--- title: "Chat and follow-ups" description: "How a prompt becomes a dashboard, what the agent does at each step, and how follow-ups revise the same dashboard." --- > Documentation Index > Fetch the complete documentation index at: https://slashobserve.com/llms.txt > Use this file to discover all available pages before exploring further. # Chat and follow-ups The chat on the **Observe** page turns a sentence into a saved dashboard. The server runs the agent against your configured datasources, checks every query it writes, and saves the result as a version of a new dashboard. It needs a model; see [LLM providers](/docs/connect/llm-providers/). For a first run, follow [Your first dashboard](/docs/get-started/first-dashboard/). ## Start a chat Type into the box labelled "Message the agent" and press Enter (Shift+Enter adds a line; on a touch device Enter adds a line and the arrow button sends), or choose one of the example prompts, which sends it at once. Pressing `/` outside a text field moves the cursor to the box. A prompt is 1 to 4000 characters and cannot contain control or bidirectional-text characters. The box shows "Describe the dashboard you want first." when it is empty and "Use at most 4000 characters." when it is too long. Starting a dashboard needs Edit on a folder, because the agent creates a dashboard for you. Without it the chat home shows "Starting a dashboard needs Edit on a folder. Ask an Admin if you need it." See [Roles](/docs/teams/roles/). The chips under the box list the healthy datasources and the name of the model; the agent can only query datasources the server has configured, see [Prometheus](/docs/connect/prometheus/) and [ClickHouse](/docs/connect/clickhouse/). Without a model the box is replaced by "No model is connected to this server." **Import from Grafana** next to the send button opens [Import Grafana dashboards](/docs/use/import-grafana/), which needs no model. Under the example prompts, the chat home lists your saved prompts and those shared with everyone, and typing `/` as the whole first line of the box opens a picker for them; see [Saved prompts](/docs/teams/saved-prompts/). Each person may have 3 new-dashboard requests queued or running, and the instance 20. Past that the server answers `429` with `too_many_requests` or `agent_busy` ("The agent is busy with other requests; try again in a moment."); wait for a request to finish. Follow-ups do not count toward the 3. Your recent chats appear on the Observe page under "Recent chats", and the **History** button lists older chats and Grafana imports. ## What the agent does The agent is a loop of model calls and tool calls run by the server. It does not see your browser, and it can only act through its tools. Its instructions hold the Spec guide, the schema and an example, and tell it to use only the datasources it is given and only the metric, label, table and column names it discovered, so it does not invent names. If the prompt itself contains a Grafana dashboard JSON, the agent imports it with the importer instead of rewriting it, which ends the run like a save; this works for a new dashboard only, and the JSON has to fit in the 4000 characters of the prompt. Its order of work is: 1. Pick a datasource, then discover what exists: list datasources, metrics, labels and label values on Prometheus, and tables, schemas and sample rows on ClickHouse. 2. Draft the dashboard as an ObserveSpec document. 3. Lint every new or changed query against its datasource. 4. Optionally run a query to see its data. 5. Save. The conversation shows these as phases, each with a message from the server: | Phase | Examples of what you see | | ---------------- | -------------------------------------------------------------------------------------- | | Discovering | "Listed 14 metrics on prom", "Read the otel_logs schema on ch" | | Checking queries | "Checked a query on prom: 0 errors" | | Running queries | "Ran a query on prom: ..." | | Saving | "Saved version 1" | | Repairing | "Save rejected: 2 issues (...)" | A save is the real check. It runs every panel query and rejects a panel whose query is invalid or returns no data, except for alert list panels. The rejection comes back to the agent with the problem for each panel, and the agent fixes the draft and saves again. The instructions allow three repair attempts; if the agent still cannot save, the request fails. A successful save ends the run, and the version summary is a sentence the agent wrote, "Create dashboard" if it wrote none. Queries the agent runs while drafting are capped at 15 seconds, 200 series, 100,000 points and 1,000 rows, so a heavy query may fail where the saved panel would work. A run gets 3 minutes and at most 20 model calls. When it runs out of either, the conversation ends with "The dashboard could not be built" and a **Try again** button. A run that a server shutdown interrupts goes back to the queue and starts again, and one that a crash interrupts gets one more attempt. Only the server's step messages and errors are shown. The model's own text is never displayed, so there is no commentary to read and no answer to a question. A dashboard has 2 to 8 panels unless you ask for another number, and the agent adds variables only when the request asks for a breakdown people will filter on. It uses the time range you state ("last 6h" becomes `now-6h` to `now`), and otherwise the last hour. ## Read the result When the run ends the progress list collapses into a line such as "Ran 14 steps, repaired 1 query". Open it to see every step. The result card shows the title, "N panels from ``" and the footer "Saved as version 1". Choose **Open dashboard** to go to it. The dashboard lands in the folder you last created a dashboard in, or General when you have not created one; see [Folders](/docs/teams/folders/). A failed request shows "The dashboard could not be built" (or "The dashboard could not be updated" for a follow-up) with the reason the server stored, and **Try again** starts a new request with the same prompt. When the request already has a dashboard, an **Open dashboard** button is shown next to it. Every request has an id, shown as "Request" and its first eight characters at the bottom of the conversation, which you can quote when reporting a problem. The page keeps polling the server. It shows "Reconnecting..." after two failed polls, and "Progress stopped updating." with a **Retry** button when polling stalls. The run itself goes on server-side either way. ## Follow-ups Once the first dashboard exists, the box under the conversation reads "Ask for another change to ``", with the note "Each change is saved as a new version of the dashboard." Until then it is disabled with "Wait for the dashboard, then ask for changes." An empty follow-up says "Describe the change you want first." A follow-up is not a continuation of the chat. The agent starts a fresh run that sees the datasources, your new message and the current Spec of the dashboard, not the earlier messages. Say what to change in the message itself, for example "only show 5xx" or "add a panel with memory by pod". The agent revises the latest version of that dashboard and saves the result as the next version. It keeps panel ids, positions and everything else you did not ask to change, and it does not delete panels marked unverified unless you ask. It saves again only after the same checks as a new dashboard. If someone saves a version while it works, it restarts once from the newest one. Requests for one dashboard run one at a time, in the order they were sent, and up to 5 can be queued or running. A sixth is refused with `too_many_revisions`. You can also send a follow-up from the dashboard. In its activity card, write the change in "Write a comment, or tell the agent what to change" and choose **Send to agent**, or reply to a comment thread and choose **Send to agent** there. These requests carry on in the activity card; see [Comments, agent requests and versions](/docs/use/review/). In the conversation, a turn sent from the dashboard shows "Sent from the dashboard". A follow-up needs Edit on the dashboard's folder, which Editors and Admins have by default; see [Roles](/docs/teams/roles/). It does not need to have been your dashboard. ## What the agent cannot do - **Answer questions.** It only produces dashboards, and its text is not shown. - **Use a datasource that is not configured or not healthy**, or names it did not discover. - **Work on a dashboard over 128 KiB of Spec.** It fails with "the dashboard is too large for the agent to revise". A dashboard holds at most 200 panels. - **Change a dashboard you cannot edit.** It cannot reach folders you have no Edit access to. - **Add panels with no data.** The save refuses them, and the agent does not create placeholders. - **Make the same choices every time.** Results depend on the model; a clearer prompt with the metric, service or time range you want gives better dashboards. To work on a dashboard in your own editor and model, use [MCP](/docs/use/mcp/) or the [CLI](/docs/use/cli/). To change one panel by hand, see [Edit a panel](/docs/use/edit-panel/).