Roivenue

List the roivenue accounts this connection can access, as [{id, name}]. Call this when no account is pinned by server config: present the options and ask the user which account to explore, then pass the chosen id as `account_id` to the other analyst tools. If an account is pinned server-side you don't need this — the tools use it directly. In answers and citations name accounts by NAME — ids are internal.

list_accounts

List the distinct values of a dimension for the account (e.g. the sites in `site`), ranked by `measure` (default netRevenue) over [date_from, date_to]. Use this to enumerate scope options (sites, markets, conversion types) before querying or before asking the user to choose. Omitted attribution_model defaults to the user's Explorer model; without one, attributed measures sum across EVERY model and read ~Nx high. Relative windows ('last 7 days', 'year to date') exclude today and end YESTERDAY; get exact app-matching dates from `resolve_date_range` instead of computing them. Omit dates/attribution/currency to use the caller's Explorer defaults; `defaults_applied` echoes what was used — disclose it. Pass `account_id` (from list_accounts) unless the server pins one; a pin always wins.

list_dimension_values

Load focused guidance for the kind of task you are about to do, and follow it. Call this BEFORE answering when the question matches one of the skills — path-analysis (customer-journey / touchpoint-path questions, channel overlap, order drill-in), reconciling-numbers (the user's figure disagrees with yours), attribution-and-dimensions (zero revenue with non-zero cost, or a name that could live under several dimensions), budget-reallocation (where to shift spend), data-health (stale or broken-looking data), charts (an explicit chart request). Needs no account or token. Most questions need no skill at all.

get_skill

Call this FIRST, before your first data answer, to mirror the calling user's Explorer defaults: their selected attribution model, currency, default date window (as an object plus human-readable text) and time granularity, and the accounts they can reach. Compute data answers with these defaults so the numbers match what the user sees in the Explorer app — and say so when you do. Works with no account selected (it is about the user, not an account); `account_id` is optional and used only to resolve the attribution model to its canonical code. Name accounts by NAME in answers, never ids.

my_settings

Build a deep link that opens a roivenue Explorer page with a chosen scope — give the user a clickable follow-up after a data answer. `page` is one of: performance (Performance analysis), dimension-breakdown (Dimension breakdown; takes `dimension`), path-analysis (takes `dimension` for the journeys table), attribution-analysis (takes `dimension`), transaction-history, budget-optimizer, dashboard. Optionally scope with account_id, date_from/date_to (+ `granularity`: daily|weekly|monthly|aggregate, default daily), `attribution_model` (a model **code**, e.g. from account_overview — never a display name) and `filters` like [{"name": "site", "values": ["kytary.cz"]}] — the page opens exactly as if the user had picked those in the app. Pure URL building — needs no token; a server-pinned account wins over account_id. Returns {url, note}: present the url as a Markdown link, e.g. [Open in Explorer](url). Call this BEFORE you start writing the answer, so the answer is written once, whole, after the last tool call.

explorer_link

The Explorer **Path analysis** page's KPIs for an account over [date_from, date_to]: average path length, average days to convert, the share of conversions the attribution model redistributes away from last-touch, and the touchpoint-type mix (amount plus a 0-1 share). `attribution_model` takes a model **code** from describe_metrics, never a display name (the AI Data Driven code varies by account: 'RnnProbability' or 'aiDataDriven') — and must be MULTI-TOUCH: under last-touch every path collapses to one touchpoint and path metrics are degenerate. An empty result carries a `note` — follow it. Money is EUR-pinned. `Visit` is the effective click+visit bucket (the ETL remaps clicks to Visit, so Click is ~0). `exclude_stitched_paths` defaults **true** to match the app's fresh-open view; set it false to include cross-device-stitched journeys (which raises the redistribution share). include_impressions / include_synthetic_impressions add those buckets and raise the totals. Scope with `filters` like [{"name": "platformName", "values": ["Google Ads"]}] (optional "operator": in|notIn|contains|notContains; contains takes one "value"). Relative windows ('last 7 days', 'year to date') exclude today and end YESTERDAY; get exact app-matching dates from `resolve_date_range` instead of computing them. Omit dates/attribution/currency to use the caller's Explorer defaults; `defaults_applied` echoes what was used — disclose it. Pass `account_id` (from list_accounts) unless the server pins one; a pin always wins.

path_kpis

A distribution histogram from the Explorer **Path analysis** page for an account over [date_from, date_to]. `kind`: 'days_to_convert' (first touch to conversion) or 'touchpoints_to_convert' (touchpoints per converting path). Returns {kind, bins:[{title, amount, share}]}. Bins are FIXED server-side — days: 1, 2, 3-7, 8-15, 16-30, 31+; touchpoints: 1, 2, 3, 4, 5-10, 10+ — custom bins are not possible. `attribution_model` takes a model **code** from describe_metrics, never a display name (the AI Data Driven code varies by account: 'RnnProbability' or 'aiDataDriven') — and must be MULTI-TOUCH: under last-touch every path collapses to one touchpoint and path metrics are degenerate. An empty result carries a `note` — follow it. Money is EUR-pinned. Takes no toggles. Scope with `filters` like [{"name": "platformName", "values": ["Google Ads"]}] (optional "operator": in|notIn|contains|notContains; contains takes one "value"). Relative windows ('last 7 days', 'year to date') exclude today and end YESTERDAY; get exact app-matching dates from `resolve_date_range` instead of computing them. Omit dates/attribution/currency to use the caller's Explorer defaults; `defaults_applied` echoes what was used — disclose it. Pass `account_id` (from list_accounts) unless the server pins one; a pin always wins.

