MagisTral

Leverage Mistral AI functions and multi-agent conversations to enable interoperable applications and support a wide range of uses-cases

Comment

Inspiration

  • the project that won the previous Mistral SF hackathon: a prompt engineering assistant powered by AI, l'Invite Parfaite: if it has been done for prompts, it can be done for implementing Mistral functions too.
  • Wilfred's previous job at Samsung, where he was in charge of connecting our IoT cloud to multiple external APIs.

What it does

  • Magistral Wizard: an universal external APIs to Mistral AI functions converter, it takes those API specifications (Swagger, GraphQL, custom format) or documentation (with NLU) as input and produces the tools logic for Mistral AI. We wanted to be impactful for Mistral and the innovation ecosystem.
  • Magistral Multi-Agent conversation : tailored to support a wide range of use-cases, including calling external APIs thanks to Magistral Wizard. The scoped agent discussion helps to preserve data privacy, leakage and biases

How we built it

  • Python
  • Mistral AI client API
  • NiceGUI library
  • Prompt-Engineering

Challenges we ran into

  • The solutions provided by AI are not 100% perfect
  • Short-time

Accomplishments that we're proud of

  • First working prototype

What we learned

  • Mistral AI functions
  • Prompt-engineering
  • During the hackathons : AWS SageMaker, Weight & Biases, and HuggingFace workshops

What's next for MagisTral

  • Integrate it natively in Mistral.AI "La Plateforme"
  • Experiment converting Amazon Alexa Voice Skills
  • Leverage Snowflake for the multi-agent discussion

Other idea we had

  • Integrate vision
  • Prune and quantize the Mistral open-source model for targeting mobile inference and (edge) AI accelerator

Built With

  • graphql
  • mistral
  • nicegui
  • python
  • radiofranceapi
Share this project:

Updates