The Mezmo AI Assistant is designed to help you make faster, smarter decisions with your telemetry data. It works best when you guide it with clear, scoped prompts.
✅ Optimized For
Troubleshooting production issues
Detecting patterns across telemetry data
Explaining anomalies in plain language
Summarizing large volumes of logs
Identifying likely root causes faster
🧠 How to Use the AI Assistant Effectively
Use the built-in prompt suggestions to get started quickly.
Start broad, then narrow your prompt as you refine the investigation.
Think of the AI as a teammate — ask questions as you would with an engineer.
Include time, scope, and intent in your prompt for best results.
🛠️ It can also generate Mermaid visualizations
For example, you can ask:
“Show me a pie chart of service usage for app:restaurant-app in the last 24h”
pie showData
title _service for app:restaurant-app (last 24h)
"restaurant-performance" : 18269007
"metrics" : 4035581 "events" : 2200573
"performance" : 2200502
"access" : 373064
"app" : 27167
"server" : 24
"errors" : 5
🎯 Prompt Structure: What to Include (in order)
Element | Description |
1. Goal | What you're trying to achieve: RCA, reduce noise, fix errors, etc. |
2. Scope | Pipeline ID, app/service name, environment, cluster, etc. |
3. Time Window | Use UTC, e.g. “from 12:00 to 12:30” or “last 15 minutes” |
4. Symptom/Impact | E.g., “timeouts”, “dropped events”, “5xx errors” |
5. Known Filters | E.g., level:error, trace_id, host, customer ID (avoid PII) |
6. Desired Output | E.g., top patterns, timeline, likely root cause, next steps |
🧾 Sample Prompts
🔍 Root Cause Analysis (RCA)
“Run RCA for auth-service from 12:00 UTC to 12:30 UTC.
Symptoms: timeouts and 5xx errors.
Return: trigger → impact → top log clusters → recommended action.”
🔄 What Changed?
“Compare checkout-service logs between
09:00–10:00 UTC (baseline) and 12:00–13:00 UTC (incident).
What new errors or increases occurred?”
📉 Reduce Log Noise / Cost
“Identify top noisy logs in prod during last 1h.
Group by app and message. Suggest filters or redactions (include risks).”
🔁 Detect Duplicate Logs
“Check for duplicate logs in orders-service (prod) during 10:00–10:30 UTC.
Show dominant patterns and causes like retries or multi-shipping.”
📐 Build a Query
“Help write a Mezmo query to find timeouts in cart-service (prod)
excluding known retry noise. Fields: status_code, latency_ms, trace_id.”
🔎 Explore What’s Happening
“Show me the top services by volume in prod during last 30 minutes.”
⚠️ Usage Considerations
📦 Large Log Sets May Exceed Capacity
What may happen:
AI returns a generic error instead of a result.
Why:
The data has too much variation to summarize in a single pass.
Workaround:
Narrow your scope by time, app, namespace, host, or pod.
What’s coming:
A multi-agent system to analyze larger and more diverse datasets.
💬 Long Chat Sessions May Fail
What may happen:
Responses appear to load, then show a “Please try again” message.
Why:
Conversation history is stored in the browser and can reach limits.
Workaround:
Start a fresh session if the assistant stops responding properly.
📎 Tool Results Aren’t Remembered Between Messages
What may happen:
The assistant can’t recall previous tool output details.
Why:
To keep performance high, only the conversation text is retained.
Workaround:
Ask the AI to restate key results, or re-run queries as needed.
What’s coming:
Session memory to persist important outputs across turns.
🧵 Long Conversations Aren’t Summarized
What may happen:
Very long conversations may become unstable or unresponsive.
Why:
There’s no automatic compression of older messages yet.
Workaround:
Start a new session for each investigation or incident.
What’s coming:
Automatic summarization and history management.
⏳ No Chronological Log View (Yet)
What may happen:
The AI shows grouped patterns but not event-by-event timelines.
Why:
Current tools optimize for clustering and deduplication, not sequence.
Workaround:
Ask for grouped logs, then manually explore related timelines.
What’s coming:
Timeline-based tools that preserve both order and grouping.