The article introduces Grafana Assistant, a large language model integration inside Grafana that lets users chat with existing data sources, extract insights, and generate custom dashboards. It targets engineers working with CAN and LIN data recorded via CANedge devices and stored as Parquet data lakes on Amazon, Google, or Azure clouds. With Grafana Assistant, users can explore months of vehicle or machinery data across many devices without writing SQL, turning queries and findings into shareable, navigable Grafana panels.
Compared with general chatbots like ChatGPT, the piece notes practical hurdles for analyzing time series at scale: manual file uploads, static outputs, and limited shareability. While model context protocol setups can extend chatbot tooling, they are complex for most users. Grafana Assistant, by contrast, runs directly against configured data sources, offering zero additional setup, the ability to query terabytes, conversational exploration, rapid dashboard creation, and easy retention and sharing of insights.
Getting started is straightforward: use the provided dynamic Grafana dashboard template, select the relevant data source as context, and paste a refined system prompt to improve consistency. The authors observe that supplying guidance in-chat often works better than setting universal rules, though results can vary.
Seven showcases demonstrate the workflow using a 1 GB Kia EV6 dataset via the Grafana-BigQuery integration. First, the assistant inventories tables, columns, and time ranges with iterative queries, though it shows limited recall of earlier outputs in longer chats. Second, it runs ad hoc analyses in chat and converts them into dashboard panels, with the tip to specify signal units to avoid guessing. Third, a high-level prompt produces a functional but inconsistent multi-panel dashboard, while a detailed prompt yields a repeatable, production-ready layout without requiring SQL from the user and at a notable time savings.
Further tests highlight boundaries and workarounds. For kWh per 100 km, a multi-query-plus-transformations approach proves brittle, whereas a single SQL approach works more reliably but embeds assumptions. The assistant also migrates an Athena dashboard to BigQuery with about 95 percent accuracy, and it can automate repetitive panel creation, though it currently cannot reference panel outputs as data sources, leading to redundant queries.
Support varies by backend: BigQuery is fully supported; Athena is partial since queries cannot be executed; Synapse is unsupported. Grafana Assistant is available in Grafana Cloud with rate limits and, at the time of writing, is included in the free tier. The article cites documentation stating data is encrypted, access is limited to the assistant team, and data is not used for model training, and advises checking company policies. Overall, it is a powerful concept that excels at visualization and light exploration, with occasional inconsistencies and clear room for growth.
