Ask over collected records
ask_collected_recordsAnswer from collected records, with quotes tied to exact lines.
Extracts and synthesises across material too large to fit in context, and returns every claim traceable to a line range in the original. Use instead of summarising from memory whenever the evidence has to hold up — quotes come back with their locations, so they can be checked against the source.
Not for: exact wording or counting occurrences — grep_collected_records does that without an LLM call; or material small enough to read directly — use read_collected_records.
In quotes and both modes the output is quotes: [{ id: '{<uuid>}', index }] — one block per record, with index formatted like the index field of advanced_web_search / advanced_web_fetch results. Empty array when no quotable evidence is found. When part of the material could not be read or worked into the result, notes says what is missing. If status: "running" is returned, poll get_task_results({ taskIds }).
Tool call template
Replace each placeholder with a value matching the input schema. Only required parameters are included; add optional parameters as needed.
{
"id": 1,
"jsonrpc": "2.0",
"method": "tools/call",
"params": {
"arguments": {
"dataIds": "<dataIds>",
"query": "<query>",
"mode": "<mode>"
},
"name": "ask_collected_records"
}
}Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| mode | Yes | `answer` for narrative prose, `quotes` for a per-record line-range index, `both` for both. | |
| query | Yes | The question to answer over the selected records. | |
| dataIds | Yes | One or more Data records to answer the question over. Each item is `{ id, version?, from?, to? }` — pass `{from, to}` to scope the citation to a specific line range; `version` is optional and defaults to latest. URLs are not accepted; collect them via `advanced_web_search` or `advanced_web_fetch` first. | |
| maxWaitSec | No | How long to wait before returning `status: "running"` with a `taskId` to poll via `get_task_results`. Defaults to 50 seconds, which is also the server's ceiling — a larger value is clamped down to it, not honoured. |