Upgrading¶

From 1.1 to 1.2¶

What changes when you move from FunctAI 1.1 to 1.2.

import functai
from functai import ai, module

Most of what is new adds to the API: conversations, tools that ask first, serving, plugins, a new bake. These are the changes that need you to change code, then the ones that change what your code does without any change.

Code to change¶

Settings that no longer exist are refused by name (TypeError: ... unknown setting(s) ['stateful']), so nothing is silently ignored.

Baked models. Bake was rebuilt (Bake it into a small model):

• A model baked with 1.1 is refused, a head included (baked.json format 1): bake it again. A saved program that names one cannot load until then.
• fn.bake(rows) now decides what to train (method="auto"): a head when every output has a fixed set of answers, as before; for any other function, a generative student, where 1.1 refused. That needs pip install "functai[bake]" and a GPU here, or a training service you have set up ("functai[tinker]"); plan_only=True says what it would do and what it would cost, and spends nothing.
• The teacher is no longer run again on the test rows by default (compare_teacher=True to compare).

Inputs are checked, and converted when the meaning is clear¶

Every call, of an AI function or a module, now binds each input to its declared type before anything else runs. A value whose meaning is clear is converted; anything else is refused with functai.InterfaceError (code interface-input), and nothing is sent:

@module
def double(n: int) -> int:
    return n * 2

double("5"), double(5.0)

(10, 10)

try:
    double("five")
except functai.InterfaceError as error:
    print(error.code, error.field, "·", error)

interface-input n · double: input 'n': "five" does not bind to {"type":"integer"}

• Text takes a number (42 is "42"), True/False, a list or a record (as JSON), or a value with a text of its own (a date, a data frame); <object at 0x…> is refused. An integer takes 5.0 and "5"; a number takes "2.5"; a boolean only True or False.
• A record input keeps only the members it declares.
• A missing value (None, NaN, pandas' NA) for an optional input is that input left out: it gets its default. That is what a table's empty cell now means.
• Outputs are checked, never converted, and a record holds only its fields: a model's reply with another member does not fit, and is asked again.
• InterfaceError is a TypeError and a ValueError, so an except TypeError you already have still catches it.

What changes without a code change¶

• Versions. Every function with a default, and every module, gets a new version once (a default now counts by what is written, a module's version includes its interface). The call log's Versions shows a new one. Ratings are matched by signature, which defaults do not change, so they still apply.
• Ratings made with 1.1 in a notebook. rated(fn) now tells two notebooks' functions of the same name apart by the notebook's file. A call logged by 1.1 from a notebook names another file (the kernel's cell file), so some of your earlier ratings may be missing from rated(fn): functai.rated(fn, any_file=True) takes them from every file.
• Who rated. A rating made without by= is recorded under the computer's account, and kept on its own: on a shared account, one person's rating no longer replaces another's (a disagreement is marked disputed).
• A reply cut off at the token limit is sent again with twice max_tokens only when you set max_tokens. Without it, the reply already had the model's whole limit, so it fails at once, and the error says how many tokens went to thinking and what to change.
• Models that only run at temperature 1 (GPT-6; Claude Opus, Sonnet, Fable and Mythos 5) no longer fail when you set temperature=0: the setting is left out of their requests, with one warning.
• The reply cache never keeps a reply that could not be read, and the optimizers' teacher is never answered from it.
• A tool that returns a record (a dataclass, a pydantic model) is shown to the model as JSON, not as Python's str().
• Error messages quote the value at fault (cut after 80 characters), unless log_content keeps that field out of the log.
• The call log is format 2. FunctAI reads both formats; if you parse the JSON lines yourself, records have new fields (program.interface, saw, request_hash...): contract/calls.md says each.

From 0.x¶

What changed in 1.0, and how to move code over.

1.0 kept the way you write functions (@ai, _ai, configure, all=True, stateful, tools, module="cot", .opt, undo_opt, @module, phistory); some of these have changed since (see From 1.1 above). What runs underneath is new: prompts are laid out by lmcc and sent by lm15. Data, metrics and optimizers are new too.

Behaviour changes in 1.0¶

• Automatic instruction writing is opt-in. In 0.x every new function asked the model to rewrite its own instruction (autoinstruct), and the first calls refined it again. That spent money at import time and made prompts change by themselves. Now @ai(autoinstruct=True) or @ai(instruction_autorefine_calls=2) turns them on; they run at the first call, not at definition.
• Prompts are laid out by lmcc, so their text differs from 0.x. Re-run your evaluations after upgrading.
• Unknown settings raise instead of being ignored.