Package index • functai

Writing an AI function

A formula, a sentence saying what it does, and a codebook of its fields.

ai_config() with_ai_config() local_ai_config(): Settings for every AI function
update(<functai_fn>): Change an AI function's settings
model_capabilities(): What FunctAI declares a model can do

predict(<functai_fn>) augment(<functai_fn>): Predictions for a data frame
ai_problems(): The calls that failed in the last call of an AI function

evaluate(): How often is a model right?
tidy(<functai_evaluation>) glance(<functai_evaluation>) augment(<functai_evaluation>): Tidy an evaluation
exact_match(): Is an answer right?
score_interval(): The mean of scores and its 95% interval
normalize_text(): Text as exact_match compares it
casefold_full(): Unicode full case folding

labeled_few_shot() bootstrap_few_shot(): Worked examples from rows with known answers
gepa(): Rewrite the instruction from the function's mistakes (GEPA)
random_search() instruction_search(): Try several sets of worked examples, keep the best
ai_trials(): The search that improved a function
with_demos() with_instructions() ai_demos(): Worked examples and instructions
compare(): Two versions compared row by row

ai_model() update(<ai_model>) tunable(<ai_model>): An AI model, for tidymodels
worked_examples(): How many worked examples, as a tuning parameter

ai_interface(): What a program takes and gives
ai_version() ai_instructions() ai_signature() ai_signature_id() ai_render(): What an AI function sends, and its version

Your own code around AI functions, as one call; a call watched while it is made.

ai_program() program_from_interface(): Your own code around AI functions, as one program
opaque(): Any R value
predict(<functai_program>) augment(<functai_program>): Predictions of a program for a data frame
ai_stream() stop_stream() ai_text(): Watch a call while it is made

Calls that remember each other, kept in a store; tools that ask a person first.

ai_conversation(): A conversation with an AI function or a program
all_turns() last_turns(): Which earlier turns a turn is shown
remember(): What an AI function inside a program's conversation remembers
ai_turns() ai_turn(): The turns of a conversation
continue_from(): Continue a conversation from an earlier turn
move_head(): Make a turn the conversation's head, for everyone who opens it
wait_turn() stop_turn() abandon_turn() resume_turn(): What can be done with a turn
approve() deny(): Answer a person's question about a tool call
earlier(): The conversation so far, inside a program's turn
folder_store() memory_conversations(): Where conversations are kept

Hooks that change calls, every change recorded.

ai_plugin() on_hook(): A plugin: hooks that change what programs do
ai_change(): What a hook changes
ai_ask(): Ask a person whether a tool may run
keep_entry() entries(): A plugin's entries in a conversation
branch_turns(): The done turns of an ended turn's branch, for a turn_end hook
compaction(): Keep a long conversation short: older turns folded into a summary
delegate(): Another program as a tool
load_plugin(): A plugin from a file

The reply cache, a smaller call log, a judge’s evidence.

Observers, journals and stores, by the contract’s rules.

ai_journal(): Keep each call tree's log while it is written
memory_store() store_refuse(): A store of call trees' logs
settle(): What a journal kept of an end it could not confirm
follower() follow_event() follow_state() follow_recover(): Follow call trees' logs live
outside_view(): What a caller outside a program sees of its log
turn_events(): A turn's events, from its store

The examples every trainer reads, and students trained anywhere.

bake_examples() export_examples(): The training conversations of an AI function
baked(): A model baked anywhere, served by an OpenAI-compatible server

ai_model_fit() ai_model_predict(): Fit an AI model (parsnip's fit function for the "functai" engine)
as_field(): Turn a type into a field
reexports augment tidy glance tunable: Objects exported from other packages
functai functai-package: functai: Functions Whose Body a Language Model Writes, for the Tidyverse