Performance issues with Ollama on first message with disabled MCP servers

[nicolaskrier]

Hello,

As discussed in this issue, I am experiencing performance issues while running Dive locally with all configured MCP servers disabled on my MacBook Pro M4 with 48 GB RAM. As @kevinwatt suggested, the slowness occurs only after sending the first message; the rest of the conversation is fine.

Tests Conducted

I performed some tests by typing only a greeting message: "Hello!". Here are the results:

1. Test with Ollama only:

ollama run mistral-small:latest

Result: The response appears quickly.

2. Test with OpenWeb UI and Ollama:

ollama serve

Result: The response appears after 7-8 seconds.

3. Test with Dive and Ollama:

ollama serve

The response appears after 11-12 seconds.

Logs

Here are the logs for the last two tests:

• ollama-debug-with-dive.log
• ollama-debug-with-openweb-ui.log

Thanks

Thank you in advance for your assistance.

[kevinwatt]

Hi Nico,

Thanks for the testing. I double-checked the log files. They look the same, except for the run times.

Based on our internal discussions, this issue might be related to the Desktop's Default System Prompt. Similar to Claude Desktop, Dive also has a System Prompt of about 1000 tokens. This is currently the default setting.

https://github.com/OpenAgentPlatform/Dive/blob/main/services/prompt/system.ts

As a result of our discussions, we will provide an option to disable it in the version(0.7.3) in the system settings. However, turning it off might result in worse performance for some less capable models.
Nevertheless, this turn-off prompt feature is still necessary when conducting basic tests like this or when users simply want to use the model directly.

We greatly appreciate your bringing this issue to our attention. Your feedback is instrumental in helping us improve Dive.