Skip to content

Mistral AI

To use Mistral AI, 2 ways:

Configure your .env with Mistral API key

# Mistral API
MISTRAL_API_KEY=your-api-key

Azure Mistral API

# Azure Mistral API
MISTRAL_AZURE_API_BASE=<your-azure-api-endpoint>
MISTRAL_AZURE_API_KEY=<your-azure-api-key>

Coverage

For the assessment, mistral models have been tested on the most challenging feature: "ANSWER_USER" on the Mojodex platform.

See blog post for a better understanding of the challenge: Advanced Prompting Strategies for Digital Assistant Development

mistral-large ✅

🎉 full coverage

mistral-large has been tested on all features and provides a full coverage of the Mojodex cognitive functions.

mistral-medium 🚧

The results are promising, but the full coverage is not yet available.

Some results so far

Overall mistral-medium gets the idea behind the data/prompts/tasks/run.txt with the following limitations observed:

  • 👍 Straightforward implementation of the API
  • 👍 Follows the overall logic of the prompt
  • 😕 Big variations between runs in interpretation
  • 😕 Not reliably steerable to defined in the prompt

Work in progress

  • Prepare the Mistral AI provider implementation
  • Run basic tests with existing prompts
  • Complete the assessment on all features for mistral-large
  • Complete the assessment on all features for mistral-medium
  • Prompt Tuning: mojodex_core/prompts/* has been crafted for GPT-4 and GPT-4 Turbo. We need to craft prompts for Mistral AI more specifically.
  • Explore guardrailing strategies to steer the model to the desired output
  • For non-supported features, we can explore fine-tuning the model on our own data gathered in /data/prompts_dataset