Structured Outputs From Mistral’s Models

Summary

The author is exploring Mistral's models on "La Plateforme," noting their proficiency in generating structured outputs and the potential for integration into complex workflows, despite current limitations in documentation and token consumption monitoring.

Abstract

The author has recently gained access to "La Plateforme" and has begun experimenting with Mistral's models. They have made an initial observation that the models perform well, although "La Plateforme" is still in development, lacking detailed documentation and the ability to monitor token consumption, despite having a feature to set a maximum budget. The author has successfully tested the models' capability to produce structured outputs by using pydantic to define the output schema. They acknowledge that while the initial version of the code works, it will require refinement in the coming days and weeks. The author is optimistic about the future improvements of "La Plateforme" and believes that Mistral's models will become increasingly useful for creating sophisticated workflows.

Opinions

The author finds Mistral's models to be of high quality and useful for generating structured outputs.
There is a recognition that "La Plateforme" is not yet fully evolved, with specific mention of the need for better token consumption monitoring and more detailed documentation.
The author is confident that the current shortcomings of "La Plateforme" will be addressed in the near future, enhancing its utility.
The author has a positive outlook on the potential of Mistral's models within complex workflows, indicating a belief in their versatility and applicability to advanced use cases.

Structured Outputs From Mistral’s Models

I just got access to “La Plateforme” and immediately started playing with Mistral’s models.

First observation: These models are pretty good. Now “La Plateforme” still needs to evolve. For now, I can’t monitor my token consumption, even if I can already set a max budget. Plus the docs do not contain enough details about the models. I have no doubt this will be different in a few months.

One thing I wanted to test tonight is the ability of Mistral’s models to generate structured outputs. I quickly wrote some lines of code using pydantic to define the schema of the structured output. Here is a quick first version of the code. I will certainly have to make it better in the following days and weeks. Turns out Mistral’s models are pretty good at generating structured outputs, which makes them very useful to develop complex workflows.