Execution Workflow
End-to-End Flow
Section titled “End-to-End Flow”sequenceDiagram participant Service as Docker Service participant RAG as LlamaIndex participant DLP as DLP Scrubber participant Agent as LangChain participant LLM as Groq / DeepSeek participant ntfy as ntfy Server participant Admin as SysAdmin participant SSH as Paramiko Service->>RAG: 1. Crash detected RAG->>RAG: 2. Vector search logs + runbooks RAG->>DLP: 3. Raw retrieved text DLP->>Agent: 4. Sanitized context Agent->>LLM: 5. Reasoning request LLM->>Agent: 6. Root cause + fix command Agent->>ntfy: 7. Push alert with action button ntfy->>Admin: Notification on mobile Admin->>Agent: Confirm and Execute Agent->>SSH: Trigger remediation tool SSH->>Service: Apply fix on affected node
Step-by-Step
Section titled “Step-by-Step”1. Trigger
Section titled “1. Trigger”A critical service crashes on the main Proxmox node — for example, a Docker container exits unexpectedly or a health check fails.
2. Retrieval
Section titled “2. Retrieval”LlamaIndex queries the local vector database to pull the exact stack trace and cross-references it with local runbooks (restart procedures, known failure modes).
3. Sanitization
Section titled “3. Sanitization”Raw text passes through the Python DLP Scrubber. All MAC addresses, internal IPs, hostnames, and tokens are replaced with [REDACTED_*] tags.
4. Reasoning
Section titled “4. Reasoning”LangChain sends the sanitized logs and runbook context to the Groq or DeepSeek API.
5. Formulation
Section titled “5. Formulation”The cloud LLM analyzes the error and formulates a fix — for example, “Restart the database container” — returning a structured response to LangChain with the proposed command.
6. Alert
Section titled “6. Alert”A local POST request hits the ntfy server. The sysadmin’s phone receives the analysis summary and a Confirm & Execute action button.
7. Remediation
Section titled “7. Remediation”The admin taps Confirm & Execute. The webhook triggers LangChain to call the Paramiko_SSH tool. Paramiko logs into the failing node over SSH and applies the fix.
Design Principles
Section titled “Design Principles”- Human in the loop — analysis can be automated; execution requires explicit approval.
- Sanitize first, reason second — no cloud API call receives unsanitized infrastructure data.
- Local retrieval, cloud reasoning — RAG stays on-prem; only masked context leaves the network.
See Core Components for per-component details.