Skip to content

CHANGELOG.md

Latest commit

 

History

History
100 lines (62 loc) · 6.83 KB

CHANGELOG.md

File metadata and controls

100 lines (62 loc) · 6.83 KB
Raw

Changelog

All notable changes to this project are documented in this file. Release notes are grouped by theme rather than listing every commit.

[0.3.2] — 2026-06-22

Compared to 0.3.1. A maintenance release focused on caching correctness and CI/test coverage. No breaking changes.

Bug fixes

  • Structured-output cache now keys entries by model identity — the model name and API base_url are folded into the cache key alongside the messages, schema, and generation params, so switching models or endpoints no longer returns another model's cached completion (#336).
  • Structured-output cache eager loading and eviction now treat each on-disk entry as a directory (the layout PydanticModelDumper actually writes). The previous is_file() filter matched nothing and silently disabled eager cache loading; eviction now uses rmtree instead of unlink (#331).
  • Offline embedding cache key is now stable: when HF_HUB_OFFLINE is set, the commit SHA is read from the local Hugging Face ref file ($HF_HUB_CACHE/<repo>/refs/<rev>) so the key matches the online path instead of falling back to the bare revision string. The embedding cache key now also incorporates the model name, preventing cross-model cache collisions for non-local models (#337).
  • Augmentation console scripts basic-aug and evolution-aug now resolve to the correct module paths (autointent.generation.utterances._basic.cli / ._evolution.cli); the entry points were left dangling by an earlier module rename (#330).

Tooling, CI, and tests

  • Coverage regression floor: the combined coverage total is now gated against an 85% floor, so a drop below it fails CI (#333).
  • Manual coverage dispatch workflow added and the broken coverage configuration fixed (#325).
  • Windows test jobs restored — a bash-only shell step had been breaking them on Windows runners (#328).
  • Added test coverage for the Generator dump/load round-trip and the async empty-response guard (#332), the basic/evolution augmentation CLIs (#330), and the JSONFormatter plus macro retrieval metrics (#329).

[0.3.1] — 2026-06-16

Compared to 0.3.0. A maintenance release focused on bug fixes, reproducibility of default Hugging Face downloads, and CI/test stability. No breaking changes.

Bug fixes

  • LLM description scorer now persists generator_config across dump/load, so a reloaded LLMDescriptionScorer rebuilds its Generator with the original settings instead of defaults (#302).
  • Tunable decision threshold test fixture updated for Optuna 4.9 sampling order — confirms compatibility with optuna>=4.9 (#293).
  • faiss typing tweak for IndexFlatIP to satisfy stricter mypy runs (#292).

Reproducible Hugging Face defaults

  • New canonical SHA pin map autointent.configs._pinned_revisions.DEFAULT_REVISIONS, applied automatically by HFModelConfig when revision is left unset. Default embedders, rerankers, and NLI models now resolve to fixed commit SHAs instead of main, making installs deterministic and avoiding Hub 429 rate limits under parallel CI load. To opt out, pass an explicit revision (including "main") on the relevant config (#294).

Public API additions

  • New RetrievalMetricFnWithOOS and ScoringMetricFnWithOOS protocols (exported from autointent.metrics) describing the OOS-aware metric callables returned by the existing ignore_oos decorators (#316).

Tooling, CI, and tests

  • Full OS × Python matrix is now gated behind the full-ci PR label and dev pushes; ordinary PRs run a faster default matrix (#318).
  • mypy --strict is enforced on tests/ to prevent type drift between fixtures and production code (#316).
  • Soft mypy profile added for docs/ and user_guides/ so tutorials and Sphinx helpers are type-checked without forcing them onto strict mode (#320).
  • Live-API tests (OpenAI, OpenSearch) are mocked in CI — no network calls or credentials required to run the suite locally (#301).
  • HF cache prewarm workflow plus a .ci/warm_hf_cache.py helper keyed off DEFAULT_REVISIONS removes the long-standing Hub rate-limit flakes in the matrix (#294).
  • GitHub Actions runners bumped to Node 24 images (#300).
  • Pipeline interruption tests restored after the sampler API removal in #296 (#303).
  • Test layout cleanup: the legacy setup_environment helper is gone — tests now use the standard pytest tmp_path fixture (#317).

[0.3.0] — 2026-05-19

Compared to 0.2.0.

Breaking: slimmer default install and optional extras

Heavy integrations are now optional and loaded lazily where possible. A plain pip install autointent pulls a smaller core set of dependencies; features that need transformers, OpenSearch, MCP, OpenAI, vLLM, and similar stacks require the matching optional dependency groups (see pyproject.toml). If you relied on everything being importable after a minimal install, pin extras explicitly in your environment.

Embeddings: OpenAI, vLLM, and training

  • OpenAI embeddings path with safer batching and tokenizer fallbacks when the API does not expose tokenizer metadata.
  • vLLM optional extra for serving-compatible embedding workflows.
  • Embedder fine-tuning utilities and training-oriented improvements for custom embedding models.

Scoring and search

  • GCN-based scorer for graph-style intent scoring in the AutoML search.
  • OpenSearch optional backend for vector index storage and retrieval alongside existing options.

Serving and integration

  • MCP (Model Context Protocol) interface plus HTTP serving pieces (FastAPI / related extras) for running AutoIntent behind an API or MCP host.

Data, splits, and augmentation

  • Out-of-scope (OOS) intents are always kept separate in dataset splits so evaluation stays honest.
  • Adversarial augmentation support for more robust pipelines under optimization.

Configuration and validation

  • OptimizationConfig.from_preset for constructing optimization settings from named presets.
  • Split readiness checks (including multilabel-aware validation) so you catch bad label or split geometry before long runs.

Tooling and dependency bounds

  • uv-centric developer workflow (dependency groups, reproducible installs).
  • Python supported through below 3.15 (requires-python = ">=3.10,<3.15" in pyproject.toml).
  • OpenAI Python SDK v2 (openai>=2,<3) in the openai extra.
  • datasets pinned below 5 (datasets>=3.2,<5) for compatibility with current loaders and tests.

[0.2.0]

See the v0.2.0 release on GitHub for earlier changes.