AI Chat
Ask anything about the fleet in plain English. The assistant queries live data โ metrics, alerts, issues, cluster topology, billing, logs โ and streams its answer as it works, showing each step it takes in a panel above the answer (see Watching It Work).
What To Askโ
- "What's wrong with acme-prod?" โ cluster names work; no need for UUIDs
- "Which clusters had the most firing alerts this week?"
- "Show me the topology of acme's biggest cluster"
- "Did anything change on cluster X before the latency spike yesterday?"
- "Which Tier A customers have clusters below score 60?"
- "Is anything wrong with this table design?" followed by the pasted
CREATE TABLE(no cluster needed) - "This dashboard query on poc-retail-01 runs every minute and is slow. Would a materialized view help?"
The agent can search the internal knowledge base too โ past incidents, known bugs by StarRocks version, and proven fixes.
Watching It Workโ
While an answer is being written, a steps panel above it shows what the assistant is doing, much as claude.ai and Claude Code do: how it read your question, the evidence it gathered, what it thought, which tools it called and how it checked the answer.
- While it works the panel is open. Its header names the current step and the time so far, for example Collecting a snapshot of kkclster1โฆ ยท 0:12, and the timeline below adds each step as it happens (the one under way has a spinner). Click the header to collapse the panel: the current step stays in the header, so a collapsed panel still tells you what is going on. The panel remembers, in this browser, whether you last left it open or closed while an answer was streaming, and the next answer starts that way.
- When the answer is done the panel collapses to a short summary, for example Worked for 42 s ยท gathered evidence (5 lookups) ยท thought for 8 s ยท 6 tool calls ยท revised the answer. Failed tools, a spent lookup budget and a cut-off reply show in amber; a long summary wraps to a second line, and hovering it shows it all. If you opened or closed the panel while the answer was streaming, or opened one of its rows, it stays open. Click the summary to see the steps again.
- Short answers get no panel. A greeting or a general question answered without lookups has nothing to show, so it has no panel.
- Rows with a chevron expand. Click one for its details; rows without a chevron say all there is.
- Reading while it works. The chat follows the answer as it streams while you are at the bottom. Scroll up, or open a step to read it, and it stops following; scroll back to the bottom and it follows again.
| Step | What it shows |
|---|---|
| Understood the request | The kind of question the assistant took it for (for example Table design review, Known-issue check) and the facts it picked out of your message: cluster, databases, tables, error, goal, what you pasted. Only what you wrote: pasted lines that read like instructions are counted, never quoted. Not shown for a general question with nothing to pick out. |
| Gathered evidence | The server's read-only lookups before the assistant started writing, each with its time and result size, the parts that went into the evidence, and what was not available. |
| Some data sources are unavailable | The sources this kind of question uses (Admin Console, Knowledge Lake, fleet database, fleet metrics, direct SQL to clusters) that were not available when the answer started, and why: unreachable from this server, stale (the newest data is old), not configured or no credentials on this server. |
| Thought for 8 s | A summary of the model's thinking. Only the Claude 5.x models (Opus 5.5, Fable 5.1, Sonnet 5.5, Haiku 5.5) show thinking. On the other models (Sonnet 4.6, the shipped default, and Opus 4.7, Fable 5, Haiku 4.5) the panel has tools and notes only, and its last line says so: pick Opus 5.5 or Sonnet 5.5 in the model menu to see how the assistant plans. |
| A note | What the assistant wrote before a tool call (for example Let me check the logs first). While the assistant writes, its text appears here first; it moves to the answer bubble as soon as it reads like the answer (a heading, a list, a second paragraph, a few lines) or the assistant finishes. Notes before tool calls never show in the bubble. |
| A tool call | A sentence such as Read the topology of acme-prod, the input in short and the time. Expanded: the tool's name, its input (SQL and DDL shown as written), the size of its output and its first 600 characters, with email addresses and phone numbers masked. Failed, timed-out and skipped calls are marked; a call the reply was cut off in the middle of is not started and is not counted. |
| Tool budget used up | The lookup budget ran out, so the assistant answered from what it had. |
| Reviewed the answer | The server's checks on the reply: no problems found, which problems a rewrite fixed, or why a rewrite was dropped. |
| Reply cut off or declined | The reply reached its length limit (ask the assistant to continue), or the model declined the request. |
| Stopped | The answer ended with an error, including a connection that closed before the answer finished. Try again. |
- Output previews show only for users who see every customer. If your account is limited to particular customers, tool rows show the tool, its input, size and status, and say the preview is hidden.
