How Do I Parse Financial Reports and 10-K Tables for AI?

The finance-specific answer

Parse financial reports and 10-K tables for AI by separating narrative sections from tables, preserving units and reporting periods, and attaching every extracted value to its page and table context. Use an LLM for retrieval and explanation. Use deterministic code for arithmetic.

Where financial extraction breaks

Financial tables are context-heavy

The value 42.8 has no meaning without its row label, column period, currency, scale, and footnotes. “In millions” printed above a table can change every number below it.

Basic extraction often loses exactly those details. It may also mix two columns of narrative or repeat page headers inside table rows.

Parse structure before analysis

Use DocuShell Parse PDF to retain headings, tables, pages, and source coordinates. Store the original table cell text as well as any normalized numeric value.

For each table, capture:

• title and section heading
• reporting period and comparison period
• currency and scale
• row and column headers
• footnotes and continuation pages
• source page and coordinates

Separate retrieval from calculation

An LLM can find the right table and explain why a number matters. It should not be the only calculator for totals, growth rates, or reconciliations.

Retrieve the structured rows, run calculations in code, and pass both the result and source evidence to the model for a readable answer.

Handle long and continued tables

Annual reports often continue a table across pages. Repeated headers are helpful for reconstruction but should not become duplicate data rows.

Join continuation pages only when the column schema matches. Keep page boundaries in metadata so each cited value still points to its original location.

Build a review step

Test tables with parentheses for negatives, em dashes for missing values, nested row labels, and multi-line headers. Review material figures against the source PDF.

AI can make financial reports easier to search. Trust comes from keeping the arithmetic deterministic and the evidence visible.

Example: calculating regional growth

Suppose a table reports 2025 and 2026 revenue in millions. The pipeline should extract the two periods, currency, unit, region labels, original cell strings, and normalized numbers. Application code calculates growth. The LLM receives the result plus the relevant rows and explains the change.

If a footnote says a business unit was reclassified, retrieve it with the table. Otherwise the arithmetic may be correct while the comparison is misleading.

function growthRate(current, previous) {
  if (previous === 0) return null;
  return ((current - previous) / Math.abs(previous)) * 100;
}

Finance rule: let code calculate from validated structured values. Let the model explain the calculation, identify caveats, and cite the source. Do not reverse those roles.

Financial extraction tests worth keeping

Build fixtures for:

• parentheses used as negative values
• dashes that mean zero, unavailable, or not applicable
• values split across continued tables
• multiple currencies in one report
• units stated only in a caption
• nested row labels and subtotals
• restated prior-year figures

Check balance or subtotal relationships where the document provides them. These rules catch cell shifts that look plausible in isolation.

Separate filings, periods, and amendments

Index the filing date, reporting period, document type, and amendment status. Retrieval should not mix an amended filing with the original without telling the user.

When a question asks for a trend, fetch the correct periods and show which filing supplied each value. Keep source page links for every material number.

For the extraction layer, use the PDF table guide for LLMs. For output design, compare JSON, Markdown, HTML, and text. When the answer will be generated, follow page-level citation practices and evaluate support with RAG faithfulness checks.