Use Custom AI Models via Callbacks

Overview

Starting with SDK v4.1.0, ComPDF Conversion SDK exposes a callback-based extension point that lets you plug in your own AI inference engine for OCR, Layout Analysis, and Table Recognition. Instead of relying on the built-in DocumentAI model loaded by set_document_ai_model, you can:

• Run inference with any model or runtime you choose (e.g. your in-house engine, PaddleOCR, a cloud OCR API).
• Return the result to the SDK as a JSON string with a well-defined schema.

When the relevant callback pair is registered on ConvertCallback, the SDK skips its built-in model invocation for that capability and consumes your JSON output instead. If a pair is left unset, the SDK falls back to the built-in DocumentAI model (when available).

Callback Pairs

Each AI capability uses two callbacks: a trigger (invoked by the SDK with the path to a page image saved as PNG in a temp directory) and a result getter (invoked by the SDK immediately afterwards to retrieve the JSON string).

Capability	Trigger field / setter	Result getter field / setter	Triggered when
OCR	ocr	get_ocr_result	OCR is enabled
Layout Analysis	layout	get_layout_result	layout analysis is enabled (or implicitly when OCR is enabled)
Table Recognition	table	get_table_result	table recognition is enabled and a table region is detected by layout analysis

Rules:

• The trigger receives a UTF-8 path to a PNG file. Return true if your inference succeeded, false to make the SDK ignore the result for that page.
• The getter must return a UTF-8 JSON string. The SDK copies the string into an internal buffer before consuming it.
• Both callbacks for a capability must be set together. If only one is provided, the SDK falls back to the built-in path.
• Coordinates in your JSON must be in the pixel space of the image the trigger received (top-left origin, X right, Y down).
• Confidence filtering: OCR spans with confidence < 0.1 and layout objects with confidence < 0.45 are discarded by the SDK.
• When all three capabilities you need are covered by your own callbacks, set_document_ai_model does not have to be called.

Sample

def on_ocr(image_path: str) -> bool:
    # Run your OCR engine on `image_path`, cache the JSON result.
    return True

def get_ocr_result() -> str:
    return ""  # return UTF-8 JSON

def on_layout(image_path: str) -> bool:
    # Run your layout engine on `image_path`.
    return True

def get_layout_result() -> str:
    return ""  # return UTF-8 JSON

def on_table(image_path: str) -> bool:
    # Run your table engine on `image_path`.
    return True

def get_table_result() -> str:
    return ""  # return UTF-8 JSON

callback = ConvertCallback()
callback.set_ocr(on_ocr)
callback.set_get_ocr_result(get_ocr_result)
callback.set_layout(on_layout)
callback.set_get_layout_result(get_layout_result)
callback.set_table(on_table)
callback.set_get_table_result(get_table_result)

word_options = ConvertOptions()
word_options.enable_ocr = True
word_options.enable_ai_layout = True
word_options.enable_ai_table_recognition = True
CPDFConversion.start_pdf_to_word("input.pdf", "password", "path/output.docx", word_options, callback)

JSON Schemas

OCR Result JSON Schema

Returned by get_ocr_result. The SDK populates each text_spans[].chars[] either from words[] if provided, or by uniformly splitting the span rect.

{
  "text_spans": [
    {
      "text": "Hello World",
      "confidence": 0.98,
      "rotation": 0.0,
      "rect": { "left": 120, "top": 80, "right": 320, "bottom": 110 },
      "style": {
        "font_size": 18.0,
        "font_color": { "r": 0, "g": 0, "b": 0 }
      },
      "words": [
        { "text": "Hello", "rect": { "left": 120, "top": 80, "right": 200, "bottom": 110 } },
        { "text": "World", "rect": { "left": 210, "top": 80, "right": 320, "bottom": 110 } }
      ]
    }
  ]
}

Field	Type	Required	Description
text_spans	array	Yes	Recognized text spans on the page.
text	string	Yes	UTF-8 text content of the span.
confidence	number	No	0.0 – 1.0. Spans below 0.1 are discarded.
rotation	number	No	Text rotation in degrees. Default 0.
rect	object	Yes	Bounding box in image pixels (left/top/right/bottom).
style.font_size	number	No	Estimated font size in pixels.
style.font_color	object	No	{ r, g, b } 0 – 255.
words	array	No	Per-word boxes. If omitted, the SDK splits the span rect evenly. Strongly recommended for CJK + Latin mixed lines for correct glyph spacing.

Layout Analysis Result JSON Schema

Returned by get_layout_result. Objects with confidence < 0.45 are discarded.

{
  "objects": [
    { "type": "title", "confidence": 0.95, "rect": { "left": 60, "top": 50, "right": 540, "bottom": 90 } },
    { "type": "paragraph", "confidence": 0.97, "rect": { "left": 60, "top": 100, "right": 540, "bottom": 220 } },
    { "type": "figure", "confidence": 0.92, "rect": { "left": 80, "top": 240, "right": 520, "bottom": 460 } },
    { "type": "table", "confidence": 0.93, "rect": { "left": 60, "top": 480, "right": 540, "bottom": 700 } }
  ]
}

Supported type values:

Value	Meaning
paragraph	Body text paragraph
title	Heading
figure	Image or figure
figure_title	Figure caption header
figure_caption	Figure caption text
table	Table region. Whether the table is bordered or borderless is determined by the table recognition stage, not by the layout label.
table_title	Table caption header
table_caption	Table caption text
ordered_list	Ordered list
unordered_list	Unordered list
catalogue	Table of contents
formula	Math formula
code	Code block
algorithm	Algorithm block
header	Page header
footer	Page footer
page_number	Page number
reference	Reference or citation

Objects with a type value that is not listed above are ignored. Use the values in this table as the canonical layout labels in your custom output.

Table Recognition Result JSON Schema

Returned by get_table_result once per detected table region. Polygons use 8 integers [x0, y0, x1, y1, x2, y2, x3, y3] in the order top-left, top-right, bottom-right, bottom-left.

{
  "type": "table_with_line",
  "position": [60, 480, 540, 480, 540, 700, 60, 700],
  "rows": 3,
  "cols": 2,
  "angle": 0.0,
  "height_of_rows": [40, 60, 60],
  "width_of_cols": [200, 280],
  "table_cells": [
    {
      "start_row": 0,
      "end_row": 0,
      "start_col": 0,
      "end_col": 0,
      "cell_background_color_r": 240,
      "cell_background_color_g": 240,
      "cell_background_color_b": 240,
      "position": [60, 480, 260, 480, 260, 520, 60, 520]
    }
  ]
}

Field	Type	Description
type	string	table_with_line for bordered tables; any other value is treated as a non-standard (borderless) table.
position	int[8]	Table polygon in image pixels.
rows / cols	int	Row / column counts.
angle	number	Skew angle in degrees.
height_of_rows	int[]	Per-row pixel heights (length = rows).
width_of_cols	int[]	Per-column pixel widths (length = cols).
table_cells[]	array	One entry per merged cell.
start_row / end_row	int	Inclusive row span of the cell.
start_col / end_col	int	Inclusive column span of the cell.
cell_background_color_*	int	Cell background color components (0 – 255).
position	int[8]	Cell polygon in image pixels.

Tip: Validate Your JSON

If you need a reference output to compare against, run a conversion once with the built-in DocumentAI model. The SDK uses the same JSON shape internally, so your custom output should follow the same structure.