v0.3.1

Feature

• Migrate from Granite 3 to Granite 4 hybrid models (#357) (8f9e18c)
• Add MelleaTool.from_smolagents() for smolagents integration (#430) (0471006)
• Add tool calling argument validation (#364) (840a02d)
• Instrument telemetry (#355) (b2e5a52)
• Add query clarification RAG intrinsic support (#391) (d38698a)
• Add mellea react agent (#402) (7884b8d)
• Optimize example test discovery and execution speed (#372) (e9aefaf)
• New MelleaTool class and adoption across mellea (#380) (ffb8b6c)
• Add code coverage tracking with pytest-cov (#353) (b45a4b6)
• Add pytest markers for test categorization (#322) (#326) (0d8d020)

Fix

• Lint/format issues (#536) (781bb6b)
• Tools in examples (#535) (a49bdf8)
• Quick fix to get the role / content from specifically parsed messages (#533) (2f54cc8)
• Migrate from IBM alora to PEFT 0.18.1 native aLoRA (#422) (c6a3e64)
• Flag more tests that require ollama (#420) (b06851f)
• Guarantee proper ordering of decompose subtask dependencies (#407) (f0b1346)
• Astream output (#358) (9cafe05)
• Update ci for merge-queue (#417) (5cf8eee)
• Some examples needed update (#408) (3d5ab56)
• Formatting model_ids for better readability (#386) (318a962)
• Update agents.md to strongly encourage using uv (#388) (8b2e2cf)
• Restrict transformers version to 4.x (#379) (67f8bc0)
• Friendly error messages for optional backend dependencies (#343) (4f7091f)
• Add missing await keywords in async tests (#346) (a7442a6)
• Use repr for helpful debug display in Message/ToolMessage (#339) (f15fadb)
• Add skip to timeout test for python < 3.11 (#333) (2cc3352)
• Don't overwrite user-configured logging levels (#298) (119ea86)

Documentation

• Bedrock example. (#410) (3204b3a)
• Add decompose to tutorial with example (#366) (ef0a964)
• Create contributing doc (#369) (1cacbf9)
• Add security policy (#363) (afbda1d)
• Add code of conduct (#365) (94b21d9)
• Add discord badge to readme (#362) (168ccca)

Performance

• Use module-scoped fixture for RAG tests (#337) (d72ecfd)

v0.3.0

Feature

• SOFAI Sampling Strategy (#311) (cbf3913)
• Reorg of codebase (#310) (cbc456b)
• Add typed components; add typing to model output thunks and sampling results (#300) (2eb689d)

Fix

• Tool calling code sample in tutorial (#313) (a42a487)
• Adds granite-common[transformers] to Mellea's huggingface depedency group. (#330) (87a8166)
• Rename file from test_* (#332) (6512b32)
• Readd init file for mellea/stdlib (#328) (cb156a7)
• ImageBlocks are CBlocks (#323) (8a4c910)
• Additional tests optimization when running on github actions. (#293) (c5398e4)
• Import times by not exporting RichDocument at module level (#321) (565b27f)
• Add logging for start_session details (#299) (6e68f57)
• Typos in READMEs and documentation (#303) (9f6a086)
• Add explicit exports to init.py (#317) (6e7b09b)
• Mify protocol issues (#304) (7013b04)
• Linting error (#302) (c6e3b08)
• Add double quotes around brackets used in pip install (#301) (2d017f1)

Documentation

• Add AGENTS.md to guide AI coding assistants (#320) (a89256a)
• Improve contributor instructions in README. (#314) (2be67c8)

v0.2.4

Fix

• Fix gc in instructions and add exception to generate walk (#295) (5fc7df0)
• Marks span tests as qualitative & removes chat error message. (#294) (5ce6360)

v0.2.3

Feature

• Allow forcing a release through test failures (#292) (14b55a3)
• Lazy Spans and KV Blocks (#249) (b9e4a33)
• Switch to new RAG intrinsics repo (#289) (94c35ad)

Fix

• OpenAI base_url default and reasoning effort model option. (#271) (9733df8)
• Unpin granite_commons version from 0.3.5 (#287) (0b402bd)

v0.2.2

Feature

• Add langchain / message interop example (#257) (9b1f299)
• Add better error messages for incorrect genslot args (#248) (9d875d6)

Fix

• Uv-lock package changes (#261) (cb0623f)
• Lock granite-common version to avoid arg changes (#260) (03716c1)
• Docstrings to have code blocks (#256) (94a7b40)

v0.2.1

Feature

• Test-based Evaluation with LLM-as-a-judge (#225) (0f1f0f8)
• Add a code_interpreter tool (#232) (b03c964)

Fix

• Add simple lock to hf generation to prevent using incorrect weights (#237) (6b2a527)
• Collection of small fixes (#238) (2120112)
• Fix unused litellm import (#246) (633bfd7)
• Minor updates to answer relevance (#245) (bde9b4d)
• Pre-commit file selection (#243) (e70d307)

Documentation

• Fixed copyright in LICENSE (#210) (3087051)

v0.2.0

Feature

• Change backend functions to use async; add generate_from_raw (16b8aea)
• Updates for intrinsics support (#227) (52953a5)
• Add requirements and preconditions to gen slots (#226) (f73d8e2)
• MelleaSession.register for functional interface and MelleaSession.powerup for dynamic mixin (register all methods in a class) (#224) (662cfcc)
• Add secure Python code execution with llm-sandbox support (#217) (9d12458)
• Adds think budget-forcing (#107) (a2e29e6)
• Making generate_from_raw public (#219) (7eae224)
• Conda/Mamba-based installation script (#138) (6aea9dc)
• Adds a vllm backend (#122) (21908e5)
• Add the ability to run examples with pytest (#198) (e30afe6)
• Ollama generate_from_raw uses existing event loop (#204) (36a069f)

Fix

• Vllm format issues (abbde23)
• Some minor fixes (#223) (7fa0891)
• Watsonx self._project_id not getting set (#220) (10f6ffa)
• Decomp subtask regex (#218) (5ac34be)

Documentation

• Adding pii m serve example (#215) (54f13f4)

v0.1.3

Feature

• Decompose cli tool enhancements & new prompt_modules (#170) (b8fc8e1)
• Add async functions (#169) (689e1a9)
• Add Granite Guardian 3.3 8B with updated examples function call validation and repair with reason. (#167) (517e9c5)
• Majority voting sampling strategy (#142) (36eaca4)

Fix

• Fix vllm install script (#185) (abcf622)
• Watsonx and litellm parameter filtering (#187) (793844c)
• Pin trl to version 0.19.1 to avoid deprecation (#202) (9948907)
• Rename format argument in internal methods for better mypiability (#172) (7a6f780)
• Async overhaul; create global event loop; add client cache (#186) (1e236dd)
• Update readme and other places with granite model and tweaks (#184) (519a35a)

v0.1.2

Feature

• Making Granite 4 the default model (#178) (545c1b3)

Fix

• Default sampling strats to None for query, transform, chat (#179) (c8d4601)
• Docstrings (#177) (6126bd9)
• Always call sample when a strategy is provided (#176) (8fece40)

v0.1.1

Fix

• Bump patch version to allow publishing (#175) (cf7a24b)