v0.11.1: Fix engine import; upgrade to Outlines v1.x; temporarily disable vLLM

✨ New features and improvements

• Migrate to support Outlines >= 1.0 (#146)

🔴 Bug fixes

• Resolve #145: direct engine object import didn't work (#146)

⚠️ Backwards incompatibilities

• vllm support has been temporarily disabled (#146). This is because vllm currently doesn't support outlines >= 1.0. Will be re-enabled once vllm migrates to outlines v1.

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.11.0...v0.11.1

v0.11.0: Distillation, caching, making engine dependencies optional

✨ New features and improvements

• New Distillation task that allows to distill a local model using setfit or model2vec. Currently only implemented for the classification task. (#138)
• Caching support: Pipeline objects now optionally cache documents to ensure that no document is processed twice unnecessarily. (#140)
• Missing third-party dependencies (like vllm, instructor, outlines, ...) at import time are now allowed. This enables running sieves with the engine of your choice without having to have all engine dependencies allowed. If you still want to install all dependencies, install with pip install sieves[all]. (#137)

🔴 Bug fixes

-

⚠️ Backwards incompatibilities

• Engine dependencies (except for outlines, which is used as a default engine) are not installed by default. If you want to install all engine dependencies, install with pip install sieves[all].

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.10.0...v0.11.0

v0.10.0: vLLM, Hugging Face datasets conversion, single-label classification

✨ New features and improvements

• vLLM support (#134)
• Add Doc.from_hf_dataset() to generate Doc instances from HF datasets (#130)
• Add single-label prediction mode for Classification task (#132, #134)
• Simplified interface for chunking tasks (#125)
• Speed up test suite (#121)
• Set default model for Engine (#129)

🔴 Bug fixes

-

⚠️ Backwards incompatibilities

-

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.9.0...v0.10.0

v0.9.0: OCR simplification, NER task

✨ New features and improvements

• Unify tasks for doing OCR (Docling, Marker) in single OCR task (#111, #114)
• Add task for OCR with Marker (#106)
• Extend classification task to support label descriptions (#114)
• Speed up test suite (#121)

🔴 Bug fixes

-

⚠️ Backwards incompatibilities

-

📖 Documentation and examples

-

👥 Contributors

@adgianv, @rmitsch

---

Full Changelog: v0.8.0...v0.9.0

v0.8.0: PII masking, API simplifications, doc fixes

✨ New features and improvements

• Add task for masking personally identifiable information (#101)
• Simplify usage of engines: no need to use specific engines like Outlines anymore, instead the new Engine class deals with this based on the passed model class (#108)
• Simplify config and usage of GliX, Ollama, Instructor engines (#100)

🔴 Bug fixes

• Fix errors in documentation (#93, #94)

⚠️ Backwards incompatibilities

• PredictiveTask instances now expect an engine of type Engine instead of GliXEngine | HuggingFaceEngine | ....

📖 Documentation and examples

-

👥 Contributors

@rmitsch, @nsorros

---

Full Changelog: v0.7.0...v0.8.0

v0.7.0: Sentiment analysis, installation fixes

✨ New features and improvements

• Add sentiment analysis task (#90)

🔴 Bug fixes

• Fix bug in fetching version in Config (#90)

⚠️ Backwards incompatibilities

-

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.6.1...v0.7.0

v0.6.1: Fix PyPi package installation

✨ New features and improvements

-

🔴 Bug fixes

• Fixes No module names sieves error on package installation (#88)

⚠️ Backwards incompatibilities

-

📖 Documentation and examples

-

👥 Contributors

@nsorros, @rmitsch

---

Full Changelog: v0.6.0...v0.6.1

v0.6.0: Summarization, translation, multi-question answering

✨ New features and improvements

• Add translation task (#81)
• Add summarization task (#82)
• Add multi-question answering task (#83)
• overwrite flag determining whether original text should be overwritten for tasks with a textual output (e.g. translation) (#81)

🔴 Bug fixes

-

⚠️ Backwards incompatibilities

• Bridge moved from sieves.task.predictive.core into sieves.task.predictive.bridges

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.4.0...v0.5.0

v0.5: Documentation, batching, `instructor` engine

✨ New features and improvements

• Online documentation at https://mantisai.github.io/sieves (#68, #69, #70)
• Add new engine for instructor (#75)
• Add batching support for all engines (#75)
• Standardize task tests (#71)

🔴 Bug fixes

• Fix bug due in DSPy engine due to dspy update (#67)
• Fixed typing for engines and tasks (#65)

⚠️ Backwards incompatibilities

• sieves.engine.core.TemplateBasedEngine is now named sieves.engine.core.PydanticEngine

📖 Documentation and examples

• Online documentation at https://mantisai.github.io/sieves (#68, #69, #70)

👥 Contributors

@rmitsch

---

Full Changelog: v0.4.0...v0.5.0

v0.4.0: Export as HF `datasets.Dataset`; `unstructured` preprocessing task

✨ New features and improvements

• Support for exporting task results in datasets.Dataset format for easy distillation/model training (#63)
• Add new task: preprocessing documents with unstructured (#61)
• Introduced strict mode raising errors on unsuccessful result parsing (#57)
• Use reasoning traces/CoT for existing tasks (#59)
• Simplify serialization implementation for tasks (#61)

🔴 Bug fixes

• Fixed bugs in serialization mechanism (#61)

⚠️ Backwards incompatibilities

• tasks.parsing and tasks.chunkers have been merged into tasks.preprocessing

📖 Documentation and examples

-

👥 Contributors

@rmitsch

---

Full Changelog: v0.3.0...v0.4.0