← All articles · Maintain
Resetting Blueprint data
How to clean up models, calibration runs, the Python sidecar, or the entire ~/.blueprint directory — including what survives a reset and what doesn't.
3 min read
What lives in ~/.blueprint/
Everything Blueprint writes goes under ~/.blueprint/ (or
%USERPROFILE%\.blueprint\ on Windows). The current layout:
~/.blueprint/
├── models/ # downloaded GGUF weights (often 100+ GB)
├── runtime/ # llama.cpp binary (~300 MB)
├── python/ # uv-managed venv + Python features (up to 5 GB)
├── calibrations/ # imatrix + eval results per run (varies)
├── loras/ # trained LoRA adapter files (50 MB each)
├── hosts.json # host registry (SSH targets)
├── remotes.json # remote-server registry (OpenAI endpoints)
├── svc-token # bearer token for the local supervisor
├── service-config.json # current serve config
├── service-status.json # current supervisor status
├── logs/ # blueprint.log + svc logs
└── .first-run-done # marker for "skip the welcome overlay"
The Maintain tab gives you scalpel-like control over each.
Per-feature cleanup
Delete a model
Maintain → Models → Delete. Removes the .gguf file. Frees
disk immediately. The model can be re-pulled at any time.
Delete a calibration run
Maintain → Calibration runs → Delete. Removes the entire run directory (prompts, eval set, imatrix, calibrated quants, results, report). Free disk; non-recoverable.
Uninstall a Python feature
Dashboard → Python runtime → Uninstall. Removes the packages from the venv but leaves the venv itself + uv + the Python interpreter. To wipe those, use the Reset Python option from Maintain.
Reset Python sidecar
Maintain → Python runtime → Reset. Removes
~/.blueprint/python/ entirely. Next time you install a Python
feature, the venv + Python install rebuild from scratch.
The nuclear option
Reset Blueprint data
Maintain → Reset Blueprint data. Removes the entire
~/.blueprint/ directory. After this:
- Every model you've pulled is gone — re-pull
- Every calibration run is gone — re-run
- Every host registration is gone — re-add
- Every LoRA adapter is gone — re-train
- The Python sidecar is gone — re-install features as needed
- Logs are gone
- The first-run welcome overlay shows again
This is irreversible. The card confirms via a typed-input modal ("type RESET to confirm") to prevent accidental clicks.
When you'd use it: you're handing the machine to someone else, you suspect a corrupted state you can't isolate, or you genuinely want a clean slate for testing.
What ISN'T in ~/.blueprint/
- The desktop app binary — that's in your Programs folder / Applications. Reset doesn't uninstall the app.
- The installed
blueprint-svcWindows service (if you set one up) — Reset data doesn't unregister the SCM entry. Use Maintain → Uninstall service first if you want to fully wipe. - Wails WebView2 cache — under
%LOCALAPPDATA%\Blueprint\on Windows. Reset doesn't touch it. Usually safe to leave; if you want it gone, delete that directory manually.
Common pitfalls
- Reset while serving: the active llama-server holds open file handles on the model. On Windows, this can block the delete. Stop the server first.
- Forgetting that LoRAs are tied to base models: if you delete the base model but keep the LoRA, the LoRA is dead weight — it only loads against its specific base.
- Re-pulling the same models from scratch is slow + bandwidth-
expensive: if you're going to reset, consider backing up
~/.blueprint/models/to an external drive first.
Need help with this in production?
Inspire AI Lab runs LLM optimization engagements end-to-end using Blueprint. If you'd rather hand the work to us instead of running it yourself, book a 30-minute review.