The update also includes regex validation for easier output validation without handling parsing, and surreal database integration for traditional database use cases. RAG (Retrieval-Augmented Generation) has been improved with better chunking strategies and incremental indexing. Performance improvements have been made with the llama implementation being rewritten and optimized, and several new models have been added. Future plans include adding support for fine-tuning models and training new heads for existing models, and improving the performance of the language models and adding support for more models.
Key takeaways:
- Kalosm v0.2.0 has been released with new features such as Tasks and Agents, Task Evaluation, Prompt Auto-Tuning, Regex Validation, and Surreal Database Integration.
- Improvements have been made to the RAG (Retrieval-Augmented Generation) system, including new chunking strategies and incremental indexing.
- Performance has been optimized, with the new implementation being 7-25% faster than the previous version and support for batched loading in constrained generation.
- The release also includes support for several new models, including Dolphin Phi v2, Solar-11b Models, and Tiny Llama 1.0.