Local AI Setup

Configure GGUF models through the NeuralScio worker or a compatible local endpoint.

NeuralScio has two separate local AI paths:

1. The built-in NeuralScioAI worker with GGUF models. 2. The Ollama-compatible provider path for existing Ollama setups or compatible runtimes.

The built-in worker does not require Ollama. It launches as a separate process, connects through QLocalSocket IPC, and loads GGUF models with a backend you choose in settings.

Built-in worker path

Use this path when you want the app-managed local model workflow.

What it uses

  • NeuralScioAI.exe
  • AiProcessManager
  • IpcProtocol over newline-delimited JSON
  • GGUF models stored on disk

Backend choice

You can choose the backend that matches your machine:

  • Auto for automatic selection
  • CPU for maximum compatibility
  • CUDA for NVIDIA systems
  • Vulkan for supported GPU stacks

Typical flow

1. Start the app. 2. The desktop app launches the worker when local AI is enabled. 3. The worker connects through local IPC. 4. The selected GGUF model loads. 5. The UI becomes ready for chat, summarization, and other local AI actions.

When to use it

  • You want a local path that does not depend on Ollama.
  • You want app-managed backend selection.
  • You want a single desktop workflow for AI and research.

Ollama-compatible provider path

Use this path if you already run Ollama or another compatible local runtime.

  • Install and start the Ollama-compatible runtime you want to use.
  • Point NeuralScio to that provider in AI settings.
  • Use this path only when you specifically want the compatible runtime behavior.

This is separate from the built-in worker. The built-in worker does not need Ollama.

Common problems

Worker unavailable

  • Check whether NeuralScioAI.exe started successfully.
  • Verify the local model path exists.
  • Restart the desktop app to retry worker startup.

Model load timeout

  • Use a smaller GGUF model.
  • Switch to CPU if your GPU backend is overloaded.
  • Check whether the machine has enough RAM and VRAM for the chosen model.

Out of memory

  • Reduce the model size.
  • Lower the context size if available.
  • Switch from CUDA or Vulkan to CPU for a more conservative load path.

CUDA or Vulkan unavailable

  • Install or update your GPU drivers.
  • Switch the backend to Auto or CPU.
  • Confirm the chosen backend is supported by your hardware.

Ollama unreachable

  • Verify the Ollama-compatible runtime is running.
  • Recheck the provider endpoint in settings.
  • Confirm that the runtime is listening on the expected local port.