Structured Extraction: Tables, Headers, and Numerical Data

Update November 2, 2026 — Le Chat becomes Vibe. As of May 28, 2026, Le Chat is now Vibe (still at chat.mistral.ai): a single agent for work and code, carrying over your conversations, settings, and subscriptions. The legacy Le Chat experience remains accessible in Chat mode; in Work mode, agents have become skills (invoked with @), memories form the knowledge base, and deep search is the Deep Research skill. Plans are now called Mistral Free, Pro, Team, and Enterprise. This lesson uses the name Vibe; some menus described here may have changed names or locations.

From Raw Documents to Organized Data

Extracting text from a document is good. But real value appears when you get structured data: clean tables, identified figures, and classified headers. This is the difference between receiving a photocopy of a financial statement and receiving that same statement in a usable spreadsheet. Mistral’s Document AI excels precisely at this task: it doesn’t just read the text; it understands the document’s structure and returns information in a directly usable format.

Tables: The Sweet Spot

Tables are everywhere in business documents—financial statements, price lists, schedules, study results—and this is historically where classic OCR failed most miserably, mashing columns and rows into a mess. The OCR 3 engine, by contrast, handles them with remarkable precision. When you upload a document containing a table to Vibe, the model first identifies the table boundaries, even when they are subtle or colored. It then recognizes headers and distinguishes them from data, associating each cell with its column and row. It even handles difficult cases: merged cells, subtotals, footnotes—all classic traps that a poorly converted spreadsheet turns into silent errors.

To get a usable result, make it a habit to specify the desired format in your prompt. “Extract the table from page 3 in Markdown format” gives you a table readable directly in Vibe; “Convert this table to CSV so I can paste it into Excel” prepares it for import into a spreadsheet; “Extract only the ‘Amount’ column and the ‘Date’ column” filters what you need at the source. Markdown is often the most practical format for daily use: immediately readable and easy to copy-paste into other tools.

Headers, Hierarchy, and Structure

Beyond tables, Document AI can reconstruct the document hierarchy: titles and subtitles are identified and classified by level (h1, h2, h3), paragraphs are distinguished from one another even without visible line breaks, bulleted or numbered lists are correctly reconstructed, and footnotes are linked to their references. This capability is invaluable when working with long reports or legal documents, where structure carries as much meaning as the text: knowing a clause appears in the “Warranties” section rather than “Miscellaneous” changes its interpretation.

You can therefore make structural requests: “Give me the table of contents for this document,” “List all section titles with their page numbers,” or even “Extract only the paragraphs discussing warranties”—the latter combining structure and meaning.

Numerical Data, With Context

Financial documents, quotes, invoices, and activity reports are full of numbers: financial amounts (revenue, totals, VAT, discounts), signature or due dates, market share or margin percentages, and identifiers like invoice numbers, customer references, or SIRET codes. What sets Document AI apart is that it doesn’t just extract numbers—it understands their context. If you ask, “What is the annual revenue?”, it specifically looks for the number associated with that concept, even if it’s buried in a 50-row table. A classic OCR would have handed you fifty unlabeled numbers; here, you get the right one.

This understanding allows you to go beyond extraction: you can request calculations and verifications. “Is the total of the ‘Amount Excl. VAT’ column correct?” makes the model recount; “What is the difference between the budgeted and actual figures?” cross-references two columns; “Which expense item represents the largest share?” turns the table into an analysis.

For advanced users, JSON export structures the output for your tools: “Extract the following information from this invoice in JSON format: number, date, amount excl. VAT, VAT, amount incl. VAT, supplier” or “Create a JSON object with the data from each row of the table.” This format is ideal if the data needs to feed a spreadsheet, database, or other software later.

In Practice

See how these capabilities come together in three common scenarios. You receive a stack of PDF invoices: for each one, ask, “Extract from this invoice: supplier name, date, invoice number, amount excl. VAT, VAT, and amount incl. VAT. Present as a table”—and your accounting entry is prepared. Facing a 50-page annual report: “Extract all financial tables from this report and present them in Markdown. Identify revenue, net income, and operating margin.” To decide between two quotes, upload both and ask: “Compare the prices in these two quotes for the same services. Identify rate differences.”

A few methodological rules to wrap up. Be specific about what you’re looking for—a particular table, a column, a data type—and always specify the desired output format (Markdown, CSV, JSON, list). For scanned documents, a resolution of at least 150 DPI is recommended for reliable extraction. If the document is very dense, process it page by page: “Extract the table from page 7” will yield better results than a global “Extract all tables.” And for any critical use, cross-check extracted data against the original document—extraction is an accelerator, not a liability waiver.

Key Takeaways

• Document AI faithfully reconstructs tables with headers, columns, and merged cells
• You can request extraction in Markdown, CSV, or JSON depending on your needs
• The model understands the context of numerical data and can answer specific questions
• The document’s hierarchical structure (titles, subtitles, lists) is preserved
• Always verify extractions for critical uses