Building Grafana Assistant
summary
Grafana Assistant embeds an LLM chat in the frontend with tool calls that actually run queries against Prometheus and Loki, generate mermaid diagrams and deep links, and keep a 20k‑token system prompt plus ~20k of tool descriptions. Tools execute in the browser while a Go backend handles orchestration. A single manage_investigation tool replaces many tiny tools to reduce confusion, memory stores context, and an agentic swarm of lead, specialist and reporter agents runs concurrent probes; evaluation uses the LLM as judge and the feature proves useful in demos and internal use.
AI-written summary · about
chapters (13)
- 0:00Introduction to Grafana Assistant
- 2:06What is Grafana Assistant
- 6:01Live demo of Grafana Assistant
- 9:28Deep links and navigation
- 11:15Trust and hallucination handling
- 13:01Context and structured input
- 16:31Investigating performance issues
- 19:28User experience and visualization
- 21:13APIs and extensibility
- 23:19Evaluating and improving the assistant
- 25:50Tool crafting and simplification
- 26:53Multi-agent swarm investigations
- 29:23Key takeaways and conclusion
related talks
23mAI/ML + Grafana: How to Create Intelligent Grafana Apps Leveraging LLMs
21mAgent Observability Beyond Vibes
44mIntroduction to Grafana Loki
53mGetting started with Grafana Loki, a modern logs database
23mAI/ML + open source observability
05mGrafana Loki Log Aggregation for Incident Investigations
26mExploring Observability with MCP Servers
49mKeynote: Grafana 11, Loki 3.0, Alloy, Golden Grot Awards, and more