Native processing of RO-crates in Galaxy via LLM/Agent

[anuprulez]

Background

RO-Crates provide a structured and machine-readable package containing research data, metadata, provenance, workflows, software information, contributors, licences, and relationships between research objects.

Galaxy already provides support for importing and exporting RO-Crates. However, understanding the contents of a large or domain-specific RO-Crate can still require knowledge of JSON-LD, RO-Crate profiles, workflow provenance, and the scientific context of the included files.

This discussion proposes adding an RO-Crate-aware LLM interface to Galaxy, allowing users to explore and interpret an RO-Crate using natural-language questions.

Related discussion: usegalaxy-eu/issues#908

Proposed functionality

A user should be able to select an RO-Crate file from file uploader in a Galaxy history and perform actions such as:

• Generate a human-readable overview of the crate.
• Recommend relevant Galaxy tools or workflows for processing the included data.
• Ask questions about its datasets, workflows, tools, parameters etc.
• Explain how a result was produced.
• Identify the inputs and outputs of a workflow execution.
• List the software and tool versions required to reproduce an analysis.
• Compare the metadata or provenance of multiple RO-Crates.
• ...

Example questions could include:

• What is contained in this RO-Crate?
• Which workflow generated the final result?
• What input datasets were used?
• Which Galaxy tools and versions are required to reproduce the analysis?
• Are the licence and contributor metadata complete?
• Summarize the main processing steps in non-technical language.
• Which files correspond to the final analysis outputs?

Suggested processing architecture

Galaxy should first process the crate by:

• Validate the archive and identify its RO-Crate profile.
• Parse ro-crate-metadata.json.
• Resolve the entities and relationships represented in the JSON-LD graph.
• Extract relevant metadata, including:
  • datasets and files;
  • workflows and workflow executions;
  • inputs and outputs;
  • tools, software, and versions;
  • provenance relationships.
• Construct a compact, structured context for the LLM.
• Submit that context and the user’s question through the Galaxy LLM infrastructure.
• Save the generated answer as a new Galaxy history dataset.
• ...

Possible Galaxy integration ?

Several integration levels could be considered.

Use Orbit ?

Use Orbit to upload RO-crate file and write necessary skills and route job processing to Galaxy

Add a file uploader to GalaxyAI

GalaxyAI can accept RO-crate files and then using suitable prompts.md file, process them

Dataset action ?

A more native interface could add an “Explore with an LLM” action to recognised RO-Crate datasets. This could open a question interface and automatically provide the parsed crate context.

Interactive crate explorer ?

A longer-term implementation could extend the existing RO-Crate or ZIP explorer with:

• a natural-language question panel;
• clickable entities and relationships;
• metadata completeness checks;
• generated summaries;
• provenance visualisation;
• links between answers and source entities.

Galaxy tool ?

An initial implementation could be provided as a Galaxy tool accepting:

• an RO-Crate archive;
• a natural-language question;
• an LLM provider or model;
• optional files or entity types to include;
• optional output format, such as Markdown, JSON, or HTML.

Other approaches ?

[anuprulez]

ping @bgruening @paulzierep @davelopez