daaf — decision-as-a-function

Write a typed Python function with a docstring. Laya answers it, locally, in any of 100+ languages, in tens of milliseconds.

from enum import Enum, IntEnum
from typing import Literal
from daaf import decide

@decide
def is_scam(message: str) -> bool:
    """Is this message a scam trying to get money or credentials?"""

class Dept(Enum):
    billing = "invoices, payments, refunds"   # the value describes the option to the model
    support = "bugs, outages, login problems"
    sales = "pricing, buying a product"

@decide
def department(message: str) -> Dept:
    """Which department should handle this message?"""

is_scam("Votre colis est bloqué, payez 1,99 € ici")  # True
department("I was billed twice")                     # Dept.billing

The function body is never run: the docstring is the question, the return type is the set of possible answers.

Install

pip install daaf

The first call downloads the Laya model (a few hundred MB) and loads it, which takes a few seconds. Later calls are fast.

Return types

class Urgency(IntEnum):
    not_urgent = 0
    soon = 1
    blocking = 2

@decide
def urgency(message: str) -> Urgency:
    """How urgent is this message?"""

urgency("Production is down!")  # Urgency.blocking

Arguments

• One argument: its value is the text Laya reads.
• Several arguments: Laya reads them as named fields, for example def urgent(subject: str, body: str) -> bool.

Many inputs at once

is_scam.batch(["message 1", "message 2", ...])         # one-argument function
urgent.batch([("subject", "body"), {"subject": "s", "body": "b"}])  # several arguments

Inputs share forward passes, which is much faster than a loop. Results come back in input order. batch_size= caps how many go through the model at once.

Options

@decide(min_confidence=0.8)      # raise daaf.Undecided below 80% instead of guessing
def is_scam(message: str) -> bool:
    """Is this message a scam trying to get money or credentials?"""

@decide(model="multilingual", max_len=8192)  # anything else goes to laya's Router.predict
def is_contract(document: str) -> bool:
    """Is this document a contract?"""

@decide(router=my_router)        # your own Router, or any object with Laya's .predict / .predict_batch
def is_spam(text: str) -> bool:
    """Is this message spam?"""

Undecided carries Laya's raw answer in .answer, so you can hand the case to a person or a larger model. In .batch(...), the first undecided input raises it.

is_scam.question shows the exact question sent to Laya.

Accuracy

daaf adds no intelligence: answers are as good as Laya's zero-shot checkpoints, which can make mistakes, especially on subtle questions. Test on your own examples, and use min_confidence where an error costs something. Clear docstrings and Enum descriptions help.

License

Apache-2.0, like Laya.

Tests

python test_daaf.py   # no model needed, a fake router stands in for Laya