How to work with cases in Studio

When you need this

Use this when you want a flow to run on saved, realistic inputs instead of test input you retype after every change. A case is one input with its context, the upstream outputs a range of the flow needs, an optional expected output and tags. Cases live in datasets/<dataset_id>.yaml. The Cases tab is where you read them, add to them and run them. Experiments select their cases from these same files by tag.

Steps

• Open a flow and click Cases in its tab bar, next to Graph and Runs. The picker at the top lists every dataset in the project, searchable by ID, with its case count and the flow it belongs to. A line under the heading says which kind you picked:
  • Cases of this flow can run from here.
  • Cases of another flow run only from that flow’s tab.
  • Cases without a flow feed experiments whose subject is a flow local to the experiment, not runs from this tab.
• Filter by tag, then by name. Filter lists the tag dimensions the cases use, such as length, channel or regression, then the values of the one you pick, with a count for each. Each filter you add shows as a chip. The search box narrows the list by case name, and the counter says how many cases are shown of the total.
• Read the list. One row per case: its name, its tags, a mark when it has an expected output, and the upstream nodes whose outputs it saves in node_outputs. Click a row to open the case in place. You see its input, context, expected output, node_outputs and tags. You also see Experiments using this case: every experiment whose dataset and tag filter select it, with a link to each.
• See the case’s media. An image, audio, video or document in a case shows in the open case, whether the case points at a file in the project (file) or at a blob (blob_id). A file is read straight from the project folder, so what you see is the file on disk.
• Run one case. An open case of this flow has a Run this case panel. Drag the stage picker to choose the first and last stage, or click one node to run only that stage. Studio checks live that the case has what the range needs, and earlier stages take their outputs from the case. Start run runs it once and opens the run on the Runs tab.
• Run several cases as a look. Tick the cases, or tick the header to select every shown case. A selection bar appears at the bottom with the count, Clear, a Stages picker (from stage, to stage) and Run. The bar says whether the range starts from the whole flow or takes earlier stages from node_outputs. When a start can’t work, it names the cases that lack the outputs it needs. Run starts a look series: every selected case runs once over those stages, with no verdict and no finding, and Studio opens the series. A case with an expected output is checked against it. A look takes at most 500 cases.
• Add cases with the chat or a CSV. Ask the agent to add cases (or Ask the agent to write cases, when the flow has no dataset yet) sends a prompt to the chat. The agent first asks which situations the cases should cover and how many. It gives each case a unique name, tags with the same keys as the other cases, and an expected output when the answer is known, then runs aqven check. Upload CSV opens a dialog with a guide to the columns this flow expects and a downloadable example. Check CSV previews the import before anything is saved: matched columns, valid rows, the nodes each row can start from and the problems to fix. Create dataset turns on only once the check passes.
• Edit cases in the file. A saved case isn’t edited in place on this tab. Change datasets/<dataset_id>.yaml, or ask the chat to. Never rename a case that experiments already use: the server splits working and held-out cases by the case’s name, so a renamed case is a new one.
• Attach a media file to a case. Studio writes the file into datasets/<dataset_id>/, next to the dataset, with a name that doesn’t clash with the files already there, and saves a file reference in the case instead of a blob. The file goes into git with the dataset. For private media, see how to keep case media as files in the project. Media from a CSV import and cases drafted from a run point at blobs. Run aqven datasets materialize to turn them into files.
• Lock a fix with a case. From a run that went wrong, To cases drafts a dataset case from that run and hands it to the chat. See How to investigate a run.

Example

Open the support_case flow’s Cases tab and pick support_case_cases: twelve cases, tagged by length, channel, lamp_kind, action and regression. Add the filter regression: yes and five cases remain. These are the cases the reply_look experiment selects, and each one lists reply_look under Experiments using this case.

Select the five, set Stages to polish through polish, and click Run. The bar reads “Earlier stages come from node_outputs”, because each case saves the outputs of prepare, triage, route, search_kb and panel that the revision loop reads. Studio starts a look and opens it: five attempts, one per case, each with its outcome, cost and a link to its run.

See also

• How to use Research in Studio: the experiments that select these cases, and the series that answer them.
• How to follow and read a series in Studio: the page a look opens on.
• How to write an experiment: selecting cases with cases.tags.
• How to keep case media as files in the project: the two path forms, private media and moving old blob-backed cases to files.
• Datasets reference: every key of a case.
• How to investigate a run: the run that Start run opens, and the To cases action.