Loading docs
Review generated Qiskit, Cirq, and OpenQASM output beside the visual workflow before routing or exporting.
Generated code is not a separate product. It is a view of the same workflow record, useful for review, export, and collaboration with SDK-oriented teammates.
Use the code tab to verify that visual edits produce the expected source. Then route or export only after the code view matches the team’s objective.
Section 01
When a block property changes, return to generated code and confirm the source reflects the intended operation.
This is the best way to keep visual builders from becoming vague. The canvas shows purpose; generated code confirms exact execution structure.
Section 02
The AI builder can draft a workflow path from plain English, but the output should remain inspectable. Treat generated source as a review surface, not as invisible automation.
Teams should verify the circuit, code, route fit, and review state before execution.
Workflow AI saves each conversation inside its workflow. Reopen a chat to continue it, search past messages from History, or use New chat for a fresh conversation in the same workflow. Recent messages and a short summary of older turns give the assistant bounded context; the live circuit and editor take precedence. Draft chats follow the workflow when it is first saved. Cancelling a response or reaching the provider timeout closes its usage reservation; retries stay attached to the original usage receipt. A single oversized prompt is rejected before provider execution.
OpenAI Responses requests disable provider-side storage and reasoning summaries. Model routing uses the enabled provider and plan configuration maintained by administrators; generated code still requires your review before a run.
Research defaults to Automatic in both workflow and code assistance. The assistant retrieves published QFlow, Academy and Hub references, searches directly for changing facts, and checks the web when local references do not cover a technical error or algorithm evaluation. When tool support and quota allow it, the model can also decide to look up a remaining factual gap. Searches preserve the actual question; official SDK references and primary research sources are preferred. Choose Search the web to request a lookup, or Off to skip source retrieval. Retrieved links appear under Sources and remain with saved chat history. A failed lookup is shown as unverified rather than presented as current evidence.
Ask for an explanation, review, diagnosis, algorithm comparison or implementation plan in your own words. The assistant chooses a suitable response without requiring a mode selection. Plans include dependencies, relevant resource assumptions, a baseline experiment and concrete validation steps. A cited source supports an explanation; it does not prove that the proposed code ran, that an algorithm achieved a benchmark, or that hardware is accessible. Applying code and submitting runs remain explicit actions.
Closing the assistant or browser does not stop an accepted response. Reopen the same conversation to recover its answer and sources; use Stop when you intend to cancel. Source-based code changes still appear as proposals for inspection before applying them.
Administrators can enable web research and choose public search, Brave Search API or Tavily under Admin → AI → Sources and web research. Public search is best effort. Dedicated API keys are encrypted, saved values are never displayed, and blank key inputs preserve an existing key for the same service. Search API charges belong to that account.