Building ZORGAX: From a Local AI Model to a Server Operations Agent for MyZubster

Описание и применение

ZORGAX успешно функционирует как разговорный помощник, но при отсутствии активных автоматизаций повторно вызывал функцию list_automations, что выявило проблему надёжности вызовов инструментов.

Задача

ZORGAX repeatedly called the same tool function when no automations were available, exposing inefficiency and potential danger.

Результат

ZORGAX now functions as a conversational assistant with validated connectivity and controlled tool calling; repeated tool calls issue isolated for further work.

Рабочий процесс

• Configure Docker networking to expose Ollama at 127.0.0.1:11434 via systemd socket proxy.
• Update Open WebUI to use reachable Ollama endpoint http://172.22.0.1:11434.
• Verify Ollama API response with HTTP 200 and model list.
• Test text generation via /api/generate endpoint and confirm healthcheck passes.
• Simulate tool interaction: user asks for active automations, ZORGAX calls list_automations, receives empty result, and replies no automations.
• Identify repeated tool calls in Open WebUI path and isolate issue for further investigation.

Доказательства

We asked ZORGAX to identify which tools it could actually use.

Источник

Instead of terminating the process and reporting that no active automations were found, ZORGAX repeatedly called the same function.

Источник

We tested ZORGAX directly through Ollama, without Open WebUI. Then we simulated a complete tool interaction: ... There was no repeated tool call in that controlled test.

Источник