Openhosta

- **Enhancement**
- Changed examples syntax in the system prompt for better performances.

- **Doc**
- Added an example for a custom Model class with an Llama model.

- **Added**
- Introduced `BaseEncoder` as an Abstract Base Class (ABC) for all specific encoders, providing a standardized interface.
- Added `BoolEncoder` to handle encoding of boolean values.
- Introduced `PredictConfig` data class to encapsulate parameters for the `predict` function.
- Initial implementation of the `HostaCache` class in `core/cache.py` with generic type support.
- Initial implementation of the `ModelCachedData` and `HostaModelCache` classes in `predict/cache.py`.
- **Changed**
- Refactored `IntEncoder` and `FloatEncoder` to inherit from `BaseEncoder` and implement the encode method.
- Updated `HostaEncoder` to use a dictionary (`self.encoders`) for mapping data types to their corresponding encoders. This allows for easier extension and maintenance.
- Improved type handling in `HostaEncoder.encode` method for better extensibility and readability.
- Refactored `predict` function to accept a single `PredictConfig` object instead of multiple parameters. This change improves readability and maintainability of the function.
- **Fixed**
- Enhanced exception handling in `FloatEncoder` to provide more informative error messages.
- Removed unnecessary constructors from encoder classes, streamlining the code.

- **Features**
- Added `max_tokens` args for emulate
- Added `use_locals/self_as_ctx` args for emulate for clarity and modularity
- `thought` is now used to add chain of thoughts in a emulated function
- `PromptManager` is now available for users to change all meta-prompt.

- **Changes**
- `thought` function become `thinkof`
- `suggest` is removed
- `creativity` become `temperature` and `diversity` become `top_p`
- There's no more `hostacache` for emulated functions

- **Fixes**
- OpenHosta now can't be used with Python 3.13 due to the incompatibility with pytorch

- **New Features**

- **`predict` Function**
The `predict` function is now available, allowing you to create internal models—currently supporting linear regression—based on user-provided training data. This simplifies model generation without relying on external APIs. Key functionalities include:
- **`.retrain`**: Retrain models with specified parameters.
- **`.continue_train`**: Continue training with existing model weights.
- **`.emulate`**: Run predictions through an LLM or create a model directly using internal linear regression based on training data.

- **TrainingSet Management**
Manage training datasets effortlessly with new tools:
- **`.visualize`**: Inspect current data visually.
- **`.add`**: Add new examples.

- **Enhancements**

- **Expanded Dataset Support**:
`load_training_example` (previously `load_examples`) supports JSON, JSONL, and CSV formats for easier integration.

- **Verbose Mode in `predict`**:
Track detailed model training and define target losses with `get_loss`.

---

- **Features**
- Added `_last_request` attributs to the `Model` object

- **Optimization**
- Reduce the `emulate`'s user prompt to only the function call string and move the rest in the system prompt.
- Added markdow header in `emulate` prompt and enhance strucure
- Remove unecessary setence like confidence level and transition

- **Internal**
- Refacto the prompt building function in `emulate`
- Remove the validate function in LLM call response handler

- **Fixes**
- Added code block syntax for the function definition.
- Re-added `diagramm` attributs but decrepated
- Added explicitly a neutral response in the `emulate` prompt (None)