Ask: find sessions by describing them

Filters find sessions by facts you can name. Some behaviour has no single event to filter on. Ask is for that behaviour.

Describe what happened, in order. Use the words that your events use: button text, page paths such as /settings/billing, and custom event names such as plan_changed.

Examples of behaviour that Ask can find:

• "Opened billing settings, started changing the plan and left without saving."
• "Clicked Send invite several times and nothing happened."
• "Hit an error at checkout, then opened the help widget."

• A candidate has a session id and at least 3 events in your date range that match your applied filters. Ask takes up to the 2,000 most recently active sessions.
• If background judging is on for your project, the sessions that Watch ranks highest come first. The rest follow, newest first.
• A page holds up to 200 sessions.
• The judge gives two answers for each session: how likely the session shows what you described, and which event shows it best. It writes no summaries or explanations.
• For each session, the judge reads a summary of the events. The What the judge reads block lists what the judge gets.
• Sequential trims long sessions to 255 events. It keeps errors, rage clicks, dead clicks, form submits and page views ahead of custom events and ordinary clicks.
• A narrow filter can drop sessions, but for a session that stays, the judge still reads the other events.
• Ask reads events, so it also finds sessions without a recording.

The judge reads the same event summary that background judging uses (Judgments):

• page paths, without domains or query strings
• click text
• form names
• error messages
• custom event names, and properties whose keys do not look personal
• session facts: duration, device, browser, OS, country, entry page, page count, whether the user was identified, and any DevTools error counts

Sequential does not send the session's user id.

• The Match meter shows the confidence as a percentage.
• Each result also shows the session: id, user id, start time, duration and event count. The id opens the replay.
• The start time and event count cover only the events in your range that match your filters.
• If you cannot play the recording now, the id and an Open session button both open the Events page. A chip shows Processing · ready by and a time while the recording is being processed, or Expired after the replay period. A Processing result checks again 5 seconds after the time in its chip. When the recording is ready, the chip goes away and the id opens the replay. You do not have to search again. If Sequential cannot read the states in time, a result counts as ready when Sequential has an index of its recording. When that result comes into view, Ask asks for its state. If Ask cannot get the state either, also on a second try, a result without an index shows a muted Couldn't check recording chip. Ask cannot ask for a result that has no usable start time or duration. Search K more keeps the newer states of the results already shown.
• Watch from here opens the replay on the Insights tab. That tab judges the recording when you open it, unless the recording has a saved judgment. It uses a separate allowance (Session insights).

• N is the number of sessions that the judge judged. Ask skips sessions with fewer than 3 readable events.
• M is the number of candidates. If M is 2,000, Ask may have left out older sessions in your range. To include them, narrow the range or add filters.
• Search K more judges the next page of up to 200 sessions. K is the number of sessions that Ask counts as left to search. It adds new matches below the current results.
• "· some sessions could not be judged" means that part of the page failed.

With background judging on, the first page favours the sessions that Watch ranks highest. So to find an ordinary behaviour, such as a smooth first purchase, you may need to click Search K more.

• Sequential does not save an answer when part of the page failed. If you run the same question again, the judge judges the whole page again.
• Each click on Search K more is a new run.
• K can be higher than what is left, because Ask counts skipped sessions as remaining.
• Ask orders the list again on every click. If new sessions arrive, or background judging judges sessions, between clicks, Ask can judge some sessions twice and skip others. A session judged twice counts twice in N.

• Ask in the app has its own allowance: 2,000,000 input tokens per project per UTC day by default. Background judging, Try it in the rubric editor and Ask from an agent each have a separate one.
• Your Sequential server sets the allowance, with ASK_DAILY_TOKENS.
• If Sequential cannot read the allowance, it does not run Ask and spends nothing. The server answers 503 budget_unavailable with Retry-After: 30. A saved answer still returns.
• The allowance resets at midnight UTC. The "resets …" message shows the time in your local time.
• Sequential saves an answer for one hour, unless part of the page failed. The same question over the same sessions returns the saved answer. It does not judge again or use the allowance.

• Retry returns the saved answer only if the judge judged the whole first page within the hour and the same sessions are still first in line.
• Ask ignores double quotes, backticks and extra spaces in the question.
• Ask does not ignore apostrophes and capitals. A question that differs only in these is a new question.
• If new sessions change the page, Ask does a new run.

Readable events are clicks with text or an id, page views with a path or screen name, errors with their message, and named custom events (Events, Send custom events).

Each result has thumbs-up and thumbs-down buttons next to Match. A click saves your verdict with your question. It does not judge again or change the order of the results.

• Your verdicts count towards "Right X of Y you checked" in the Watch header. That line appears when Watch has judged sessions.
• Your verdicts do not count towards the calibration page (Feedback and calibration).
• Only your latest Ask verdict for each session counts.
• The thumbs start blank each time the results load.

• When you run Ask, Sequential adds the question to the page address. For example: /sessions?mode=ask&q=gave+up+at+checkout. If someone opens that link, Ask runs at once.
• The link carries only the question. The recipient's selected project, date range and filters apply.
• Unless a saved answer matches, the link starts a new run on the allowance.

• An empty results with near_misses above 0 means that the judge read sessions and none matched.
• A near miss has a match of at least 0.2 and less than 0.4.
• After a stop, Sequential did not check the rest of the page. Call again with next_cursor. For budget_exhausted or token_daily_limit, call after midnight UTC. For ai_unavailable, call again in a minute.
• A cursor used hours later, for example after midnight UTC, continues in a list that has changed. Some sessions can be skipped or come again.
• Each result has session_id, user_id, session_start, match (0–1), state (on or unclear), moment, the session's judgment in beliefs, and replay_url.
• replay_url and the url of a moment name the project, so they open that project. See Links that open a project.
• The app shows the same two counts under its results. See What a result shows.
• For each refusal, see Troubleshooting.