Gibberish answers and Saving files to AppData. Issue with custom install path

[sleveo]

Title: Issue with custom install path: Gibberish output and files still saving to AppData

Description

I am attempting to move my local AI setup to a non-system drive (D:). Previously, a default C: drive installation of Ollama and AnythingLLM worked flawlessly on my machine (i5-14400, 32GB RAM, Intel iGPU).

I used your script to install to a custom path: D:\zAI\Ollama. While the installation completes without hard errors, I am running into two major issues.

Steps to Reproduce

1. Ran the script targeting D:\zAI\Ollama.
2. Set the system environment variable: OLLAMA_MODELS=D:\zAI\OllamaModels (Note: Please update this path to match your exact setup before posting!).
3. Started AnythingLLM and connected to Ollama.
4. Downloaded and ran gemma4:2b. I verified the Hash.
5. Sent a simple "Hello" query.

Expected Behavior

1. The model should output coherent responses.
2. All application data, models, and cache files should stay confined to the designated D: drive path.

Actual Behavior

1. Gibberish Output: The model responds with grammatically correct, real words and full sentences, but they are completely unrelated to the prompt (e.g., answering "Hello" with a random string of unrelated thoughts).
2. Data Leakage to C: Drive: Even with the script and environment variables active, roughly 2GB to 7GB of data is still being written to C:\Users\<myUser>\AppData. (Note: Specify if this is in AppData\Local\Ollama or AppData\Roaming\anythingllm-desktop).

I have clean-uninstalled and retried this process 4 times with the exact same result. I even tried a different model: Phi-3.5 Mini 3.8B with the same results.