v3.0.0

🚀 Features / Enhancements

• new llama model
• add lm_eval extension
• add fine-tuning with deployments

See 📚 Examples for more information.

🐛 Bug Fixes

• huggingface: fix not working default tool
• llama-index: avoid batching in embeddings

⚠️ Warning

We have removed deprecations introduced after v2 release. See the V3 Migration Guide for detailed information.

Full Changelog: v2.3.0…v3.0.0

v2.3.0

What's Changed

• Add Text Experimental module (see client.text.experimental)
  • Sentence Similarity (client.text.sentence_similarity).
  • Rerank (client.text.experimental.rerank).
  • Classification (client.text.experimental.classification).
• Add Folder module (client.folder).
• Add Tags module (client.tag).
• Add Tasks module (client.task).
• Add Request Feedback Service (client.request.feedback).
• Update Moderations.
  • Add Social Bias model.
  • Remove Implicit Hate and Stigma model.
• Add tune creation from an existing file (client.tune.create_from_file).
• Allow to upload arbitrary files (remove old constraints for json files only).
• Add support to update the file content (see client.file.update method).
• Add support for the latest LangChain / LLamaIndex (migrate to 0.10.x) / Transformers versions.
• Unify schemas between various services.

Documentation: https://ibm.github.io/ibm-generative-ai/v2.3.0/index.html

Full Changelog: v2.2.0...v2.3.0

v2.2.0

What's Changed

• fix: correct typo in url by @SOTAkkkk in #310
• feat: improve types generation by @Tomas2D in #312
• docs(langchain): add sql agent example by @Tomas2D in #314
• feat: less strict typings by @David-Kristek in #315
• feat(llama-index): add embeddings by @David-Kristek in #316
• fix: allow the remaining limit to have a negative value by @Tomas2D in #317
• typo: fix bam api url in documentation by @adi-g15-ibm in #319
• chore: fixes and updates by @Tomas2D in #318
• fix: improve http error handling by @Tomas2D in #320
• docs: add simple text generation example by @Tomas2D in #323

New Contributors

• @SOTAkkkk made their first contribution in #310
• @adi-g15-ibm made their first contribution in #319

Full Changelog: v2.1.1...v2.2.0

v2.1.1

What's Changed

• fix: point readme documentation links to latest version by @jezekra1 in #306
• fix: make SharedResource threadsafe by @jezekra1 in #307
• feat(langchain): validate peer dependency by @Tomas2D in #308
• docs: update changelog by @Tomas2D in #309

Full Changelog: v2.1.0...v2.1.1

v2.1.0

🚀 Features / Enhancements

• Add new System Prompts module (example)
  Add LangChain Embedding Function and extend embedding service with truncation parameter (general example, example for ChromaDB)
  LangChain - LLMInterface construct now also accepts dictionaries next to Pydantic Models.
• Add an example for creating LangChain Agent with Tools ( llama-2-70b-chat model) (example)
• Add an ability to inspect service method for metadata (for instance, to retrieve underlying endpoint) (example)
• Add support for the latest LangChain / LLamaIndex / Transformers version.

🐛 Bug Fixes

• LangChain - handles generation/streaming with custom prompt_id and data (example)
• Improve handling batching for large payloads (tokenization)
• Improve handling concurrency limits (text generation / embeddings)

🔧 General Changes

• Schemas are now exported from genai.schema (the old way of importing remains to work, but you will receive a warning; all examples are updated)
• Updated Documentation
• Added a version selector located in the left sidebar
• Added a copy button for examples
• Added the Changelog page (contains grouped list of commits and list of used API endpoints)

⬆️ How to upgrade?

Run pip install ibm-generative-ai --upgrade

📚 Documentation

v2.0.0

On November 22nd, 2023, the API (v2) was announced. We reflected this change on the Python SDK by rewriting its core to be faster, more reliable and mainly in sync with the API. The new SDK brings the concept of the central client, which gives you access to the API very straightforward. This concept was recently integrated into OpenAI SDK / Cohere SDK, and more are joining.

To seamlessly migrate from V0.X versions to 2.0, we have prepared the Migration Guide. The reborn documentation with a lot of examples will help you get started.

Here is a little sneak peek.

• Very Performant.
• Generated Typings directly from the API.
• Smart Requests Concurrency Handling.
• Retry Mechanism in case of network or API failure.
• Batching Large Requests automatically.
• Easy to extend.

Full Changelog: v0.6.1...v2.0.0

v0.6.1

What's Changed

• fix: correct llama-index import for new version by @David-Kristek in #243
• fix(examples): correct Hugging Face example prompt by @David-Kristek in #244
• fix: prevent duplicating template with same name by @Tomas2D in #245

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

v0.6.0

What's Changed

• feat(extensions): add support for llamaindex by @David-Kristek in #238
• fix: update aiohttp to support python 3.12 by @Tomas2D in #239
• fix: add missing init.py in package to fix broken import by @jezekra1 in #241
• fix: update maximal local concurrency limit based on API response by @Tomas2D in #242

New Contributors

• @jezekra1 made their first contribution in #241

Full Changelog: v0.5.1...v0.5.2

v0.5.1

🐛 Bug fixes

• Add missing rate-limit check for tokenize methods
• Unify error messages between sync and async methods

Full Changelog: v0.5.0...v0.5.1

v0.5.0

🚀 Features / Enhancements

• Added integration for LangChain Chat Models; see an example of generation and streaming.
• Added support for LangChain Model Serialization (saving and loading models); see an example.
• Added support for the Chat endpoint in Model class; see an example.
• Added support for new moderation models (HAP, STIGMA, Implicit Hate) - not released on API yet but will be available soon.
• Added type validation for input_tokens property in generate response.
• Extend LangChain generation information / LLM Output (token_usage structure, generated tokens, stop_reason, conversation_id, created_at, ...).
• Add optional raw_response=True/False parameter to generate_stream / generate_as_complete and generate methods to receive a raw response instead of unwrapped results.

🐛 Bug fixes

• LangChain extension now correctly tokenizes the inputs (previously, the GPT2 tokenizer had been used).
• Improve general error handling.

Full Changelog: v0.4.1...v0.5.0