ask_cosmos_kappa_transact
ChatGPTQuery Consumer Edge Kappa Transact data using natural language. Answers questions about Consumer Edge Kappa Transact Panel. Returns a task_id — poll get_cosmos_result(task_id='<returned task_id>') every ~10s until ready. Question formatting rules: - Preserve exact time ranges, date constraints, and metrics from the user. Do NOT broaden scope (e.g. 'last 4 weeks' stays 'last 4 weeks'). - Preserve stock tickers verbatim (e.g. keep 'LULU', don't expand to 'Lululemon'). Cosmos resolves tickers itself. - Do NOT add analytical details the user didn't specify (comparison windows, aggregations, bucketing). Cosmos picks the right approach. Sessions: pass a session_id (from a previous get_cosmos_result) to continue a conversation. Only one query per session at a time — wait for it to finish before reusing the id. Omit session_id to start fresh. For follow-ups, send only the current request and brief facts needed to resolve references; never send a conversation transcript. Pass that request in original_user_question so Cosmos can interpret the question and show the request as User ask in the submission response.
ask_cosmos_orion
ChatGPTQuery Consumer Edge Orion data using natural language. Answers questions about Consumer Edge Orion 2022 Panel. Returns a task_id — poll get_cosmos_result(task_id='<returned task_id>') every ~10s until ready. Question formatting rules: - Preserve exact time ranges, date constraints, and metrics from the user. Do NOT broaden scope (e.g. 'last 4 weeks' stays 'last 4 weeks'). - Preserve stock tickers verbatim (e.g. keep 'LULU', don't expand to 'Lululemon'). Cosmos resolves tickers itself. - Orion merchants are a two-level hierarchy: a parent merchant is the owning company (e.g. 'The Gap', 'Yum! Brands, Inc.'); a child merchant is a banner under it (e.g. 'The Gap: Old Navy', 'Taco Bell'). The same name can exist at either level and the two give different numbers, so you MUST disambiguate. Explicitly say 'parent merchant X' or 'child merchant X' in the question; if the user is ambiguous, ask before calling this tool. - One child merchant can roll up to more than one parent (e.g. 'Kroger' appears under both 'Kroger' and 'Kroger Ex Fuel'). If a child-merchant query returns multiple parent rows, don't sum them — pick the parent matching the user's intent or surface both distinctly, and explain the split. - Do NOT add analytical details the user didn't specify (comparison windows, aggregations, bucketing). Cosmos picks the right approach. Sessions: pass a session_id (from a previous get_cosmos_result) to continue a conversation. Only one query per session at a time — wait for it to finish before reusing the id. Omit session_id to start fresh. For follow-ups, send only the current request and brief facts needed to resolve references; never send a conversation transcript. Pass that request in original_user_question so Cosmos can interpret the question and show the request as User ask in the submission response.
get_cosmos_result
ChatGPTPoll for the result of a previously submitted Cosmos query. Call this after submitting a question with an ask_cosmos_<dataset> tool. Automatically waits up to 40 seconds for completion before responding, so you only need to poll every ~10 seconds. Returns the query result on success, a status update if still running, or an error message on failure.
get_cosmos_usage
ChatGPTReport your organization's Cosmos usage against its limits. Returns, for each limited resource, the share of its limit drawn, how many units are left, and (when relevant) a suggested action if you are close to or over it.
query_cosmos_data
ChatGPTRun a DuckDB SQL query against tabular data from a completed Cosmos query. Available after get_cosmos_result returns data tables. The result includes an "Available Data" section listing table names like data_0, data_1, etc. that this tool queries with standard SQL. Strongly preferred for any calculation, aggregation, or comparison on Cosmos results — averages, totals, growth rates, ranking. It computes exact values over the full returned rows, which is more reliable than estimating from the table preview.
refresh_cosmos_task
ChatGPTRefresh an earlier Cosmos answer by running its query again, without asking the question again. Use this whenever figures from an earlier answer are needed again, such as a dashboard tile or a recurring report: asking again can write a different query and return different numbers. Pass the task_id of the answer, or of an earlier refresh of it. - No as_of_date or dates: re-runs the task's own date windows. - as_of_date "latest" or a date (YYYY-MM-DD): moves the windows to that data date the way they were meant, e.g. week to date stays week to date and the same week last year moves with it. - dates: sets the named dates to exact values and keeps the rest. Returns a new task_id right away; poll get_cosmos_result with it. A refreshed task can be queried with query_cosmos_data and refreshed in turn, so keep its task_id. A refresh that runs uses one credit, like a question. One that is refused, fails, or ends in needs_choice uses none. Each answer has an "Outcome:" line: - submitted: poll get_cosmos_result. - refreshed: the refreshed results, listing the dates each one used. - needs_choice: the windows can move more than one way, so nothing ran. Ask the user which is meant, then refresh that choice's task_id with the as_of_date given. - refused (<reason>): nothing ran. not_supported_yet: that kind of result can't be refreshed yet, so ask again. latest_only: the result can only be refreshed with as_of_date "latest" or with dates. baseline_invalid: the answer's query doesn't match its question, so ask again. no_dates_given: lists the task's date ids. Other reasons: not_found, wrong_variant, no_access, not_refreshable, source_unavailable, needs_date_rules, windows_not_movable, unknown_date_ids, invalid_dates, rate_limited, out_of_credits. - failed (<reason>): the refresh stopped while running: validation_failed, execution_failed, sql_row_cap_reached, error, or a refused reason found while running, such as baseline_invalid. The text after the line says what happened and what to do.
search_cosmos_tasks
ChatGPTFind your earlier Cosmos answers by the words of the request. Use this when you need an earlier answer but no longer have its task_id, for example in a new conversation: refresh it with refresh_cosmos_task instead of asking the question again, or show it again with get_cosmos_result. Returns up to 10 of your answers from this server, best match first. Each gives the task_id to use, which for an answer refreshed before is its latest refresh, with the question, its dataset, when it was asked and last run, the data date it read, and whether it can be refreshed and, if not, why.
ask_cosmos_kappa_transact
Claudeask_cosmos_orion
Claudeget_cosmos_result
Claudeget_cosmos_usage
Claudequery_cosmos_data
Claude