December 2024 Summaries

Ollama has introduced support for structured outputs, allowing users to format a model's output according to a predefined JSON schema, enhancing reliability and consistency over traditional JSON mode. Updated libraries for Python and JavaScript support this feature, enabling use cases such as parsing data from documents, extracting data from images, and structuring language model responses. To implement structured outputs, users can set the format parameter in cURL requests or employ Pydantic (Python) and Zod (JavaScript) for schema serialization. Example applications include extracting structured data about pets or analyzing images, with outputs returned as JSON objects matching the schema. The update also features compatibility with OpenAI and tips for optimizing structured outputs, such as setting the model's temperature to zero for deterministic results. Future developments aim to improve performance, introduce GPU acceleration, and expand format support.