Zum Hauptinhalt springen

Create Config

POST 
/configs

Create a new config in the API key's project.

The name must be unique within the project. kind selects the config type and the shape of parameters, which is validated against that kind. For invoice-v2: the inline extraction_config.schema must be a well-formed JSON object schema, and every matcher's dataset_id, column and resolved scalar field path must exist. For extraction: record_schema_id must name a record schema in the same project, and every matcher's dataset_id, column and resolved scalar field path must exist — scope and field resolve against that referenced record schema, not an inline one. For redaction: label_set_id must name a redaction label set of the project (null picks the project's only one; several is a 422), and every label_ids entry must be an entity of that set (null selects all). A project holds one redaction config, the selection a redaction job uses without label_ids; creating a second is a 409, change the existing one with PUT instead.

An invoice-v2 config's extraction_config carries two opt-in switches, both false by default, that widen what a job created with it returns:

  • include_metadata — the result's metadata.extraction mirrors the structure of data with per-field extraction metadata (confidence etc.). It costs extra tokens, so it is null without the switch.
  • read_qr_codes — the QR codes on the document's page images are decoded and their raw payloads returned in the result's qr_codes (a Swiss QR-bill, an EPC Girocode, any other QR content — verbatim, for the client to interpret). Report-only: they never steer extraction. With the switch off qr_codes is null, which says the pages were not read — not that the document carries no code.

Mapping roles (matchers[].mappings[].role, optional — defaults to auto):

  • identifier — must-match: a row qualifies only if the extracted value equals the column exactly (raw comparison, no normalization). Multiple identifiers AND together; an identifier whose extracted value is empty constrains nothing.
  • descriptive — scored text signal; abbreviations, word order and language differences are expected.
  • numeric — scored number signal; small rounding differences are tolerated.
  • auto — scored signal weighed at the matching engine's discretion; never a hard constraint.

A row counts as a match when its score reaches the matcher's min_score; rows are ordered by score and capped at limit.

Request​

Responses​

Config created successfully