Pulse Documentation

name: tool-process
kind: tool
description: Execute one pre-built processing request against a cohort.
type: reference
applies_to: process, compose, predict, mcp

When to use

CALL TO RUN ONE VALIDATED REQUEST AGAINST A COHORT. The body comes from pulse_examples_get (or is hand-authored and run through pulse_predict). For multi-question batches use pulse_compose. For stage-chained workflows use pulse_process_chain.

Input

request (string): JSON-encoded types.Request. Top-level keys: cohort, aggregations, groups, attributes, filterers, windows, features, regressions, crosstab, joins, overlays, post_tests.

Output

descriptor.Envelope wrapping pulse.Response. data carries Data (rows or matrix), Metadata (cohort filename, partial-cohort reasons), and Components (universal floor + per-operator state — see response-components). errors / warnings use coded entries.

Gotchas

• Slot keys differ from manifest catalog names: groupings go under "groups" (NOT "groupers"); aggregations under "aggregations" (NOT "aggregators"). Unknown top-level keys → PULSE_REQUEST_UNKNOWN_FIELD with a “did you mean” suggestion.
• DO NOT author bodies from external docs or source code — they may be out of date. Always seed from pulse_examples_search + pulse_examples_get or pulse_manifest.
• Smart defaults: omitting Type on a slot that names a field auto-infers from schema type. Disable via Options.DisableDefaults or CLI --no-defaults.
• models: [{name, path}] loads saved model files by name for scoring references; see compose-fits (Loading).

See

• request-envelope — slot map, smart defaults, normalization.
• response-components — universal floor + per-operator components emitted in Response.Components.
• aggregation-design, grouper-design, attribute-composition — operator catalogs.