- After a reload (or in another window) each answer keeps its steps panel, collapsed. The evidence lookups keep no output preview, and long tool inputs are shortened. Answers saved before the panel existed (October 2026) show fewer details: no durations, no single evidence lookups, no question type, no data sources row and no review step.
How An Answer Comes Togetherโ
- Evidence first. For a question about one cluster, an error, a query or a table design, the assistant collects the read-only evidence the question needs before it starts writing: the cluster's fleet record; for a cluster registered for direct read-only SQL, a live overview (version, nodes, warehouses) and a snapshot of the databases you named with the configuration checks run against it; known-issue matches for pasted errors and logs, and checks on any pasted
CREATE TABLEorALTER TABLE. These lookups appear under Gathered evidence in the steps panel and take at most about 35 seconds; anything unfinished by then is named in the answer as not collected, and a snapshot still being collected is named so the assistant can read it when it finishes instead of taking a second one. They count toward the answer time, not toward the tool-call count. Fleet-wide questions ("most problematic clusters") skip this step. A snapshot is reused within the conversation for 15 minutes, except when you say you changed something, ask whether a change worked or ask for a re-run; pre-flight checks always take a fresh one. - At most one corrected rewrite. Before the reply is final, the server checks it for patterns the assistant must not send: a value or command for a setting that is not approved for change during a customer evaluation, "root cause" without the query's profile or plan, concrete sort or distribution keys when you gave no queries, proposed table DDL that fails the checks, a cluster-wide
SET GLOBALto fix one query, a data expiry (TTL) when you haven't said how long data must be kept, a promise to read a query profile or the query history (no tool can yet), and asking again for a goal you declined to give. If it finds one, the assistant writes the answer once more, and the corrected version replaces the streamed draft in place, unless the rewrite brought in a new problem of its own: then the draft stays. This happens at most once per reply, only on cluster, query, table and error questions, and the conversation keeps only the version you see. The Reviewed the answer step in the steps panel says what was fixed, or why a rewrite was dropped. - Always an answer. Each reply has a lookup budget: eight rounds of tool calls and a time limit (150 seconds by default, set by an admin). When it runs out, the assistant stops looking things up and answers from what it has: the verdict, the evidence, what it could not check and why, and the next step.
- Unavailable sources are named, not waited on. The assistant knows which data sources it can reach right now (the Admin Console for live logs, the Knowledge Lake, the fleet database and metrics, and direct SQL to registered clusters). When the Admin Console or the Knowledge Lake is down, it does not call it: it answers from the sources that are up, says which evidence it could not collect, and asks you to paste logs when logs would decide the answer. If one of the two fails during a conversation, it is skipped for the rest of that conversation and tried again after 15 minutes.
- Few questions, stated defaults. It looks things up before asking, asks at most three questions (each a direct question right after the one-line verdict), and states the defaults it assumed instead of asking: concurrency 1, cold runs not scored, and the goal read from your wording. A few things it always asks first: your queries before it picks sort or distribution keys (until then it shows placeholders such as
<distribution key>), whether counts must be exact (it then gives the exact options, and the approximate ones only for the case that approximate counts are acceptable, saying how the result changes), how long data must be kept before it proposes a TTL, and, for a query id with no profile, the profile or its counters pasted from a re-run with profiling on (it cannot fetch profiles itself). It does not ask for what you already said (goal, concurrency, latency target, how often a query runs), and if you'd rather not say what an evaluation must prove, it does not ask again: it states the defaults it assumed. - Live cluster over billing records. For a cluster's nodes, warehouses and settings it uses the live snapshot rather than billing metadata, gives the snapshot's age, and says which source a number came from when two disagree.
