Status: Runnable

Extraction output

An extract step reads one media slot and returns JSON whose shape depends on the file’s format. The same shape reaches assertions, Agent and code inputs, and every SDK’s result types.

steps:
  - id: document
    action: { kind: extract, from: inputs.policy }

Every field is always present. A value that is unavailable is null only where the type below allows it; an empty list is [] and empty text is "". Counts come from list lengths: there is no page_count or slide_count.

Formats

The TypeScript notation describes the JSON on the wire.

PDF

type PdfOutput = { format: "pdf"; pages: { number: number; text: string }[] };

{"format":"pdf","pages":[{"number":1,"text":"Refund policy\nRefunds are issued within 30 days."},{"number":2,"text":""}]}

number is one-based.

PPTX

type PptxOutput = {
  format: "pptx";
  slides: {
    number: number;
    title: string | null;
    paragraphs: string[];
    tables: { rows: string[][] }[];
    notes: string;
  }[];
};

{"format":"pptx","slides":[{"number":1,"title":"Q3 refunds","paragraphs":["Refund rate fell to 2%"],"tables":[{"rows":[["Region","Rate"],["US","2%"]]}],"notes":""}]}

Slides are in presentation order; number is one-based.

DOCX

type DocxOutput = {
  format: "docx";
  blocks: ({ kind: "paragraph"; text: string; style: string | null } | { kind: "table"; rows: string[][] })[];
};

{"format":"docx","blocks":[{"kind":"paragraph","text":"Refund policy","style":"Heading1"},{"kind":"table","rows":[["Window","Days"],["Standard","30"]]}]}

Blocks are in document order. style is the stored paragraph style ID.

XLSX

type XlsxOutput = {
  format: "xlsx";
  sheets: {
    name: string;
    cells: { address: string; value: string | number | boolean | null; formula: string | null; error: string | null }[];
  }[];
};

{"format":"xlsx","sheets":[{"name":"Refunds","cells":[{"address":"A1","value":"Days","formula":null,"error":null},{"address":"B2","value":30,"formula":null,"error":null}]}]}

• Sheets keep workbook order; cells keep row, then column order.
• address is the worksheet address, such as B2, not a position in the list. Empty cells are omitted, so cells[1] above is B2.
• value is the stored value or the cached result of a formula. Formulas are never recalculated.
• Dates stay Excel serial numbers.
• An error cell has value: null and the error, such as #DIV/0!, in error.

CSV

type CsvOutput = { format: "csv"; rows: string[][] };

{"format":"csv","rows":[["region","rate"],["US","0.02"]]}

Every row is kept, including the first. Neither headers nor value types are inferred: every cell is a string.

JSON

type JsonOutput = { format: "json"; value: JsonValue };

{"format":"json","value":{"refund_days":30,"regions":["US","EU"]}}

value is the parsed document.

Text, Markdown, HTML

type TextOutput = { format: "text" | "markdown" | "html"; text: string };

{"format":"markdown","text":"# Refund policy\n\nRefunds are issued within 30 days.\n"}

The source text, unchanged. Markdown and HTML are not parsed into structure.

Tables

In PPTX, DOCX, and similar tables, each cell holds its extracted text. A merged cell’s value is in its top-left position; the positions it covers are "", so every row of a table has the same length.

Addressing the output

References use dotted paths with zero-based indexes and no $. prefix:

requirements:
  - id: titled
    statement: The policy opens with its title
    check: { kind: assert, path: steps.document.output.blocks[0].text, operator: equals, expected: Refund policy }
  - id: thirty-days
    statement: The standard refund window is 30 days
    check: { kind: assert, path: steps.workbook.output.sheets[0].cells[1].value, operator: equals, expected: 30 }
  - id: no-errors
    statement: No cell in the workbook is an error
    check: { kind: assert, each: "steps.workbook.output.sheets[*].cells[*].error", operator: equals, expected: null }

[*] is allowed only in each, which passes when every element passes. List indexes are zero-based; PDF page and PPTX slide number values are one-based.

Citations into extracted files

A citation names the original slot, never the extract step:

For DOCX and XLSX, the JSON pointer addresses this page’s extracted JSON for that slot, such as /blocks/0/text or /sheets/0/cells/1/value. It is not a byte offset into the file. The server checks every citation against the pinned file and the extraction version that produced it; a pointer or page that does not exist makes a passing Agent check inconclusive.

Limits

• Extraction is text and structure only. There is no OCR, no rendering, and no extracted image files.
• To check how something looks, give an Agent the original file as a media input; it reaches a model that supports that format natively. A caller supplies a PDF or image rendition of an Office file when a visual check needs one.
• ppt, doc, and xls are distinct from pptx, docx, and xlsx and are not extracted.
• Pages, slides, and extracted text are bounded by WYRD_VERIFICATION_EXTRACT_MAX_PAGES and WYRD_VERIFICATION_EXTRACT_TEXT_MAX_MIB; see Verification runtime limits. A file over a limit fails the step, and requirements that need it are inconclusive.

Last updated October 16, 2026