We released Mellea 0.4.0 with three Granite libraries: granitelib-rag-r1.0, granitelib-core-r1.0, granitelib-guardian-r1.0. Together, these releases make it easy to build structured, verifiable, and safety-aware AI workflows on top of IBM Granite models.
Mellea is an open-source Python library for writing generative programs that replaces probabilistic prompting behavior with structured, maintainable AI workflows. Unlike general-purpose orchestration frameworks, Mellea is designed to make LLM-based programs maintainable and predictable through constrained decoding, structured repair loops, and composable pipelines (new to Mellea? Start with our introductory blog and meet the team)
Melea 0.4.0
Mellea 0.4.0 is the latest release of an open source research project initiated and developed by IBM Research. Building on the base library and workflow primitives of 0.3.0, 0.4.0 extends the integration aspects of the library and introduces new architectural patterns for building generative workflows.
Includes:
Native integration with Granite library. Provides a standardized API that relies on constrained decoding to ensure schema accuracy. Event-driven callback observability hooks to monitor and track instruction, validation, and remediation pattern workflows with rejection sampling strategies
See the complete list of Mellea 0.4.0 features and updates here
What is the Granite library?
Simply put, the Granite library is a collection of specialized model adapters designed to perform well-defined operations on parts of an input chain or conversation. Rather than relying on generic prompts, each specialized model is fine-tuned for a specific task, such as rewriting queries, detecting hallucinations, or checking policy compliance. Specialized adapters allow you to improve the accuracy of each task without disrupting the functionality of the base model and at the cost of a moderate number of parameters.
Three libraries were released today for the granite-4.0-micro model. Each library targets a distinct set of pipeline tasks and consists of a collection of LoRA adapters.
Granitelib-core-r1.0: Targets the requirements validation step of Mellea’s instruct-validate-repair loop Granitelib-rag-r1.0: Targets various tasks in the agent RAG pipeline, covering pre-capture, post-capture, and post-generation Granitelib-guardian-r1.0: Specialized model for safety, factuality, and policy compliance
Mellea 0.4.0 is the latest release of an open source research project initiated and developed by IBM Research. This release builds on the basic libraries and workflow primitives introduced in 0.3.0, expands on the integration front, and introduces new architectural patterns for building generative workflows.
Start
Melea
granite library
learn more