path_histogram

Validate and DRY-RUN a draft custom (derived) dimension on live data — nothing is written. Invalid drafts are REJECTED with the valid options; returns per-value coverage vs the 'Undefined' bucket, rule conflicts, and sample rows. `when` = OR-list of AND-groups; operators: contains|startsWith|endsWith|equals|regex|in|notIn|isEmpty|isNotEmpty|not-variants|greaterThan|lowerThan|greaterOrEqualTo|lowerOrEqualTo (case-INSENSITIVE; only equals honors case). `then.type`: constant|other_dimension|expression|regex_group. Conflicting rows: preview shows the LAST matching rule, production keeps the FIRST. Run this and show coverage/conflicts before create_custom_dimension. Pass `account_id` (from list_accounts) unless the server pins one; a pin always wins.

preview_custom_dimension

Query an account's marketing performance: aggregate `measures` by `group_by` dimensions over [date_from, date_to] (ISO-8601). Use describe_metrics for valid codes; pass attribution model **codes** (e.g. 'LastTouch') in attribution_models. Scope with filters shaped like [{"name": "site", "values": ["<a site value>"]}]. Leave `granularity` unset for a single total per group — it defaults to 'aggregate'; the API's own default is per-DAY rows, so only set granularity to 'Daily'/'Weekly'/'Monthly' when you actually want a time series; 'Weekly'/'Monthly' cover whole weeks/months, so the window is widened to those boundaries for you and the widening reported in a `_window_snapped` row — report THAT window, and don't retry with hand-trimmed dates. A Daily/Weekly/Monthly series returns up to 400 rows in one call — never chunk a year of days into multiple queries. A `_suspect` row means the result is the wrong-dimension/wrong-attribution signature, NOT real numbers: re-check before reporting it. Omitted `currency` defaults to the user's Explorer currency, converting a multi-market account to that single currency exactly like the app (per-market native-currency rows appear only if no default currency is available); use account_markets to see each market's own currency. Relative windows ('last 7 days', 'year to date') exclude today and end YESTERDAY; get exact app-matching dates from `resolve_date_range` instead of computing them. Omit dates/attribution/currency to use the caller's Explorer defaults; `defaults_applied` echoes what was used — disclose it. Pass `account_id` (from list_accounts) unless the server pins one; a pin always wins.

query_performance

Read the full text of a knowledge/ doc by its path (as returned by search_knowledge). Only paths under knowledge/ are served. Always cite the path in a 'Sources:' line.

read_doc

Resolve ANY relative window ('year to date', 'last 13 weeks', 'previous month', 'this quarter') to the EXACT dates the Explorer app's date picker would show — use this instead of computing dates yourself. Pass a `preset` (one of: this_week, previous_week, last_7_days, last_14_days, last_30_days, previous_3_months, previous_month, last_month, this_month, this_quarter, year_to_date (alias this_year), previous_year, last_13_weeks) OR the explicit triple `from_type` ('this'|'last'|'previous') + `period` (day|week|month|quarter|year) + `multiplicand` (default 1); the preset can also be a dynamic triple pattern like 'last_5_days', 'previous_2_weeks', etc. Today is always excluded, so day/aggregate windows end yesterday; `granularity` (aggregate|daily|weekly|monthly, default aggregate) snaps the range to whole weeks/months exactly like the app (a weekly year-to-date starts in the prior year and can extend past yesterday). Pass granularity = the user's defaultTimeGranularity (from my_settings) to mirror their app page — path tools consume the page's snapped range — or leave 'aggregate' for plain totals. `first_day_of_week` defaults to monday (the account default). Needs no account or token — pure date math. Returns {date_from, date_to, conceptual, granularity, note} (plus 'resolved' if preset was dynamically parsed).

resolve_date_range

Search roivenue's knowledge base. Returns matches as {path, excerpt}; cite the paths in a 'Sources:' line. If the list is empty, answer starting with 'NO_SOURCE:'. Use this before answering any roivenue question; never answer from outside knowledge.

search_knowledge

Record the user's feedback on Rai's most recent answer. Call this whenever the user reacts to an answer — e.g. 'that's wrong', 'that helped', 'this response was really bad'. Use rating 'up' or 'down' and include the comment plus the question/answer being rated so the record is actionable. After recording, acknowledge briefly; if negative feedback is vague, ask one short follow-up question and record the detail with another submit_feedback call.

submit_feedback

The Explorer Transaction-history page's order/conversion list for an account over [date_from, date_to]: one row per conversion — {pathConversionId, timestamp, orderStatus?, netRevenue, grossProfit?}. Customer identifiers are never returned (say so if asked); orderStatus/grossProfit exist only on OMS-enabled accounts. Sort with sort_by (pathConversionId|timestamp|orderStatus|netRevenue|grossProfit) + sort_dir; page with limit (default 20, max 50) + offset — a `_truncated` sentinel flags more rows. Pass a row's pathConversionId to conversion_detail for that order's touchpoint journey. `attribution_model` takes a model **code**, never a display name. An `orderStatus` filter matches the page's own column filter server-side — use this tool, not query_performance, for order counts. Scope with `filters` like [{"name": "platformName", "values": ["Google Ads"]}] (optional "operator": in|notIn|contains|notContains; contains takes one "value"). Relative windows ('last 7 days', 'year to date') exclude today and end YESTERDAY; get exact app-matching dates from `resolve_date_range` instead of computing them. Omit dates/attribution/currency to use the caller's Explorer defaults; `defaults_applied` echoes what was used — disclose it. Pass `account_id` (from list_accounts) unless the server pins one; a pin always wins.

transaction_history