Can't Connect
Can't open http://ai.local:3001 in my browser
›
1
Check you're on the same network. Your computer must be on the same local network as the Nexus One AI server — either the same office Wi-Fi or wired LAN. VPN may block local addresses.
2
Try the server's IP address instead. Ask your administrator for the IP address and go to
http://[IP]:3001 directly. Some networks don't resolve .local hostnames.3
Check the system status panel. Go to the home page and look at the status panel. If Open WebUI shows red, the service may be stopped.
4
Try a different browser. Chrome or Firefox are most reliable. If one browser doesn't work, try another.
5
Contact your IT administrator. Ask them to run
systemctl status open-webui on the server to check if the service is running.Jupyter Notebook won't open at http://ai.local:8888
›
1
Check the system status panel on the home page. If Jupyter shows red, the service needs to be restarted by your administrator.
2
You need the access token. Jupyter requires a token or password to log in. Ask your administrator for it. They can find it by running
jupyter server list on the server.3
Try the IP address. Use
http://[server-IP]:8888 if ai.local doesn't resolve.Slow or Unresponsive
The AI is responding very slowly
›
1
Switch to a smaller model. Larger models (70B) are slower to respond. Try llama3.1:8b or llama3.2:3b for faster replies. See the model library for comparisons.
2
Check if many users are active. The GPU is shared. If 10 people are all chatting simultaneously, everyone gets slower responses. Try again when fewer users are active.
3
Check GPU utilisation. Open the nvtop tool (via SSH) to see if the GPU is fully loaded. If utilisation is at 100% and you're waiting, the GPU is saturated.
4
Shorten your prompts. Very long input messages (pasting entire large documents) slow things down. Use the document upload feature instead — it's more efficient.
GPU shows 0% utilisation even when sending queries
›
1
Check if a model is loaded. SSH into the server and run
ollama ps. If the list is empty, no model is loaded. Send a message in Open WebUI to load one — the first response will trigger the load.2
GPU utilisation spikes briefly. If responses are appearing and the GPU shows 0%, don't worry — the query may process fast enough that nvtop/DCGM misses the spike. GPU% is averaged; instant spikes may not show.
3
Check Ollama is running. Run
systemctl status ollama on the server. If it's not active, run sudo systemctl start ollama.Document Upload Issues
Document upload fails or shows an error
›
1
Check the file type. Supported formats are PDF, DOCX, TXT, MD, CSV, XLSX. Image files (.jpg, .png), PowerPoint (.pptx), and compressed archives (.zip) are not supported for chat queries.
2
Try a smaller file. Very large documents (100+ MB) may time out during processing. Split the document or try uploading one section at a time.
3
Scanned PDFs won't work well. If your PDF is a scanned image (not text-based), the AI can't read it directly. It needs to be a searchable/text PDF. Use OCR software to convert it first.
4
Check ChromaDB is running. Document uploads go to ChromaDB. Check the status panel on the home page — if ChromaDB is red, uploads will fail until it's restarted.
AI is not finding the right content in my documents
›
1
Rephrase your question. RAG searches by meaning, not keywords. Try asking the same thing differently — "What is the payment schedule?" vs "When are payments due?" may give different results.
2
Be more specific. Instead of "What does the document say?" try "What does section 3.2 say about termination clauses?" Specificity helps the retrieval step find the right chunk.
3
Re-upload the document. If the upload completed but the AI can't find content, try deleting and re-uploading the document. The indexing process may not have completed correctly.
4
Use the Knowledge Base for multi-document queries. If you need the AI to search across many documents consistently, add them to a Knowledge Base in Workspace → Knowledge rather than uploading per-conversation.
AI Giving Wrong or Unexpected Answers
The AI gave me wrong or made-up information
›
⚠️ This is a known limitation of all AI models, not a system fault. It's called hallucination. AI models can generate confident-sounding but incorrect answers, especially for specific facts, numbers, dates, or people.
1
Never rely on AI alone for critical decisions. Always verify important facts, figures, regulations, or legal information from authoritative sources.
2
Use document mode for factual queries. Upload the source document and ask the AI to answer from it. The AI is far more accurate when it can read the actual text rather than relying on its training knowledge.
3
Ask the AI to cite its source. Try adding "Quote the exact sentence from the document that supports your answer." This forces the AI to ground its response in the uploaded text.
4
Try a larger model. Models like llama3.1:70b (if available on your tier) are generally more accurate than 8B models. See the model library.
Out of memory error / model fails to load
›
1
Switch to a smaller model. If the model you're trying to load requires more VRAM than is available, it will fail. Try llama3.1:8b instead of 70B. See the model library for VRAM requirements.
2
Unload other models first. SSH into the server and run
ollama stop [model-name] to unload models that are sitting in VRAM idle. Then try loading your model again.3
Check available VRAM. Run
nvidia-smi on the server to see how much GPU memory is free. If other processes are using it, that reduces what's available for your model.Services Down
A service shows red in the status panel — how do I restart it?
›
Ask your IT administrator to run the relevant restart command on the server:
Restart commands
Open WebUIsudo systemctl restart open-webui
Ollamasudo systemctl restart ollama
ChromaDBsudo systemctl restart chromadb
Jupytersudo systemctl restart jupyter
Check statussudo systemctl status [service]
Still stuck?
Contact Cezen Support
Email us at support@cezentech.com with the following information to speed up the response:
- What you were trying to do
- What happened instead (exact error message if any)
- Which tool was affected (Open WebUI, Ollama, etc.)
- Screenshot of the status panel on the home page
- Output of
nvidia-smiandollama psif GPU-related