Learning Rust gently with Mellum2-12B-A2.5B-Instruct and Mini Me (Part 2)

A few more tools to improve our little teacher

I have a short train trip coming up (and I probably won't have wifi). This upcoming event triggered the preparation and release of Mini Me v0.8.0. I wanted a simple (but effective) document search function that would be a "builtin" tool rather than just using a plain grep.

My goal was the following: provide documentation to my agent to enrich its knowledge. For example, I'm interested in a particular Rust framework, Ratatui, which makes it easier to build TUIs in Rust.

So I added a new tool: search_docs (more info here: explanation/document-search.md) which relies on BM25. It's a ranking function that assigns a score to each document for a given query. It's lexical search: it relies on the terms present in the documents, not on their meaning as represented by embeddings.

You need Mini Me v0.8.0 or later to use the new search_docs builtin tool.

A few changes in our little teacher's settings

I enabled these tools:

bashTool: true
editTools: true

• editTools because the agent will need to read the documentation
• bashTool because I'd like to make the little teacher work a bit harder

Then I enabled LSP support for Rust with rust-analyzer (we'll see a bit further down how I use it):

lsp:
  servers:
    - name: rust
      command: rust-analyzer
      extensions: [".rs"]

Of course, you need a complete Rust toolchain. But if you use the development environment described in the previous blog post "Learning Rust gently with Mellum2-12B-A2.5B-Instruct and Mini Me (Part 1)", and you start it with the commands below, you'll have all the tools you need:

sbx env create --force
zed --new "ssh://zed-mellum.sbx$PWD"

I updated the sbxenv.yaml file to take the newly released features into account.

Warning: these settings work with Mellum2-12B-A2.5B-Instruct-GGUF-Q8_0, and they probably won't work the same way (or even at all) with other models. That's why, when you look for and select a model, you'll always need an adaptation period (more or less long) to find the right settings and check whether the model will work for the use case you have in mind.

Next, I added documentation (in the ./docs folder) about the Ratatui library, which makes it easier to build TUIs in Rust. This documentation is a set of markdown files to help get started with Ratatui, which I asked Claude to generate.

Finally, I added the section below to agent.yaml, to enable the new search_docs tool:

docSearch:
  autoSearchDocs: true
  dirs: [docs]
  extensions: [.md]
  maxFileSize: 1MB

And we're ready to check whether our little teacher has gained a few extra skills.

Let's ask a bit more of our little teacher

I then launched Mini Me in TUI mode: mm --tui, and asked it the following:

> search how to ceate a rust project with ratatui, then initialize a hello-ratatui cli project

So, the agent did trigger the search_docs tool, found the information and generated a new lesson explaining how to initialize the project, but didn't create anything on disk. So I asked this:

> follow these instructions to create the program

And this time the agent used its own lesson to create the project. I was then able to check that the code worked by running cargo run in the project folder.

My next test was to ask the agent to use rust-analyzer (Rust's LSP tool) to give me information about the source code:

> use rust-analyzer to explain the hello-ratatui

And it was conclusive, the agent was able to generate a report about the source code:

Of course, since Zed Editor has an ACP client, you can use our little teacher with a GUI:

And there you go! "TUI development in Rust for dummies" on the train for me, even without wifi. See you soon for new adventures.

By the way, you'll find a guide to the keyboard shortcuts of the Mini Me TUI here: Keyboard shortcuts of the window (mm -tui)

Source code: https://rickub.com/bots-garden/model-mellum