Querying with AI
The entry point is the AI button on the left of the query field; in normal query mode you can also press ⌘ + Enter (Ctrl + Enter on Windows and Linux). Which panel it opens depends on the account: when AI SRE is enabled and you have access to AI SRE conversations, the button is Open AI SRE and opens an AI SRE conversation under the query field. When AI SRE is not enabled, or its credits are used up, the button is AI natural language query and opens the basic panel that turns one sentence into query conditions.Query and analyze with AI SRE
Say in one sentence what you want to look at; AI SRE writes the result onto the page itself — the query, the event type and the time range, and for charts the measure, the group-by field and the chart type as well. It then reads back what the page got and answers you, so you can keep asking about the same application.- Status line: while it works, it says what this step is — Filling in the query, Checking (against the real field values in the application), Building the chart, Querying. When the turn ends, the status line reports what the page got: N found, Query complete, no data, Needs your input (it is waiting for your answer), Replied, Stopped, Query failed, Page changed, not applied.
- Undo change: once AI SRE has written to the page, Undo change appears on the right of the status line and puts the query, the event type, the time range and the chart back the way they were before it started. The entry disappears once you edit the query field yourself — the page is then yours and is not undone; it is disabled while a turn runs.
- Stop and collapse: press Stop while a turn runs to end it; whatever was already written stays on the page. Click the status line to collapse or expand the conversation. A turn that wrote a query collapses the conversation when it ends, leaving the results in view; a turn that only answered in words expands so you can read it.
- Follow-up and examples: the status line offers the most likely next request — press Tab to fill it in, Enter to send. When AI SRE asks something back, the status line shows Needs your input and the input prompts you to reply. Before there is any conversation, a few example questions sit under the input: pick with ↑ ↓, press Tab to fill in, or Enter to send.
- Draft while it runs: pressing Enter before the turn ends only keeps a draft — “Still running; your draft is kept, send it when this finishes” — and you send it once the turn is over.
- Another application is another conversation: switching the application at the top right, or leaving the Explorer, ends this conversation, and the next one starts fresh; pressing the button again (it then reads End this AI SRE conversation) ends it too.
- Credits and enablement: when credits are used up (this period’s allowance spent or the wallet empty), AI SRE cannot start a turn: its status line shows Out of credits with Top up, which opens AI SRE billing. After that the button opens the basic natural-language query instead, and the top of that panel states the reason in one line with the same link beside it. AI SRE is enabled and billed separately from any On-call subscription. Accounts without AI SRE always get the basic query.
Basic natural-language query
The button opens this basic panel when AI SRE is not enabled or out of credits: describe the data you want in plain Chinese or English and AI converts it into a standard query (DQL) with the same fields and syntax as a manual query. The input border is highlighted while this panel is active. The conversion reads the sentence against the current application’s platform: native apps (Android, iOS, HarmonyOS, and Flutter / React Native apps) have no browser, page URL, or Web Vitals, and you can ask about crashes, ANR, App Hang, jank, OS version, and device model. A HarmonyOS app freeze is recorded as a crash. The example questions under the input box switch to this wording.1
Describe what you want to find
Describe the data in plain language and press Enter to generate a preview. For example:AI converts it into:You don’t need to remember field names — AI picks the queryable fields for the current event type.
2
Review the preview
The preview shows the query AI returned: the whole query when the box was empty, only the added part (prefixed with
+) when it extends the current query, and — when it replaces the current query — the old query struck through and labelled Replaces the current query with the full new query below it. When the request better fits another event type or time range (for example, “Switch event type to Errors”), that change gets its own old → new row. The preview also says how many events the query matches in the current time range; while it generates, the panel lists the steps the server is taking. If the result isn’t right, refine your description and press Enter again to regenerate.3
Apply
Click Apply, Add to query, Replace query or Switch and apply to confirm: the query goes into the query field, the event type and time range take effect in the same go, and the panel closes — the page then shows the new result. This basic query offers no undo; to get back to the page as it was before AI touched it, use Undo change in an AI SRE conversation (see Query and analyze with AI SRE).
Continuing from the current query
When the query field already holds conditions, AI takes them into account and returns one complete new query, so it can both add to them and change them:- Adding a condition (“also 401”, “only the last day”): the result is the current query plus the new condition, the preview shows only the added part (prefixed with
+), and the button is Add to query. - Changing or dropping a condition (“drop the duration condition”, “only
/incident”, “switch to Safari”): the new query no longer contains some of the old conditions, so the preview strikes out the current query, labels it Replaces the current query, and the button becomes Replace query. - Changing the subject (a request unrelated to the current conditions): AI ignores the current query and writes a new one for your description, which is applied the same replacing way.
- With an empty query field the button is just Apply; when the event type has to change it is Switch and apply, and the event type switches along with it.
- When the returned query is exactly what the box already holds and nothing else changes, the panel says “The query already says this”.
