknown keys, wrong types, empty strings, a timeout that is not positive, and an api_key_env naming an unset variable are errors (ConfigError), so a typo is never silently ignored. An api_key_env is read only if it survives the merge: a higher layer's url or key discards it first.

License

The code is MIT (see LICENSE). The dictionaries in adjudicate/dict/ keep their own licenses: en_US under the SCOWL notice in LICENSE-en_US.txt, en_GB under the LGPL as stated in LICENSE-en_GB.txt.