How Do I Extract Tables from a PDF for LLMs?

The answer in one sentence

To extract tables from a PDF for an LLM, preserve the relationships that make the table meaningful: column headers, row labels, merged cells, captions, footnotes, and page location. A flat stream of numbers is not a table. It is a pile of values the model has to guess how to reconnect.

Use DocuShell Parse PDF to produce structured JSON for machines and Markdown for model context. Keep both versions. JSON gives your application a predictable record; Markdown gives the model a readable table.

Table-extraction rule: if a value cannot be traced to a row label, column label, unit, period, and page, it is not ready for an LLM or an automated decision.

Choose the extraction strategy by document type

Why copy and paste fails

PDFs store content by position. A value can sit under a heading visually without being linked to it in the file. Basic text extraction may read across columns, repeat headers, or move footnotes into the middle of a row.

That damage is easy to miss in a short prompt. It becomes costly when an LLM compares quarterly revenue, checks an eligibility matrix, or answers from a pricing table.

A reliable extraction workflow

Start with layout-aware parsing. For every table, retain:

• the section heading and table caption
• column and row labels
• page number and source coordinates
• units, currency, and date range
• footnotes and continuation markers

Then store two representations. Use structured JSON for filters, calculations, and QA. Use a Markdown table for retrieval and generation. Link both with the same table ID and source metadata.

Chunk the table as one idea

Do not split a table every few hundred tokens without considering its shape. A chunk containing only the last four rows may omit the header that explains every value.

Keep a small table intact. For a long table, repeat the column headers in every chunk and attach the same caption, section path, and page range. Add nearby prose when it changes how the figures should be read.

Handle scanned tables carefully

A scanned PDF needs OCR before table reconstruction. OCR can confuse 0 and O, drop decimal points, or merge adjacent cells. That is acceptable for search discovery, but not automatically safe for financial calculations or compliance decisions.

Use the OCR accuracy guide to improve the source, then review high-impact cells against the original page.

Make answers auditable

Require the LLM to cite the table title and page. For important outputs, validate generated numbers against the structured JSON rather than trusting the prose response alone.

The goal is simple: a reviewer should be able to move from an answer to the exact table, row, and source page without reconstructing the pipeline.

Example: preparing a quarterly revenue table

Suppose a report contains revenue by region for 2025 and 2026. The visible page also says “USD in millions,” and a footnote explains that one region was reclassified. A weak extractor sends the model twelve numbers. A useful pipeline creates one table record containing the title, unit, periods, row labels, values, footnote, page, and source coordinates.

The LLM can then answer “Which region grew fastest?” by retrieving the table, while application code calculates the percentage change. The response cites the page and mentions the reclassification instead of presenting a misleading comparison.

This division of labor matters. Models are good at explaining evidence. Deterministic code is better for arithmetic and schema checks.

A compact table record can look like this:

{
  "table_id": "revenue-by-region-p18",
  "title": "Revenue by Region",
  "unit": "USD millions",
  "periods": ["2025", "2026"],
  "rows": [
    { "region": "EMEA", "values": [31.4, 33.9] }
  ],
  "source": { "page": 18, "bbox": [72, 216, 540, 376] },
  "notes": ["Prior-year regions were reclassified"]
}

Table-extraction quality checklist

Before indexing the output, check that:

• every value still has a row and column label
• units and reporting periods are explicit
• negative numbers and parentheses are preserved
• repeated headers are not stored as data rows
• multi-page tables have been joined only when their schemas match
• captions and footnotes remain attached
• page citations open the correct source region

For a broader ingestion design, read PDF Parse API for LLM and RAG. If the source is image-only, start with how to parse scanned PDFs into Markdown and JSON. When the table will feed a retrieval system, use the citation-preserving PDF chunking guide.

Financial filings need additional controls for periods, units, amendments, and deterministic calculations; use the guide to parsing financial reports and 10-K tables for AI for that workflow.

When not to automate the final decision

Do not let extracted tables automatically approve payments, reject claims, change pricing, or make regulated decisions without validation. Table structure can look correct while one OCR error changes a key value.

Use field-level rules where possible: expected column count, numeric formats, totals, date ranges, and allowed units. Send failures and material discrepancies to a reviewer with the original page beside the extracted record.