Models and settings¶

Choose any model from any provider, and set it for the whole program, a block, a function, or one copy.

import functai
functai.configure(lm="gpt-4.1-mini", temperature=0)   # the model behind every output on this page
from functai import ai, _ai

An AI function says what you want. Which model answers it, and how, is a setting, and settings can change without touching the function.

If you set nothing¶

functai looks at what this machine can use (API keys in the environment, saved keys, subscription logins) and picks a small, capable model, preferring API keys, in this order: gpt-4.1-mini (OpenAI), claude-haiku-4-5 (Anthropic), gemini-2.5-flash (Google), Groq, OpenRouter, then your Claude, ChatGPT or Copilot subscription. It says which, once:

functai: no model chosen, so using gpt-4.1-mini (environment ($OPENAI_API_KEY)). Choose one with functai.configure(lm=...).

For anything you'll run twice, choose explicitly: the default depends on the machine.

Naming a model¶

A model is a string. The provider is found from the name, or given as a prefix:

The litellm spellings ("openai/gpt-4o", "anthropic/claude-sonnet-4-5") work too. Requests go straight to each provider's API through lm15; no provider SDK is installed.

Where settings come from¶

Settings are looked up at every call, from the most specific place to the most general. The first place that sets a value wins:

1. one copy: fn.using(lm="…") returns a copy with other settings;
2. the function: @ai(lm="…"), or fn.lm = "…" later;
3. a block: with functai.configure(lm="…"):;
4. the whole program: functai.configure(lm="…").

@ai
def capital(country: str) -> str:
    """The country's capital city."""
    ...

capital("Australia")                       # from configure(): gpt-4.1-mini

'Canberra'

with functai.configure(lm="gpt-4.1-nano"):
    print(capital("Canada"))               # this block only

Ottawa

fast = capital.using(lm="gpt-4.1-nano", temperature=0)
fast("Brazil")                             # a copy; `capital` is unchanged

'Brasília'

functai.settings reads the effective values:

functai.settings.lm

'gpt-4.1-mini'

A with configure(...) block applies to the current thread and to the threads functai starts from it (for example in evaluate(num_threads=8)), not to threads started elsewhere.

Changing a function after the fact¶

Everything can be set when the function is defined, changed on it later, or changed on a copy, which leaves the original alone:

A bad value (an unknown layout, a template that can't be read back, a connection object where a model name belongs) is refused where you write it, before anything changes.

The model and the connection are separate¶

lm= names the model. client= is how to reach it, for the rare case you build that yourself with lm15: a second account's key, a proxy, a company gateway.

import lm15

@ai(lm="gpt-4.1-mini", client=lm15.OpenAILM(api_key=OTHER_KEY))
def f(text: str) -> str: ...

A model prefix naming another provider than the client's is refused. Credentials and connections are never saved with a program: the machine that loads it uses its own.

All settings¶

A misspelled setting is an error, not silently ignored:

try:
    functai.configure(temprature=0)
except TypeError as error:
    print(error)

configure: unknown setting(s) ['temprature']. functai settings: ['adapter', 'api_key', 'api_retries', 'approve', 'auth', 'autocompile', 'autocompile_n', 'autogen_instructions', 'autoinstruct', 'bake_where', 'base_url', 'cache_replies', 'caller', 'capabilities', 'client', 'debug', 'escalate_below', 'escalate_to', 'include_fn_name_in_instructions', 'instruction_autorefine_calls', 'instruction_autorefine_max_examples', 'instruction_lm', 'journal', 'lm', 'log_calls', 'log_content', 'max_steps', 'module', 'observers', 'on_unreadable', 'optimizer', 'plugins', 'program_observers', 'program_plugins', 'replicate', 'retries', 'teacher', 'teacher_lm', 'tool_errors']; lm15 Config fields: ['cache', 'extensions', 'frequency_penalty', 'logprobs', 'max_tokens', 'presence_penalty', 'probabilities', 'reasoning', 'response_format', 'seed', 'service_tier', 'stop', 'store', 'temperature', 'tool_choice', 'top_k', 'top_p', 'user_id']

Settings some models refuse (OpenAI's reasoning models take no temperature; the ChatGPT backend no max_tokens) are left out of those requests with a one-time warning, so one configure(...) works across models.