Today we’re releasing SuperNet-201, our new OCR model, for
v3/text, v3/pdf and the Files API. It recognizes content more accurately, fixes cases where text was previously missing, and adds confidence metrics to structured output: separate text and layout confidence scores for every region, and a confidence summary for every page.Better recognition
The recognition improvements apply automatically, with no changes to your requests. We’ve improved parsing of newspapers, multi-column pages and other complex layouts where text and images are interleaved and reading order is ambiguous. Reading order is also improved on rotated pages. The model handles vertical Chinese, Korean and Japanese text more accurately, including columns read from right to left.
The model recognizes marginalia and footnotes more accurately, improves table structure recognition, including wide tables, and better recognizes receipts and their structure. Filled-in form answers are extracted as text. The existing
include_page_info: ["margin_note"] option includes content recognized as margin notes in converted output while excluding other page information, such as headers and page numbers.Text and layout confidence for each region and page
Structured output now includes separate
text_confidence and layout_confidence fields for each region, alongside a combined confidence. Text confidence estimates how reliably the text was recognized; layout confidence estimates how reliably its structure was parsed. A page_confidence object summarizes each page.These scores can help you choose which results to review. For example, a table’s text may be recognized confidently even when its row and column structure is uncertain. The table container has layout confidence; its child text lines carry the text confidence.
Getting the scores
For an image of a page, call
v3/text with document layout enabled:{
"src": "https://example.com/page.png",
"enable_document_layout": true
}
This includes
lines_json by default. You can also request it in ordinary image mode with include_lines_json: true, without changing the recognition mode. The response contains a top-level page_confidence, also available at lines_json.page_confidence, and confidence fields on each region. Here is an excerpt from a live response. Other fields and regions are omitted:{
"page_confidence": {
"confidence": 0.6765460968017578,
"text_confidence": 0.6765460968017578,
"layout_confidence": 0.786479115486145
},
"lines_json": {
"page_confidence": {
"confidence": 0.6765460968017578,
"text_confidence": 0.6765460968017578,
"layout_confidence": 0.786479115486145
},
"lines": [
{
"type": "text",
"text": "Cache",
"confidence": 0.9093252420425415,
"text_confidence": 1,
"layout_confidence": 0.9093252420425415
},
{
"type": "table",
"confidence": 0.786479115486145,
"text_confidence": null,
"layout_confidence": 0.786479115486145
}
]
}
}
For
v3/pdf, download GET /v3/pdf/{pdf_id}.lines.json after processing completes. For the Files API, download GET /files/v1/{pdf_id}.lines.json. Each entry in pages includes its own page_confidence; it describes that page, not the entire document.If you only need the recognized text and page summary from
v3/text, set include_lines_json: false to omit the larger region output.Reading the scores
Combined region confidence takes the lower of the required text and layout scores. Non-text regions use layout confidence alone. Unknown or inapplicable values are
null; an empty page with no evidence also reports null confidence values.The page summary accounts for both text recognition and page structure. One uncertain region can lower the page score, but page layout confidence also uses information from parsing the page that is not represented by individual text lines.
The scores are estimates, not probabilities that a result is correct. A score of
0.79 does not mean there is a 79% chance of an error-free page. They can miss errors, including omitted content, repeated lines or incorrect reading order, so evaluate review thresholds on your own documents.Existing confidence values in image
line_data, word data, and the top-level image OCR confidence and confidence_rate keep their meanings. include_line_data remains independent of include_lines_json. The image OCR reference and PDF reference describe the fields and calculations in detail.Preserving tables
When the parser cannot produce a valid table structure, document processing attempts to preserve the table as an image while keeping the rest of the page as text. This happens automatically in
v3/pdf and the Files API. The image preserves the table’s appearance, but its cells are no longer available as extracted text or an editable table. This handles detected parsing failures; it does not catch every incorrectly recognized table.A table preserved this way has
parse_status: "image_fallback". Its fallback_reason explains why the structure was rejected, for example invalid_structure, and text_display contains the image reference. Combined and layout confidence are zero for the table and page because structured extraction is incomplete. Other regions keep their own confidence scores.If the image cannot be prepared or its upload fails, the table is marked
failed with image_fallback_unavailable. v3/text does not return hosted crop URLs.See the changelog for release details. Please reach out to support@mathpix.com if you have any questions.