Niches · Document parsing & extraction
Get a table out with its rows and columns intact
Four or more companies' docs show a full fit.
What counts as done
Each table as HTML or JSON cells with row and column spans, every cell's text present, subheadings kept inside the table, the page it came from, and no table read as a form or paragraph.
Measured, for our own tests: cell-level accuracy (TEDS or similar) on a fixed set of merged-cell and multi-page tables. No docs we read state these figures, so they don't count against a listing here.
How agents ask for it: extract the tables from this PDF
, get this table into CSV
, pull the figures from the financial statements
, read the dosing table
Does the job 5
The docs show it meets the test above, and nothing we read says otherwise.
| Listing | Why |
|---|---|
| Adobe PDF Services / PDF Extract API C Adobe Inc. | Extract returns Table, TR, TH and TD elements with RowSpan, ColSpan, Page and Bounds, plus CSV and XLSX renditions per table; an over-complex table fails with BAD_PDF_COMPLEX_TABLE rather than loose text. source |
| Azure Document Intelligence B Microsoft Corporation | Layout returns cells with rowIndex, columnIndex, rowSpan, columnSpan, columnHeader kind and boundingRegions with page number, and HTML tables in Markdown; acc-10 uses it as the primary table engine. source |
| LandingAI Agentic Document Extraction BB LandingAI, Inc. | Tables have table_cell children in row-major order with row, column, rowspan and colspan, each with its own page and box, an empty cell shows as a zero-length range, and HTML output keeps merged cells. source |
| Nemotron 3.5 Parse B NVIDIA Corporation | Table blocks render as HTML preserving merged cells, or CSV, JSON or hierarchical JSON, each with page box and class, from a page-level parse; NVIDIA reports ParseBench results, its own run. source |
| Reducto API + MCP B Reducto, Inc. | HTML table output keeps merged cells, and with return_citations each table block lists every cell with row and column indices and its box, merged headers included, on the block's page. source |
Does the job, with a limit 11
It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.
| Listing | Why |
|---|---|
| Amazon Textract BB Amazon Web Services, Inc. | TABLE, CELL and MERGED_CELL blocks carry RowSpan, ColumnSpan, TABLE_TITLE, TABLE_FOOTER, page and boxes, matching the criterion on paper, but acc-02 and acc-08 report weak results on complex tables. source |
| Datalab B Endless Labs, Inc. | HTML tables with table_cell_bboxes (box and confidence per cell) and merge_cross_page for tables across pages, but acc-30 saw Marker 2.0 handle tables worse than before, and output can change with no pin below enterprise. source |
| Docling B Docling Project (LF AI & Data Foundation) | TableCell has bbox, row_span, col_span, column_header and row_section, but accounts show tables classed as forms (acc-24), empty cells after an update (acc-26), lost bands (acc-27) and misplaced header cells (acc-28). source |
| Extend API + MCP B CrowdView Inc. dba Extend | Table blocks come with page, bounding box and Markdown or HTML content, but the response format read does not document row or column spans or per-cell output. source |
| Google Cloud Document AI BB Google LLC | Table cells carry rowSpan, colSpan and a layout with page anchor, but the Layout Parser docs warn that tables spanning pages might be split into two tables. source |
| LlamaParse API + MCP B LlamaIndex, Inc. | HTML tables, merge_continued_tables, XLSX export and cell bounding boxes are options, but acc-34 found subheadings inside tables left out of the table node. source |
| Mistral OCR API B Mistral AI (RCS Paris 952 418 325) | table_format=html returns tables separately as HTML, but acc-39 saw unescaped pipes break Markdown table columns, and no per-cell boxes are returned. source |
| Nanonets API + MCP E Nano Net Technologies Inc. | Extract can output CSV tables with per-cell confidence, and parse can output HTML, but row and column spans and the page of each table are not in the schema. source |
| OpenDocRouter B LlamaIndex, Inc. | Tables arrive inside the Markdown, with a layout table element and its boxes, but no cell structure, spans or HTML option is documented. source |
| Unstructured API + MCP E Unstructured Technologies, Inc. | Tables come back as text_as_html with page_number and coordinates when requested, but acc-36 saw exponents in table formulas parsed wrongly, and spans are not described. source |
| Xiaomi-OCR-0 C Xiaomi Corporation | Tables come out as OTSL, which keeps spans but must be converted to HTML with the repository's script, and the 92.27 PubTabNet figure is the vendor's. source |
Doesn't do this job 4
Listings in the category that do something else, with the reason.
| Listing | Why |
|---|---|
| Clairic BB Cloudflare, Inc. | No table output type; field types are text, number, date, time, datetime, duration, list, estimate and forecast. source |
| Invofox D Invofox, Inc. (Delaware), with a registered branch in Spain | No table output in the document schema; values follow the configured document type's fields. source |
| Mindee API C Mindee SAS (Mindee Inc. outside the Eurozone) | Models return schema fields and lists, not tables with rows, columns and spans. source |
| Veryfi API + MCP C Veryfi, Inc. | Line items are returned as document fields, not general tables with rows, columns and spans. source |
6 listings in the category were added after this reading and aren't marked yet.
What stops agents finishing
- table flattened to Markdown that looks right but misaligns figures
- update regresses table output with no pinned version to fall back to
- table classified as a form or image and returned as loose text
Around it
Before: document-extraction.rag-ingest, document-extraction.scan-ocr. After: data.local-files, accounting.read-reports.
Where this comes from
Weight 21 of 100 in Document parsing & extraction: 13 of 46 accounts; the largest group of GitHub issues (Docling, Marker, LlamaParse, Mistral, Unstructured) is tables losing cells, rows or columns
The same as data: /api/v1/use-cases/document-extraction.table-extract.json, or the MCP tool get_use_case with id document-extraction.table-extract. Other jobs in this category: /api/v1/niches?category=document-extraction.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.