Metrics

All metrics in Yavio are derived at query time from the raw events stored in ClickHouse. Nothing in this page represents a separately stored value — each metric is a query over the events table.

Overview KPIs

The dashboard overview shows these high-level metrics:

Metric	How It's Computed
Total invocations	Count of all `tool_call` events
Unique sessions	Count of distinct `session_id` values, without `per_request` sessions (see Sessions); the card then says how many calls came from clients without a session key
Error rate	Ratio of failed `tool_call` events to total `tool_call` events
Avg latency	Average `latency_ms` across all `tool_call` events
Invocations over time	`tool_call` events grouped by time bucket (line chart)
Platform breakdown	`tool_call` events grouped by `platform` (pie chart)
Locations	`tool_call` events grouped by `country_code`, and within a country by `region` (table). Users are counted with the same key as the Users page, per country and per region, so a person seen in two regions is one user of the country. A call without a location is listed as Unknown: in practice only ChatGPT sends one, and only to apps that capture tool arguments; the card shows the share of calls with a location
Top tools	`tool_call` events grouped by `event_name`, ordered by count
Total conversions	Count of all `conversion` events
Total revenue	Sum of `conversion_value` across all `conversion` events

Business KPIs

Metric	How It's Computed
Session-to-conversion rate	Sessions with at least one conversion / total sessions
Widget engagement rate	Sessions with any `widget_click` or `widget_form_field` / sessions with `widget_render`

Tool Explorer Metrics

Per-tool deep-dive analytics:

Metric	Description
Per-tool invocation count	How often each tool is called
Latency distribution	Histogram of execution times for a selected tool
Error rate over time	Error ratio per time bucket for a selected tool
Error category breakdown	Distribution of error types (`auth`, `validation`, `timeout`, etc.)
Parameter fill rates	How often each parameter key appears in tool calls
Top parameter values	Most frequent values per input key
Common parameter combinations	Frequent patterns in `input_keys`
Zero-result rate	Percentage of tool calls that returned empty results
Tool discovery-to-invocation ratio	How often a discovered tool is actually called

Funnel Metrics (Coming Soon)

Funnels follow a conversation's step events in the order they happened. One conversation (session) is one attempt, however many server instances served it:

Metric	Description
Step progression	A conversation's steps, numbered by time; `step_sequence` only breaks a tie within one millisecond. Steps up to 30 days before the selected range count for the numbering; a conversation that began earlier is numbered from its first step within those 30 days
Drop-off rates	Percentage of conversations that don't progress from step N to step N+1
Example traces	Full event timeline for individual traces

Timing Metrics

Metric	Description
Time-to-first-interaction (TTFI)	Time from `widget_render` to first `widget_click`
Time-to-first-tool-call	Time from session start to first tool invocation
Latency percentiles	p50, p95, and p99 latency across tool calls

A session is the set of events sharing one session_id. Sessions are computed from the raw events at query time. Nothing is materialized, and there is no "session close" event. The dashboard counts distinct sessions. Per-session figures like duration or tool count are an aggregation over events grouped by session_id; you can run one yourself through the analytics MCP server's run_query.

A session whose session_source is per_request is not counted — not in the Sessions KPI, not in a user's session count, not in the anonymous statistics. It comes from an HTTP call that carries no session key, on any host (SDK 0.5.4 and later). Since SDK 0.5.5 every such call is a "session" of its own; on 0.5.4 a server that reuses one stateless transport put all its key-less callers into one. Its widget events carry the label too (the widget stamps its server session's source). Rows from SDKs before 0.5.4 — server or widget — have no session_source and count as before. To group such calls into conversations, set sessionKey; paths and funnels need it too, since they work per session.

Users

Users are counted with one key per person:

the user_id from .identify() on the event itself;
otherwise the user of the first .identify() in the same session, so the session's events before the call, and those of other server instances that served the conversation, count for that user;
otherwise the user_id that the same subject_id was identified with, so the person's other conversations count for the right user;
otherwise the subject_id itself.

A session is never counted as a user.

Metric	Description
Active users (DAU / WAU / MAU)	Distinct users in the day, and in the 7 and 30 days up to it
New vs returning	New: the user's first event in the data Yavio holds falls in the bucket
Retention cohorts	Users by the period they were first seen, and how many came back in later periods

"First seen" means within the raw events still stored: 13 months, or the project's shorter retention (Data retention). A user last active 14 months ago counts as new.