Skip to content

Chapter 13: AI Integration

Overview

melisai generates a structured prompt that can be fed to any LLM (GPT-4, Claude, Gemini, etc.) for automated analysis. The prompt includes the report data plus a curated database of 21 anti-patterns.

Source File: ai_prompt.go

  • Lines: ~103
  • Purpose: Generate context-aware AI analysis prompts

Prompt Structure

func GenerateAIPrompt(report *Report) string {
    prompt := "SYSTEM PERFORMANCE ANALYSIS\n\n"
    prompt += "You are an expert Linux performance engineer.\n"
    prompt += "Analyze the following system diagnostic report and provide:\n"
    prompt += "1. Root cause analysis of any anomalies\n"
    prompt += "2. Severity assessment\n"
    prompt += "3. Specific remediation steps\n"
    prompt += "4. Check for known anti-patterns\n\n"

    // Include relevant anti-patterns based on observed data
    prompt += generateAntiPatterns(report)

    // Include the full report data
    prompt += "\n--- REPORT DATA ---\n"
    prompt += jsonMarshal(report)
}

Anti-Pattern Database

The AI prompt includes known performance anti-patterns, filtered based on what's relevant to the current report:

ID Anti-Pattern Trigger Condition
P1 Single-CPU saturation Per-CPU data shows 1 core at 99%+
P2 IOWait misconception IOWait > 20% but disk util < 50%
P3 Memory "leak" (actually cache) MemFree low but Available high
P4 Swap with available memory Swap used but MemAvailable > 20%
P5 Container throttling nr_throttled > 0
P6 CLOSE_WAIT leak close_wait_count > 0
P7 TIME_WAIT exhaustion time_wait > 30000
P8 Zombie accumulation zombie_count > 10
P9 FD exhaustion Any process FDs > 5000
P10 NUMA imbalance numa_miss / (hit+miss) > 10%
P11 Memory fragmentation All high-order buddyinfo = 0
P12 Disk scheduler mismatch SSD with mq-deadline / HDD with none
P13 Cubic on WAN tcp_congestion = cubic + RetransSegs > 1%
P14 Small TCP buffers tcp_rmem max < 4MB
P15 Low somaxconn somaxconn < 128 with high PassiveOpens
P16 High context switches ctx_switches > 100K/s
P17 Steal time on VM steal > 5%
P18 SoftIRQ storm softirq > 20% on any CPU
P19 Dirty page buildup DirtyBytes > 1GB
P20 OOM in dmesg dmesg contains "Out of memory"
P21 ECC errors dmesg contains "EDAC"

Context-Aware Filtering

Only relevant anti-patterns are included in the prompt:

func generateAntiPatterns(report *Report) string {
    var patterns []string
    if hasCPUData(report) {
        patterns = append(patterns, cpuPatterns...)
    }
    if hasContainerData(report) {
        patterns = append(patterns, containerPatterns...)
    }
    // Only include patterns for which data exists
}

This keeps the prompt focused and avoids wasting AI tokens on irrelevant patterns.

Usage

sudo melisai collect --ai-prompt --output report.json

This adds an ai_prompt field to the JSON output that can be directly sent to an AI API.

Next: Chapter 14 — Report Diffing