Note: This tool is under active development. Results/outputs are not guaranteed to be stable across runs or updates.
An LLM-based tool for building custom WDL workflows using validated components from the WILDS WDL Library.
This repository is the Minimum Viable Product (MVP) of a larger planned project. The final product is intended to be a point-and-click web application running on the Fred Hutch cluster, making WDL workflow generation accessible to users without requiring any local setup or command-line experience.
The WDL Writer is not a bioinformatician. It assembles workflows only from existing WILDS WDL Library module tasks. The user must define the operations needed (e.g. alignment, variant calling) and approve potential WILDS WDL modules to consider. It uses a local LLM whose output quality varies. Always review the generated WDL carefully before running it on real data.
- Python 3.13+
- Ollama running locally (or reachable via
OLLAMA_HOST), with a model pulled (default:granite-code:8b, configurable viaOLLAMA_MODEL) numpychromadbollamarapidfuzzstreamlitsentence-transformers(only needed for benchmarking)
- Clone the repo and install dependencies:
$ git clone https://github.com/getwilds/wilds-wdl-writer $ cd wilds-wdl-writer $ uv pip install numpy chromadb ollama rapidfuzz streamlit - Make sure Ollama is installed and running, then pull the default model:
$ ollama pull granite-code:8b
To point at a different Ollama host/model, set OLLAMA_HOST and OLLAMA_MODEL before running.
- Run the interactive CLI:
You'll be prompted for your input data type, format, species, and desired analysis steps, then asked to approve which tools to include before the final WDL is generated.
$ uv run python3 wdl_writer/generate_wdl.py
The tool can also be run on Fred Hutch's Chorus GPU nodes, using an Apptainer-packaged Ollama server for inference. See scripts/ChorusGenerateWDL.sh for the full walkthrough, which covers:
- Pulling the
getwilds/ollamaApptainer image - Starting an interactive GPU session on the
choruspartition - Launching the Ollama server inside the container and pulling/configuring the model
- Running
generate_wdl.pyagainst that local Ollama server
Fred Hutch and consortium users can schedule a Research Computing and Data Management Data House Call for help getting set up on the cluster.
A web UI (wdl_writer/app.py) wraps the same pipeline in a form-based Streamlit interface instead of the interactive CLI.
Locally:
cd wdl_writer
streamlit run app.pyThen open the URL Streamlit prints (typically http://localhost:8501).
On the cluster: submit scripts/ChorusStreamlit.sbatch from rhino:
sbatch scripts/ChorusStreamlit.sbatchThis sets up a venv, launches Ollama on a Chorus GPU node, and starts Streamlit. Watch the job log for a banner with the SSH tunnel command to run on your laptop and the local URL to open in your browser. (See also scripts/ChorusStreamlit.sh for the equivalent interactive, non-sbatch version.)
- Fred Hutch and consortium users: can get hands-on help via Data House Calls.
- Missing or incorrect WDL tasks: the WDL Writer can only use tools that exist in the WILDS WDL Library. If a tool you need is missing, file an issue there.
- Bugs or feature requests for this tool: file an issue in this repository. See CONTRIBUTING.md for how to contribute changes.
