Interview-Driven Workflow Construction and Repair in Galaxy

Talk to Galaxy

The code in this folder is one team’s implementation of the pitch in the project README: turns a free-text research description into a real, mechanically-validated Galaxy workflow — reused from the IWC registry, adapted from one, or built from scratch — never just a plausible-looking guess. Full context and goals this implementation is measured against: PROPOSAL.md.

Get started

Install, credentials, running it, and cleanup are all in SETUP.md — start there. From the repository root, the short version:

cd dejian_workflow
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python3 scripts/run_pipeline.py interviews/raw/example-sarscov2-amplicon.txt

Layout

Path	What’s there
scripts/	The pipeline itself — route → fetch/build/adapt → validate. See scripts/README.md for what each script does and its current status.
interviews/	Example input: 3 synthetic + 50 real (Biostars) curated interviews in raw/, plus 100 more real ones in analysis_request_examples_100/. See interviews/README.md.
knowledge_base/	The curated ~20-workflow bacteria/virus catalog the local router matches against. See knowledge_base/README.md.
workflows/	Output of the routing decision (use/adapt/build), one subdirectory per interview. See workflows/README.md.
eval/	Pass-rate reports, the failure taxonomy, and expert-authored evaluation cases. See eval/README.md.
docs/	Presentation decks: what real interview data looks like (interview-data.marp.md), what Galaxy pipelines exist (galaxy-bacteria-virus-pipelines.marp.md), and how they’re bridged together (pipeline-architecture.marp.md).
mcp/	Config scaffold for grounding against the BRC Analytics MCP server (Goals 2–3) — not yet wired into any script, distinct from scripts/galaxy_mcp_client.py’s real, working integration with galaxy-mcp’s live IWC registry search.
PROPOSAL.md	The full project proposal this implementation is measured against.
SETUP.md	Install, credentials, testing, updating, and cleanup instructions.

What’s real vs. what’s a stand-in

Routing, fetching, static validation, byte-stable adaptation, and the live IWC registry check (via a real galaxy-mcp MCP integration) are all real, tested code — not stubs. The build (LLM-generation) path and the follow-up question loop are real but incomplete (no Galaxy Workflow Foundry grounding, no answer-collection loop yet). See scripts/README.md’s status table and SETUP.md’s “What’s not deployable yet” section for the honest, current list of gaps.