How LLM can benefit AiiDA

[khsrali]

Introduction

LLM tools has been around for a few years now, and they're getting better progressively.
We heard reports from several users, that how LLMs improved their experience with AiiDA, e.g. facilitating writing query builder.

This brought us to take the initiative and go a few steps further.
Below we explored two of such possibilities as proof of concept.

Proof of Concepts

1- verdi-llm
This repo has the full knowledge of AiiDA command line. It was designed as a learning toy for newbies for exploring verdi interface. It's particularly useful, because user only describes in natural language what they'd like to do, and verdi-llm will come with a command resolution, e.g.:

"Delete the working directories of all finished calculation jobs since the last month on cluster1 and cluster2 without asking for confirmation"
Suggests:
verdi calcjob cleanworkdir --past-days 30 --computers cluster1,cluster2 --exit-status 0 --force

In the backend, it uses a RAG database. Mapped and made from the up-to-date developments of verdi interface. This makes even the most outdated and free-to-use LLMs, to return very good results!

2- aiida-mcp
This repo, develops and utilizes an MCP servers for running a phonon workflow in AiiDA. Like verdi-llm is designed to help newbies to walk through several workflow examples that we have. User might simply ask:

"Help me to calculate phonon for silicon using AiiDA"
And it basically does everything:

• Checks if the prerequisites are met, required packages are installed and codes are defined in AiiDA
• If not, it installs them for the user and sets up the codes in AiiDA.
• Asks for all the necessarily inputs, e.g. the structure file, if user doesn't have one it can suggest.
• Writes an python script to run such calculation and submits to AiiDA, it reports back the PK.
• Is able to follow up, find out where the outputs are, parse and answers to questions like: "What's the convergence energy"

Requires Discussion

Having this two proof of concepts, we need to think for a moment.
Going further in this direction requires a team decision.

The main doubt is, a lot of these sorta solutions are age(!) specific. Meaning inmatter of months the MCP protocol might change in a way that will require maintenance. Also "hard-coding" a workflow logic in MCPs might not be the best idea. It can also happen that LLMs which are catching up quickly will be able to do these things on their own at half-a-year time. Which then means our efforts might go wasted.

@giovannipizzi @GeigerJ2 @agoscinski

[ltalirz]

Some thoughts from my side:

docs

One of the main - and I think fair - criticisms of AiiDA are that it is difficult to install and use. There is extensive documentation but it is, well, ... extensive, and there is a lot of info on github issues but... it is a lot to parse.

I believe that in the age of LLMs, the relevance of directly user-facing documentation is decreasing and what matters will be how well the LLM can help you use the program.
This means it matters less how difficult the software is to use for humans, and it matters more how well the software works / how much it can do.

This could be a chance for AiiDA, if you manage to provide the required information to an LLM in a digestible form. It could start by hosting something like a vector database of the docs + relevant issues that can be connected to the LLM the user is using.
The nice thing about this is that you can just keep adding to it - e.g. you discover an issue for in an AiiDA tutorial for how to install AiidA on a particular OS and how to work around it => it goes into the DB.
This approach also allows you to prune the DB of outdated instructions (which LLMs trained on the internet will inevitably provide in some cases).

I don't know exactly which technical form this will take, but I am convinced that this would be useful.

aiida-mcp

One strength of AiiDA is the breadth of information it records, which can be programmatically requested, such as which plugins are installed, which computers are available, which codes are available on these computers, etc.

I do think the direction of https://github.com/khsrali/aiida-mcp makes total sense, i.e. to enrich a user's LLM with this info via an MCP server. The straightforward direction that comes to mind would be that verdi mcp would start the server and offer a selection of its existing functionality as tools on the server

This indeed then enables cool features - like I can say in my VS Code chat "compute the phonons for the file I have open" and it will know from verdi mcp which codes I have, where to submit it, etc.