Check the count before applying
Once the query is generated, the server runs it once against real data in the time range and the preview shows the outcome:- N will match when there is data; Nothing matches in the current time range (in warning colour) when there is none — you can still apply it.
- On zero hits AI first tries to correct itself: it checks case and spelling against the real values of those fields in this application (for example
Mobile,Mac OS X) and, if needed, moves to a better-fitting field or event type — up to two rewrites. If it still matches nothing it returns the query as written: when your conditions are right and there is simply no such data, zero is the answer. - While it works, the panel lists the steps the server is taking: Understanding the request › Checking › Writing the query › Checking the count; a rewrite shows Rewriting the query with its reason (Nothing matched, trying another way, Switching to Pages, Fixing the query, Trying again).
- If the count is unavailable it does not change the result: the query comes back as usual, just without a count.
Analysis questions go to AI SRE
Requests that want a conclusion — “why is this page slow”, “do we still get this error after the release”, “how does this compare with last week”, “which page is slowest” — cannot be answered by the query box. AI classifies them as a question rather than a filter condition and the panel says This is a question for analysis; a query cannot answer it, with the way to AI SRE beside it: AI SRE can analyze this plus Enable AI SRE when AI SRE is not enabled, or Top up when credits are used up.- Time expressions such as “last hour,” “yesterday,” and “past 7 days” update the Explorer time picker; they are not written as DQL conditions such as
client_time. Performance-duration conditions remain DQL, for exampleview_loading_time:>2s. - The time range is limited to 14 days. If you ask for more, the preview says the range was truncated to the most recent 14 days. A relative range becomes the past 14 days. An absolute range keeps its end time and starts 14 days earlier.
Query basics
Queries support two types of terms:
A query that is just a single
* is the same as no condition: it matches every event in the time range.
Boolean operators
Full-text search
Escape special characters
When searching for field values containing special characters, you need to escape them with backslash\ or use double quotes.
The following characters are considered special:
:, ", *, -, >, <, ,, (, ), [, ], \ and spaceAttribute search
Useattribute:term syntax to search specific attributes:
Numeric search
For numeric type attributes, comparison operators can be used:Complex query examples
- Error Analysis
- Performance Analysis
- Error Requests
- Page Behavior
Search for Warning type errors that occurred in the wallet page:
Path fields and URL normalization
view_url_path and view_url_path_group (and the same-named path fields on session, error, resource, action and long task events) are derived by normalizing the page URL, and both can be filtered and charted on. Grouping by view_url_path is a high-cardinality grouping (Count and Sum only, time range up to 7 days; see Cardinality limits on the group-by field); to summarize by page, group by view_url_path_group:
- Query parameters (
?tab=1,?page=2) do not enter either field. - A hash route stays in the path: a fragment starting with
/(how hash-mode SPA routers write routes) is appended to the path and the path group, so every page of such an application has its own value. Filter or group by page with the#/...part included, for exampleview_url_path:/app/#/list. - A query inside the route is dropped (
#/list?page=2is recorded as#/list), and a bare#/or a plain anchor such as#section-2is ignored. - ID segments — numbers, UUIDs, object IDs, trace IDs, long hashes, dates — become
?inview_url_path_group, so a detail page with one address per object falls into one group;view_url_pathkeeps the original value. - Resource URLs do not follow these rules:
resource_url_pathtakes the path only, and a#...on a resource address has always been dropped.
WeChat Mini Program-Specific attributes
For View events collected from the WeChat Mini Program platform, the following queryable attributes are available in addition to the sharedview_* fields. They help diagnose cold-start latency, setData cost, and lifecycle-stage durations:
For example, to find Mini Program pages with a first-screen render longer than 2 seconds:
Session attribute search
In addition to the fields in the examples above, sessions support the following common queryable attributes, which can also be used for column display, CSV export, and link variables:
For example, to find sessions uploaded by on-error capture:
sessionOnError) and Capture additional replays on error (sessionReplayOnError).
rc_version corresponds to the published Remote Configuration version of the application, letting you filter sessions by configuration version and evaluate the rollout of a remote configuration release.
Advanced search tips
Time Range Search
Time Range Search
Perform precise searches combined with time range:
User Behavior Search
User Behavior Search
Search for user click behavior on the checkout page:
Device Type Search
Device Type Search
Search for views with loading time over 3 seconds on mobile devices:
Geographic Location Search
Geographic Location Search
Search for sessions with errors in China region:
Best practices
- Wrap multi-word phrases in double quotes for exact matching.
- Use wildcards sparingly to avoid overly broad conditions.
- Combine conditions with
AND/ORto build precise queries. - Rely on auto-complete to avoid typos.
- Save frequently used search conditions for repeated queries.
Next steps
RUM Explorer Overview
Learn about Explorer core features
Distributed Tracing
Learn distributed tracing best practices