Searching events

Press ⌘K (Ctrl+K on Windows and Linux) anywhere inside a project and type at least two characters. Alongside the pages and commands, the palette searches the project's catalog on the branch you are on: events, event types, fields, meta fields, properties, relations, tags, metrics, fact tables, scans and alert rules. Results come back best match first, grouped by type. The feature reference describes the rest of the palette.

Variants are folded into one row

A scan that names events from a data column makes one event per value it sees. A screen column gives you event_name=Screen View | screen=Home, event_name=Screen View | screen=Map, event_name=Screen View | screen=Cart, and so on. Search for "screen view" and every one of them matches equally well. Listed one by one, they pushed everything else off the list and looked like duplicates.

The palette now shows such a family as one row: the best-ranked member, followed by + N variants. Right below it, a Show N variants row expands the rest in place. Select any member to open that event. Hide N variants folds them away again.

Two events count as variants of each other when both of these hold:

• they belong to the same event type, and
• their names differ in the value of exactly one placeholder of the naming rule that produced them.

The naming rules are the project's scan event name formats (such as Screen View: {screen} ({platform})), plus the default column=value | column=value name a scan writes when it has no format. Some examples:

Names	Folded?
screen=Home | action=tap and screen=Map | action=tap	Yes: only screen differs
screen=Home | action=tap and screen=Map | action=swipe	No: two placeholders differ
screen=Home in Screen view and screen=Map in Click	No: different event types
Home Screen View and event_name=Home Screen View | screen=Home	No: the first is a hand-written name that no rule produced

When a name could vary on more than one placeholder, the one that gathers the most matching results wins. A format that cannot be read back reliably is never used to fold anything. That covers a bare {screen} with no fixed text, two placeholders with nothing between them, and a placeholder used twice. A name whose value contains the format's own fixed text is left unfolded as well: with {screen} | {action}, the name A | B | Click could be screen A or screen A | B, so tripl does not guess.

Folding only changes how results are listed. Ranking stays the same: the row sits where its best member ranked, and the members inside it keep their order. The palette asks for 12 rows, and a folded family counts as one of them, so other results are not squeezed out.

Over the API

GET /api/v1/projects/{slug}/search folds variants only when you pass group_variants=true. The default is false, so existing callers get the same answer as before. With it on:

• the representative result carries a variant_group object:
  • key: stable for the same family across searches
  • pattern: the name with the varying value shown as {placeholder}
  • placeholder
  • count: every member, representative included. Members are gathered from the top 100 matches only, so when truncated is true the count is a lower bound
  • variants: the other members, best-ranked first, each with its id, event_id, title, the value it substituted, route_path, score and confidence
• limit, total and truncated count rows, and a folded family is one row. truncated is also true when more matches exist past those top 100, even if the folded rows fit on the page.

Every other result has variant_group: null.

Search by meaning, per organization

Besides matching words, search can also match by meaning when semantic search is on for your organization. Each organization chooses its own embedding provider under Settings → Organization → Search: an OpenAI-compatible endpoint, model and API key, or the operator's, which it inherits by default. Your projects' text is only ever sent to your organization's endpoint, and a query is only compared with vectors your organization's endpoint and model produced. Another organization's vectors are never compared with yours.

• Changing the model or endpoint re-embeds your organization's projects in the background. Until that finishes, results come from word matching plus whatever is already embedded with the new model.
• With semantic search off (switched off, or the operator runs with ORG_SETTINGS_OPERATOR_FALLBACK=none and your organization has no endpoint of its own), search still works and ranks by word matching. Over the API the response reports semantic_used: false.
• The demo project keeps its built-in example vectors while semantic search is off, so meaning search can be tried without a key.

Owners and admins see where each value comes from and can go back to the operator's with Inherit. The model must produce 1536-value vectors; a model that answers with another size is refused when you save it. See Operator and organization settings.