- Read-only, always. It never changes a cluster: no
SET,ALTER, restart or benchmark run. For a setting that is not approved for change during a customer evaluation, it explains what the setting does but gives no value or command, and points you to PhoenixAI support. Pasted text that reads like an instruction (in a log, for example) is treated as data: quoted, never followed.
Good To Knowโ
- Conversations persist โ close the tab, come back, your history is there (shared across browser windows).
- Enter sends; Shift+Enter makes a newline.
- Answers are organised like claude.ai replies: the answer up front, then headed sections, short paragraphs, real lists and tables, and code blocks for SQL and log lines.
- Every message is timestamped to the second. Your question shows when you sent it; the reply shows when it finished and the model. A reply without a steps panel also shows how long it took (answered in 22.4s) and its tool calls; with a panel, its summary has them. Hover a timestamp for the full date and time.
- New conversation resets the context โ do this when switching topics; long mixed threads dilute answer quality.
- Complex questions take longer because the agent runs multiple queries โ the steps panel shows each one as it happens, and the answer bubble fills in once the assistant starts writing the answer.
- Pick the model and effort from the menu next to Send, as in claude.ai: Opus 5.5, Fable 5.1, Sonnet 5.5 or Haiku 5.5 (older models under More models), and an effort of Low, Medium, High, Extra or Max. Higher effort thinks longer: more thorough answers, but slower and costlier. Your choice is remembered in this browser and applies from your next message; until you pick one, the chat uses the model an admin set in Settings โ AI Agents. Each reply shows the model and effort that wrote it, and every conversation's cost lands on the LLM Usage page.
Configuration Advisor (POC tuning)โ
Ask for help tuning a POC or benchmark: the chat consults the Configuration Advisor rulebook, and can take a read-only snapshot of a registered POC cluster and check every rule against it.
- "Evaluate the POC on poc-retail-01: databases
tpcds_1t, goal latency, concurrency 1, cold runs not scored, sessions use the defaults." - "Which rules apply to a 4.0 shared-data cluster with a latency goal?"
- "Is
pipeline_dopsafe to change at runtime on 3.5, and what is its default?" - "Has anyone seen 'BE has overloaded with N pipeline drivers' in a POC?"
Before you evaluate a cluster, have ready the benchmark databases (and tables, if you know them), the benchmark goal (latency, throughput, concurrency, load speed or stability), the concurrency, whether cold runs are scored, and whether the benchmark sessions use the cluster defaults. You do not need all of them: the assistant takes the snapshot first, states the defaults it used for what you left out, and asks only what would change the answer (for a pre-flight check, usually what the evaluation must prove, and against which engine). The Configuration Advisor guide has a message you can copy, more example questions, a sample evaluation and how to read the results.
Known Issues (errors, crashes, logs)โ
Paste an error message, log lines or a crash stack, optionally name the cluster, and ask "is this a known issue?". The chat matches it against curated known-issue cards and tells you what is happening, how to confirm it, the fixes that worked, and whether your cluster's version already has the code fix.
- "BEs on acme-prod keep restarting with exit 139. be.out has this, is it a known issue?" followed by the pasted stack.
- "Scan the logs of acme-prod for known issues from the last 2 hours."
- "If we upgrade acme-prod to 3.5.21-ee, which known issues would that fix?"
- "Known issues about routine load on shared-data."
Paste errors verbatim: exact text matches well, while a plain-language description usually does not. When you find out whether a match was right, say so ("yes, that was it"); the assistant records it for the knowledge-base owners. It records an outcome only when you state one, never in the reply that first shows the match. The Known Issues guide has more examples, sample answers, and what each match type and version verdict means.
Both features only recommend: they never change a cluster, and everything they read from it is read-only.
If your question is about a specific triaged issue, prefer Ask AI Assistant on that issue in AI Triage โ the investigation is saved on the issue itself, uses a problem-specific specialist, and supports follow-ups in place.