AI Chat
QDash provides conversational analysis tools that use project calibration context without replacing operator judgment.
Choose a Tool
| Tool | Use it for |
|---|---|
| AI Chat | Explore a question interactively with project-aware tools and cited QDash records. |
| Analysis chat panel | Discuss the metric and filters already open on the Metrics page. |
| Agent calibration | Run a separately authorized, bounded calibration plan through the client. |
Agent calibration performs hardware-affecting operations and is documented separately in Agent Calibration.
AI Chat
Open AI Chat for a project-wide conversation. Start a new session when the chip, investigation, or goal changes substantially. QDash stores sessions so you can return to prior analysis.
The assistant can use the project tools exposed by QDash to inspect records such as chips, metrics, task results, forum threads, workflows, and provenance. Keep the active project and selected chip in mind when interpreting an answer. Open referenced records to confirm important conclusions.
Ask AI beside any page
Ask AI on the right edge opens the same chat as a sidebar next to the page you are on. The assistant is told which page is open with every message, so "what am I looking at?" or "compare this with last week" refer to the current page. Drag the sidebar's left edge to make it wider or narrower; double-click the edge to restore the default width. The width is remembered in your browser.
On the Workflow editor the assistant also sees the flow's code as it is in the editor, unsaved changes included. Ask for a change in plain terms, for example "add a T2 echo check after T1". The answer shows a diff against the editor; Apply to editor puts the new code into the editor and unlocks it, Discard leaves it as it was. Nothing is saved until you select Save, which checks the flow the same way as a hand-written one. Apply is available only while the proposal's flow is open on the same chip.
A workspace for calculations and scripts
When the deployment enables it, the assistant has a private workspace for each chat: a shell with Python, NumPy, SciPy, pandas, Plotly and Matplotlib, no network, and no access to QDash. It uses it for arithmetic, fitting and data wrangling, writing data from tool results into files. The answer's trace shows each command, the file it wrote with the code typed out as it is written, and the output of each run. Files stay for the rest of the chat and are deleted with it.
Thinking and your own instructions
The Thinking menu next to the model picker sets how much the model reasons before it answers: Auto (the model's configured default), Off, Low, Medium, or High. The choice applies from the next message and is remembered in your browser.
Settings → Copilot → Your instructions holds instructions added to every chat you start, after QDash's own rules: language, tone, and formats such as "answer in Japanese, conclusion first" or "put numbers in a table with units". They are stored with your account, so they follow you across devices. They do not apply to AI evaluations or forum replies, which stay the same for everyone, and they cannot skip the approval of write operations. A Your instructions link in the chat composer shows when they are on.
What Copilot is told, below the instructions, shows read-only what the application sends the chat model: QDash's system prompt, your instructions as they are appended to it, and the tools the chat offers. It is the running text, so it reflects the current skills, tools, and language settings; the application's part is edited in the deployment, not in Settings.
Investigate a failure
Use Investigate with AI when an execution or a task failed and you want to know why. The button appears in the Error Log of a failed execution on the Execution page and on its detail page, in the Error Log of a failed task result, and on a failed task in an execution's task details. It works when the execution failed before any task result was recorded.
- Select Investigate with AI in the Error Log header.
- The chat opens and sends the investigation for you. On the Execution pages, in task details, and in a task result preview, it opens in the floating chat window, so the details panel stays in view; on a task result's own page it opens in the chat sidebar. QDash adds the recorded error message and stack trace, and the execution, task, qubit, and chip identifiers to the chat.
- Read the answer, then ask follow-up questions in the same chat.
The assistant reads available source code, starting from the code that raised the error. The answer is short: the likely cause, the evidence (a quoted log line and cited code locations), and one next step to check or a fix to review. The assistant reads more code only when the evidence is unclear; ask a follow-up question for a deeper analysis. It does not apply fixes or re-run anything.
QDash does not record which source revision an execution ran with. The answer names the revision of the source it read, or says it is unknown, and says that it cannot confirm that this source matches the failed run. When the logs are missing or too long, the answer says what is missing.
Selecting the button again for the same failure reopens that chat instead of starting a new one. To run the investigation again from the start, select New chat while the investigation chat is open, then select Investigate with AI. The earlier chat stays in your chat history.
Run a calibration pipeline from chat
The assistant can compose a calibration from QDash's step catalog and run it as one execution with your approval and project permissions. Ask for the calibration in plain terms, for example "run the coarse one-qubit check on Q00 and Q01, then measure T1 and T2 on the ones that pass". The assistant reads the available step types and tasks, validates its plan with QDash, and shows an approval card listing the targets, the steps, and the exact tasks each step will run. Nothing starts until you approve the card.
The plan can only use the step types and task names QDash exposes; it cannot contain code, and every task result is judged by the same gates as a saved workflow. The execution appears on the Execution page tagged pipeline. See Calibration Pipeline Spec for the spec format.
Metrics Analysis
The Metrics page includes a chat panel that opens with the selected task result as its context. Use it when a question depends on the selected chip, metric, target direction, or time range. Changing those filters can change the records available to the chat; state the intended comparison explicitly.
Ask AI on a task result offers two routes before anything starts. Evaluate the figure creates a recorded evaluation with the settings described in AI Evaluations. Chat with Copilot opens the conversation described here. Viewers without edit rights go straight to the chat.
Evaluate Findings
Treat AI output as supporting evidence. Before changing calibration state, excluding a result, or starting another hardware run:
- open the underlying task result and artifacts;
- compare the finding with task knowledge;
- confirm the project, chip, target, and time context;
- use deterministic quality gates where the workflow provides them;
- record the decision in a note or a forum thread when others need the context.
Provider availability and model selection depend on the deployment configuration. A missing AI provider disables that capability rather than the rest of QDash.