Skip to main content
Classify or score answers or pages using your own custom question and labels (e.g., pricing perception, sentiment, feature accuracy). Paid: $0.01/unit judged, runs within monthly budget; quote: true prices without buying. Select by ids or filter (up to 5,000 answers, excluding shared answers); answers >90d cost $0.001 each once. Pass up to 30 labeled spotChecks to measure agreement. Creates a private cf_ facet; check progress via read_item on the cf_ id.

Who may call it, and what it costs

Price. $0.01 a unit judged. Runs when called, within the budget; quote: true prices it only and starts nothing. Answers that target.filter selects from before the last 90 days are past answers, $0.001 each the first time: never a sample, the run buys every one the workspace does not own before it judges, or leaves the past out, and a quote shows their price on its own line. Every reply says what the call cost in meta.costCredits. The prices are on Spending, tracking and stopping.

Arguments

object
required
list of string
ans_ ids up to 5000
list of string
pg_ ids of stored pages up to 500
object
the answers search_answers finds for it in the window, at most 5,000; answers other workspaces share are left out
object
default the last 28 days; “last” or from/to
one of company, product
how brands count: company (default; a company and its products are one brand) or product (each brand on its own, so a product’s numbers leave out its siblings)
object
required
string
required
one question every unit is judged on
list of string
required
the answers it may have up to 10
list of object
units with their right value, shown to the judge up to 10
list of object
your labels; Parse reports how often the judge agrees up to 30
string
required
boolean
true: the price and whether it fits, buying nothing; without it the call runs when it fits
string
the same key with the same arguments returns the same job and charges nothing more; other arguments are refused. Left out, an identical call within the UTC hour is the same call; pass a key to buy it again
Every argument the tool takes, as GET /v1/tools serves it, every shared object written out; https://agent.parse.gl/openapi.json serves the same schema with the filter and the window named once under components and pointed to by $ref. MCP tools/list leaves out the arguments few calls send, and says the filter and the window once, under search_answers; resolve_ids {"kinds": ["argument"]} lists what it left out, and a call with any of them runs:

Example call

Price a judgment of a set of answers on your own question, and buy nothing. Over REST:
Over MCP, the assistant calls the tool by name with the same arguments:
The reply has data, context, next, page and meta; follow next for the calls that come after. A refused call returns { code, message, fix }: see the error table.