Rémi Louf

**Dependency**: `outlines-core` **Version**: 0.2.3 **Release Notes**: https://api.github.com/repos/dottxt-ai/outlines-core/releases/198329713

**Dependency**: `outlines-core` **Version**: 0.2.2 **Release Notes**: https://api.github.com/repos/dottxt-ai/outlines-core/releases/197858093

**Dependency**: `outlines-core` **Version**: 0.2.1 **Release Notes**: https://api.github.com/repos/dottxt-ai/outlines-core/releases/197643911

**Dependency**: `outlines-core` **Version**: 0.2.0 **Release Notes**: https://api.github.com/repos/dottxt-ai/outlines-core/releases/197639685

Make `generate.json` compatible with OpenAI's function calling format

6

OpenAI uses a [small variation on the JSON Schema specification](https://platform.openai.com/docs/guides/function-calling), and we need to make our interface compatible.

**Dependency**: `outlines-core` **Version**: 0.1.27 **Release Notes**: https://api.github.com/repos/dottxt-ai/outlines-core/releases/190204875

This should apply to the _word_, not the tokens.

I propose to forbid sequences in generated completions using a `not_in` keyword argument to `Completion`. This can be implemented using masking and re-weighing.

Implement prompt/generation alignment

5

[Guidance](https://github.com/microsoft/guidance) implements a method called token healing, which consists in correcting for the quirks introduced by modern encodings like BPE. See [this notebook](https://github.com/microsoft/guidance/blob/main/notebooks/token_healing.ipynb) for a thorough explanation of why this...