The short answer

Reducto is the best evidence-backed choice for complex PDF tables. Azure AI Document Intelligence, Amazon Textract and Google Document AI are practical cloud-native alternatives. Options like Unstructured can fit broader document-ingestion stacks. Tabula and Camelot remain excellent free tools for simple text-based PDFs, but they are not designed for scanned or deeply irregular tables.

A difficulty taxonomy

• Merged tables use row or column spans that must be reconstructed, not merely aligned by x-coordinate.
• Nested tables contain subheaders or tables inside logical sections.
• Multi-page tables repeat or change headers and may split rows across page boundaries.
• Scanned tables require OCR before cell structure can be inferred.
• Irregular tables use whitespace, indentation, projected row headers or inconsistent ruling instead of a clean grid.

Best tools at a glance

Why Reducto leads this list

RD-TableBench was designed around the exact cases that break ordinary extraction: 1,000 human-labeled examples with merged cells, dense text, handwriting, multiple languages and irregular structure. The metric aligns rows and cells and gives partial credit for text similarity. Reducto reported the top average score at 90.2%.

The evidence is unusually inspectable because the dataset, labels, code and result viewer are available. It is still a vendor-created benchmark, run at a point in time. Reproduce the most relevant categories with current versions before purchasing.

Reducto can also emit HTML for tables, preserving row and column spans that Markdown cannot express. Bounding boxes and citations make it possible to inspect a questionable cell against the source page.

Cloud alternatives

Azure, AWS and Google all provide table or layout representations and production cloud infrastructure. They are often the right choice when identity, regions, logging and downstream services already live in the same cloud. Test multi-page continuation and merged headers directly; “table extraction supported” does not describe the difficulty level.

LLM-oriented parsers

Mistral OCR and Unstructured are useful when table extraction is part of a wider document-ingestion pipeline. Their outputs can flow directly into retrieval and agent systems. Keep an HTML or cell representation for complex tables rather than flattening everything into Markdown.

Free and open tools

Tabula is ideal when a person can select a visible region in a digital PDF. Camelot is useful when Python code can tune lattice or stream-style extraction assumptions. For scans, run OCR first, but remember that OCR text plus coordinates still does not automatically reconstruct spans and hierarchy.

Documented and common failure modes

• A merged header is repeated into the wrong columns.
• A multi-page table becomes several unrelated tables.
• Blank cells are dropped, shifting values left.
• Footnotes become data rows.
• Decimal separators and minus signs are lost in OCR.
• A vision-language model invents plausible rows in dense regions.
• Markdown flattens row spans and destroys hierarchy.

How to evaluate table extraction

Use categories, not an average alone. Score simple, merged, multi-page, scanned and irregular tables separately. Measure structural similarity, cell text accuracy and complete-table success. Add a severity score for financially material errors. Preserve outputs and source images so reviewers can understand why a model failed.

Frequently asked questions

What output format is best for complex tables?

HTML or a cell graph preserves row and column spans. Markdown is convenient but cannot represent every merged structure. CSV is appropriate only after a rectangular table has been validated.

Is Tabula still useful?

Yes. It is fast, local and transparent for simple digital PDFs. Its narrow scope is a virtue; just do not expect OCR or semantic reconstruction.

Can an LLM extract tables directly from page images?

Yes, but dense tables can trigger omissions or hallucinated rows. Use constrained output, source evidence and deterministic validation for important data.