sefia 0.1.0

Define LLM-powered steps as typed async Python functions.

sefia

Stateless Engraved Function Inference Abstraction

sefia turns typed Python functions into durable, replayable LLM-backed calls. An @infer function is an abstract method whose implementer is an LLM: the signature is the input contract, the return type is the validated output contract, the docstring is the instruction, and the body is ....

from pydantic import BaseModel
from sefia import infer


class Summary(BaseModel):
    key_points: list[str]
    uncertainty: str


@infer
async def summarize(article: str) -> Summary:
    """Summarize the article for a technical audience; note key uncertainty."""
    ...

Because model and tool steps replay on re-invocation, a call can pause, resume after a restart, and drive human-in-the-loop flows over ordinary request/response handlers — with no workflow engine or graph DSL.

Install

pip install sefia

sefia is the core framework: decorators, session, profiles, and the tool system. It is model-provider-agnostic; to talk to an actual LLM provider, add an adapter such as sefia-litellm, or install the batteries-included stack sefios.

Documentation

See the repository for the full README, tutorial, and architecture docs.

Status

Early development. APIs may change before v1.0.

License

MIT