Core Processing

Core processing settings change the primary extraction result. They are top-level request fields on /extract, separate from extensions, which add derived outputs such as chunks or alternate formats.

{
  "additional_prompt": "Preserve account labels and signed amounts.",
  "refine": {
    "modes": ["tables", "text"],
    "prompt": "Keep source row and column order."
  },
  "detect_selections": true
}

Refinement is off unless you enable it with refine. It accepts a boolean, an array of modes, or an object with modes and an optional prompt:

// Boolean — general refinement pass
{ "refine": true }

// Array — only the listed modes
{ "refine": ["tables", "text"] }

// Object — modes plus guidance in one place
{ "refine": { "modes": ["tables", "layout"], "prompt": "..." } }

Mode	What it corrects
`tables`	Table structure, headers, and cell content
`text`	OCR text, missing content, and numerical accuracy
`formatting`	Bold, italic, strikethrough, super/subscript, and LaTeX formatting
`layout`	Layout correctness pass (fixed behavior)

Refinement changes markdown, bounding_boxes, and table content in place; there is no second response tree to merge.

refine.prompt — a string applies to the whole refinement pass; an object keyed by mode ({"tables": "...", "text": "..."}) scopes guidance per mode. Prompt keys must be a subset of the requested modes, and layout does not accept a custom prompt.
refine_prompt — a sibling alias for the same guidance, convenient when you pass refine as a boolean or array. refine.prompt wins when both are present.

Selection marks

detect_selections controls the specialized pass for checkboxes, radio buttons, handwritten checkmarks, X marks, and similar controls. It defaults to true. Detected states appear on the relevant bounding-box elements and in bounding_boxes.SelectionMarks when marks are available.

Prompts

additional_prompt gives the extraction engine document or domain context.
refine.prompt / refine_prompt apply only to refinement.
custom_image_prompt applies to figure and chart interpretation.

Prompts guide extraction; they do not change the response schema.

Processing versus extensions

Use	Location	Result
Correct or steer the primary extraction	`refine`, `refine_prompt`, `additional_prompt`, `detect_selections`	Updates core markdown/layout output
Describe or return visuals	`figure_processing.*`, `custom_image_prompt`	Updates markdown and `bounding_boxes.Images[]`
Add chunks, metadata, footnote links, or alternate output	`extensions.*`	Adds fields under `response.extensions`

Extract API

Complete request and response contract.

Bounding Boxes

Table cells, words, categories, and coordinates.

Start Here

Agent Tools

Core Concepts

Security & Compliance

Help

Refinement

Refinement prompts

Selection marks

Prompts

Processing versus extensions

Extract API

Bounding Boxes

​Refinement

​Refinement prompts

​Selection marks

​Prompts

​Processing versus extensions

Extract API

Bounding Boxes

Refinement

Refinement prompts

Selection marks

Prompts

Processing versus extensions