Proposal: AI Copilot for AiiDA (Multi-Agent + RAG)

General Usage
1

Hi everyone,

After reading this year’s GSoC proposal on implementing a natural language interface for AiiDA using multi-agent AI, I’ve been thinking about a possible direction and would love to get early feedback.

Rather than approaching this as a simple chat-based wrapper around CLI commands, I’m considering framing it as an “AI Copilot” for AiiDA, an intelligent assistant that integrates into the workflow lifecycle and assists users in a structured, architecture-aware way.

AiiDA has a rich architecture (engine, workflows, provenance graph, plugins), and a natural language interface should ideally:

  • Respect provenance integrity
  • Avoid unsafe or careless code generation
  • Integrate cleanly with existing abstractions
  • Be maintainable and modular

Given the team’s concerns about low-effort AI-generated contributions, I think this project should focus heavily on architectural design and validation mechanisms rather than just LLM integration.

High-Level Concept

The AI Copilot would assist users in:

  1. Workflow Design

    • Generate WorkChain skeletons
    • Suggest input structures
    • Guide plugin usage

  2. Debugging

    • Inspect failed processes
    • Parse logs
    • Explain likely causes
    • Suggest corrective actions

  3. Provenance Exploration

    • Translate natural language into structured QueryBuilder queries
    • Summarize provenance graphs
    • Explain data lineage

  4. Optimization & Suggestions

    • Detect common misconfigurations
    • Suggest improvements based on past runs

Proposed Architecture

Instead of a single LLM, I’m considering a modular agent structure:

  1. Intent Agent

Classifies user requests (create workflow, debug, inspect provenance, etc.).

  1. Context / Retrieval Agent (RAG)

Retrieves relevant information from:

  • AiiDA documentation
  • Local workflow code
  • Execution logs
  • Node metadata / provenance graph

This would ground responses and reduce hallucinations.

  1. Execution / Generation Agent

Generates:

  • WorkChain templates
  • QueryBuilder queries
  • Explanations
  • Debugging suggestions

All outputs would be structured and constrained to AiiDA’s abstractions.

  1. Validation / Safety Layer

Checks:

  • Compatibility with AiiDA’s architecture
  • Proper engine usage (run, submit, etc.)
  • Schema correctness

This layer would help address concerns around uncontrolled LLM-generated code.

Preliminary Technical Goals

If this direction aligns with the team’s expectations, I would aim to:

  • Draft a more detailed architectural design (components, data flow, extension points)
  • Define strict output schemas for generated code
  • Prototype a minimal but well-structured implementation

Open Questions

I’d really appreciate feedback on this. Looking forward to your thoughts — and excited about the possibility of building with AIIDA.

Thanks!
Unnati Kadam

1 Like