Windows ARM64 + Node 25: qmd embed fails due to node-llama-cpp CUDA build + sqlite-vec unavailable

[MrHinsh]

Summary

On Windows ARM64, qmd is effectively unusable out of the box.

Running qmd embed triggers:

1. A failing CUDA build attempt via node-llama-cpp (LLVM for WoA required).
2. A hard failure due to sqlite-vec not being available.

This makes embedding impossible on a clean Windows ARM64 system.

Environment

• OS: Windows 11 ARM64 (WoA)
• Architecture: win32 arm64
• Node: v25.5.0
• npm global install: @tobilu/qmd@1.0.7
• GPU: none (CPU-only environment)

node -p "process.platform + ' ' + process.arch"
→ win32 arm64

Reproduction

1. npm install -g @tobilu/qmd
2. qmd embed

Observed Behaviour

First, node-llama-cpp attempts to build a CUDA variant:

[node-llama-cpp] A prebuilt binary was not found, falling back to building from source
CMake Error:
LLVM installation was not found. Please install LLVM for WoA

Then, even after falling back to CPU:

Error: sqlite-vec is not available.
Vector operations require a SQLite build with extension loading support.

Problems

1. The CLI attempts a CUDA build on Windows ARM64 even when no GPU is present.
2. No prebuilt binaries appear to be available for win-arm64 + current Node ABI.
3. sqlite-vec is not available on this platform, causing embed to fail completely.
4. There is no documented Windows ARM64 setup path.

Expected Behaviour

• CPU-only mode should work without requiring LLVM or CUDA toolchains.
• sqlite-vec dependency should either:
  • have a prebuilt for win-arm64, or
  • degrade gracefully with a clear instruction path.

Suggestions

• Publish win-arm64 prebuilds for:
  • node-llama-cpp dependency stack
  • sqlite-vec dependency
• Do not attempt CUDA builds when no GPU backend is configured.
• Document supported Node versions and platforms explicitly.

At present, Windows ARM64 appears unsupported in practice.