Inside every AI-powered pipeline, the unreliable part is the same: the LLM call itself. Silent failures, untestable outputs, no guarantees. Mellea is a Python library for writing generative programs — replacing brittle prompts and flaky agents with structured, testable AI workflows built around type-annotated outputs, verifiable requirements, and automatic retries.

See installation docs for additional options, such as installing all extras via uv pip install 'mellea[all]'. For source installation directly from this repo, see CONTRIBUTING.md.

The @generative decorator turns a typed Python function into a structured LLM call. Docstrings become prompts, type hints become schemas — no templates, no parsers:

from pydantic import BaseModel
from mellea import generative, start_session

class UserProfile(BaseModel):
    name: str
    age: int

@generative
def extract_user(text: str) -> UserProfile:
    """Extract the user's name and age from the text."""

m = start_session()
user = extract_user(m, text="User log 42: Alice is 31 years old.")
print(user.name)  # Alice
print(user.age)   # 31 — always an int, guaranteed by the schema

start_session() is the convenience entry point: it returns a MelleaSession with sensible defaults you can override as needed.

• Structured output — @generative turns typed functions into LLM calls; Pydantic schemas are enforced at generation time
• Requirements & repair — attach natural-language requirements to any call; Mellea validates and retries automatically
• Sampling strategies — run a generation multiple times and pick the best result; swap between rejection sampling, majority voting, and more with one parameter change
• Multiple backends — Ollama, OpenAI, HuggingFace, WatsonX, LiteLLM, Bedrock
• Legacy integration — easily drop Mellea into existing codebases with mify
• MCP compatible — expose any generative program as an MCP tool

We welcome contributions of all kinds — bug fixes, new backends, standard library components, examples, and docs.

• Contributing Guide — development setup, workflow, and coding standards
• Building Extensions — create reusable components in your own repo
• mellea-contribs — community library for shared components

Questions? See GitHub Discussions.

Mellea was started by IBM Research in Cambridge, MA.

Licensed under the Apache-2.0 License. Copyright © 2026 Mellea.