# Mellea

> A composable, multi-agent framework for building reliable, predictable AI agents at production scale.

Mellea is an open-source Python library from IBM Research that brings software engineering rigour to generative AI. Rather than treating LLMs as human-like entities to be prompted, Mellea treats them as computational engines — composable, testable, and verifiable.

## Core ideas

- **Generative Functions**: Python functions whose implementation is provided by an LLM, with typed inputs and outputs
- **Constrained Decoding**: Guaranteed structured output — the model cannot produce invalid responses
- **MObjects (Mellea Objects)**: Typed data structures that flow through generative pipelines
- **Requirements-Driven**: Instruct, validate, and repair — define what correct output looks like, not just what to ask
- **Unit Testable**: Generative code can be unit tested like any other Python code
- **OpenTelemetry Native**: Full observability with traces, metrics, and logs using Gen-AI semantic conventions

## Why it exists

Current agent frameworks push developers toward fragile prompt engineering and anthropomorphic agent designs. Mellea provides a programming model where LLMs are woven into traditional software seamlessly — making AI agents as predictable and maintainable as the rest of your codebase.

## Key capabilities

- Any LLM provider (via LiteLLM, direct APIs, or local models including watsonx)
- Safety and guardrails (Granite Guardian, jailbreak prevention, hallucination detection)
- Async and streaming support
- Multimodal / vision model support
- REST API deployment via m_serve
- SoFAI: fast/slow two-tier model sampling for cost optimisation
- RAG pipeline support

## Links

- Docs: https://docs.mellea.ai
- GitHub: https://github.com/generative-computing/mellea
- Blog: https://mellea.ai/blogs
- PyPI: https://pypi.org/project/mellea/