✨ New features and improvements

• NEW: More accurate Chain-of-Thought (CoT) NER prompting with spacy.NER.v3(#180)
• NEW: Task-specific component factories for llm_ner , llm_spancat, llm_rel,llm_textcat , llm_sentiment, llm_summarization(#243, #283)
• NEW: Implementation of add_label functionality to more easily work with an llm component directly in Python code and not (necessarily) through the config system (#277)
• New v2 model versions for the OpenAI models that set reasonable defaults for temperature and max_tokens (#236)
• Functionality to ignore occassional blocking errors from Cohere (#233)
• Support for Pydantic v1 and v2 (#261, #275)
• Internal refactoring, including renaming of v1 Jinja templates (#242)
• Empty the cache of torch.cuda in between calls (#242)
• Various improvements to the test suite and CI

🔴 Bug fixes

• Fix Anthropic chat endpoints (#230)

⚠️ Backwards incompatibilities

• Though significant refactoring of internal modules has happened, this release should not introduce any backwards incompatibilities for user-facing functionality.
• Check our migration guide if you want to update the SpanCat or NER task from v1 or v2 to v3.

📖 Documentation and examples

• Updated usage documentation
• Updated API documentation
• New Chain-of-Though example with GPT 3.5
• New textcat example with Dolly

👥 Contributors

@adrianeboyd, @honnibal, @ines, @kabirkhan, @ljvmiranda921, @rmitsch, @svlandeg, @victorialslocum, @vinbo8

✨ New features and improvements

• NEW: Support for LLama 2 via Huggingface (#225)
• NEW: Support for Anthropic's Claude 2 (#231)

🔴 Bug fixes

• Corrects how runtime arguments are passed to Falcon (#222)
• Fixes incorrect capitalization in example in readme (#227)

📖 Documentation and examples

We've moved most of our documentation to the main spaCy docs:

• Usage pages: https://spacy.io/usage/large-language-models
• API pages: https://spacy.io/api/large-language-models

👥 Contributors

@honnibal, @ines, @koaning, @rmitsch, @svlandeg, @victorialslocum

🔴 Bug fixes

• Guard REL against entity indices higher than the total number of entities available (#219)
• Warn instead of raising an error when API credentials are wrong for REST-based models (#218)

👥 Contributors

@honnibal, @ines, @rmitsch, @svlandeg

✨ New features and improvements

• Verify authentication details at init (instead of at run) time for Anthropic and Cohere models (#206)

🔴 Bug fixes

• Update OpenLLaMA model names after updates on HuggingFace (#209)
• Fix incorrectly spelled model names for OpenAI models in migration guide (#210)

👥 Contributors

@honnibal, @ines, @rmitsch, @svlandeg

✨ New features and improvements

• NEW: Refactored to transition from backend- to model-centric architecture. Note: this is breaking, you'll need to adjust your configs (#176)
• NEW: Support for Falcon via HuggingFace (#179)
• NEW: Extract prompt examples from component initialization (#163)
• NEW: Summary task spacy.Summary.v1 (#181)
• NEW: Sentiment analysis task spacy.Sentiment.v1 (#200)
• More thorough check for label inconsistencies in span-related tasks NER, REL, SpanCat, TextCat (#183)
• Update langchain pin (#196)
• Make fewshot file reader more robust w.r.t. file formats (#184)

⚠️ Backwards incompatibilities

• Built-in support for MiniChain was dropped (#176)
• The switch from a backend- to a model-centric architecture (#176) requires light adjustments in your config. Check out the migration guide to see how to update your config

👥 Contributors

@bdura, @honnibal, @ines, @kabirkhan, @koaning, @rmitsch, @shadeMe, @svlandeg, @vin-ivar

🔴 Bug fixes

• Use doc._context to ensure that nlp.pipe(..., as_tuples=True) works (#188)
• Fix issue with caching that prevented last doc in cache batch being cached with their LLM IO data (i. e. raw prompt and LLM response) (#191)

👥 Contributors

@honnibal, @ines, @kabirkhan, @rmitsch

✨ New features and improvements

• Make type validation optional with the new validate_types flag (#178)

🔴 Bug fixes

• Fixed nlp.pipe_labels() not working for the llm component (#175)

👥 Contributors

@honnibal, @ines, @kabirkhan, @rmitsch

✨ New features and improvements

• NEW: Optional storing of prompts and responses in Doc objects (#127)
• NEW: Optional logging of prompts and responses (#80)
• NEW: Streamlit demo (#102)
• NEW: Support for Cohere in backend spacy.REST.v1 (#165)
• NEW: Support for Anthropic in backend spacy.REST.v1 (#157)
• NEW: Support for OpenLLaMA via HuggingFace with the backend spacy.OpenLLaMa_HF.v1 (#151)
• NEW: Support for StableLM via HuggingFace with the backend spacy.StableLM_HF.v1 (#141)
• NEW: Lemmatization task spacy.Lemma.v1 (#164)

🔴 Bug fixes

• Fix bug with sending empty prompts if all Doc objects are cached (#166)
• Fix issue with LangChain model creation due to updated argument name (#162)

👥 Contributors

@adrianeboyd, @bdura, @honnibal, @ines, @kabirkhan, @ljvmiranda921, @rmitsch, @svlandeg, @victorialslocum, @vin-ivar

✨ New features and improvements

• NEW: llm component supports scoring, like other spaCy components (#135)
• Labels for spacy.NER.v2, spacy.REL.v1, spacy.SpanCat.v2, spacy.TextCat.v2 can be specified as list (#137)

🔴 Bug fixes

• Fix type comparison in type checks failing on some platforms (#158)
• Fix example 3 in readme failing (#137)

👥 Contributors

@adrianeboyd, @bdura, @honnibal, @ines, @kabirkhan, @KennethEnevoldsen, @rmitsch, @svlandeg, @vin-ivar

✨ New features and improvements

• NEW: New relation extraction task spacy.REL.v1 (#114)
• NEW: New spancat task spacy.SpanCat.v1 for entity recognition with overlapping spans (#101)
• NEW: Set prompt templates as string or read them from files (e. g. Jinja templates) using spacy.FileReader.v1 (#95)
• Improved prompt for NER task spacy.NER.v2 (#99)
• Ability to describe labels in tasks spacy.NER.v2, spacy.SpanCat.v1 (#84)
• Improved error handling and retry mechanics in REST backend spacy.REST.v1 (#110)
• spacy-llm can now (#119) be installed with all dependencies for
  • MiniChain and LangChain with spacy-llm[minichain] or spacy-llm[langchain] respectively
    for locally running models on GPU with spacy-llm[transformers]
  • Improved type checking to ensure tasks and backend play nice with each other (#83)
• Use Task abstraction instead of previously two functions template and parse to make the config easier to read (#91)

🔴 Bug fixes

• Fix improper doc identity check in caching (#104)
• Fix multiprocessing support for .pipe() (#117)

📖 Documentation and examples

• Updated the readme with information about the caching mechanism.
• New usage examples:
  • Performing multiple tasks in a single pipeline
  • Relation extraction using LLMs
  • Using OpenAI models with LangChain
  • Using OpenAI models with MiniChain

👥 Contributors

@adrianeboyd, @bdura, @honnibal, @ines, @kabirkhan, @ljvmiranda921, @rmitsch, @svlandeg