DarkGrid AI

Troubleshooting

The sidebar says "backend offline" / app stuck on "Connecting…"

The app's local AI service didn't start.

  • Quit DarkGrid completely and relaunch — a leftover process from a previous session may still be holding a port.
  • Make sure security software on your machine isn't blocking the app from running locally.
  • If it keeps happening, reinstall the latest version from the Download page.

A model won't load, or the app runs out of memory

  • The model is too large for your RAM. Use a smaller model or a lower quantization (e.g. Q4_K_M instead of Q8_0). See Choosing a Model.
  • Close other memory-heavy apps and try again.

Inference is very slow

  • Use a smaller model — a 3B model is dramatically faster than an 8B one.
  • On Apple Silicon, GPU (Metal) acceleration is used automatically — no setup needed.
  • Lower the context length in Settings if you set it very high.

"Not enough disk space" when downloading

The app refuses downloads that won't fit. Free up space (delete unused models in Settings → Models, or a large Wikipedia archive you no longer need), or pick a smaller model/archive.

Wikipedia search returns nothing

  • Confirm an archive is installed in Settings → Knowledge Base and that a .zim file exists in kb/zims/.
  • Some specialized archives are small; try a broader one (e.g. Simple English Wikipedia).

Reference indexing fails or finds no text

  • Scanned/image-only PDFs have no extractable text. Use a text-based PDF, or OCR the document first.
  • Very large files take time on first index; let the progress bar finish.

An assertion or Metal message appears when the app closes

On macOS, the local inference engine may print a Metal assertion during process shutdown when both a chat model and the embedding model were loaded. This is a known upstream teardown issue and is harmless — it happens only as the backend exits and does not affect answers or your data.

Start completely fresh

Quit the app and delete (or move) the data directory listed in Getting Started. The app recreates it on next launch. This removes all models, archives, indexed references, conversations, and settings, back it up first if you want to keep anything.