Skip to main content

Document Filters 26.3 Release

· 6 min read
Nabih Metri
Nabih Metri
Product Manager

Documents communicate through more than paragraphs. Charts reveal trends, images add context, and tables organize related values, but that meaning can be lost when content is prepared for AI and search. Document Filters 26.3 preserves more of that meaning with structured chart data in Markdown, application-generated image descriptions, and a new PDF table detection engine for recovering complex tabular structures. This gets us to richer, more accurate content for RAG, analytics, search, and document automation.

Structured Chart Data in Markdown​

Charts contain valuable information that can disappear when a document is converted for downstream processing. A rendered chart preserves its appearance, but its title, axes, series, and data points remain difficult for LLMs, search, and analytics to interpret without an additional vision step.

Document Filters can now include chart titles and data directly in Markdown as structured JSON blocks, with an optional plain-table fallback. This gives RAG and other AI pipelines access to the values and labels behind the visual, making it easier to answer questions about trends, comparisons, and individual data points while keeping the chart in its original place within the document flow.

Chart extraction supports Microsoft Office files (DOCX, XLSX, PPTX, DOC, XLS, and PPT), OpenDocument files (ODT, ODS, and ODP), and Apple iWork files (Pages, Numbers, and Keynote).

Set MARKDOWN_INCLUDE_CHARTS=ON to include both structured JSON blocks and plain-table output. For more control, use MARKDOWN_INCLUDE_CHART_JSON and MARKDOWN_INCLUDE_CHART_TABLE to enable or suppress either representation independently. All three options default to OFF, so existing Markdown output remains unchanged unless chart extraction is enabled.

See Chart Extraction in Action​

In this example, a Word document contains a bar chart, but no separate text representation of its data. The document is converted to Markdown with MARKDOWN_INCLUDE_CHARTS=ON, which enables both chart output formats. The resulting Markdown contains a JSON block with the chart title, series, categories, and values, followed by a readable table containing the same data. This structured output makes the information behind the chart directly available to AI, search, and data-analysis workflows.

A Word bar chart converted to Markdown with MARKDOWN_INCLUDE_CHARTS set to ON, producing structured JSON and table output

Add Image Understanding to AI Workflows​

Important context is often locked inside screenshots, diagrams, and other embedded graphics. Without a useful text description, AI pipelines cannot include their meaning in search indexes, embeddings, or prompts. Document Filters 26.3 introduces an image description callback that lets applications enrich each graphic during document processing.

Applications can send an image to the vision model of their choice and return the generated description through the callback. Document Filters then includes that description in the Markdown output alongside the surrounding document content. This gives RAG and other AI workflows richer text to chunk, embed, retrieve, and reason over without requiring a separate post-processing step.

The callback provides context, such as the image's location, name, type, and existing description. Pixel data is generated only when requested, helping applications avoid unnecessary processing.

Watch Image Descriptions Preserve Context​

In this example, a Word document has an image woven between paragraphs of text. A script converts the document to Markdown, and when it reaches the embedded image, the callback sends its pixel data to Hyland Knowledge Enrichment, which generates a description. The two resulting Markdown files are shown side by side. On the left, converted without the callback, the image leaves no trace beyond its surrounding text. On the right, the generated description is inserted in Markdown exactly where the image appeared, keeping the image's meaning connected to the content around it.

A Word document with an embedded image converted to Markdown, comparing output without an image description to output with a generated description inserted in place of the image

A New Generation of PDF Table Detection​

One of the most substantial enhancements in Document Filters 26.3 is an entirely new PDF table detection engine. PDFs typically store text as positioned content rather than meaningful rows and columns, leaving applications to infer where tables begin, how cells relate, and which text belongs together.

Detection is only the first step. Document Filters reconstructs each table into rows, columns, and cells so its structure remains usable in Markdown and JSON. The engine can distinguish tabular content from similarly aligned body text, keep tables separated when prose appears between them, and identify adjacent tables independently on multi-column pages.

The new engine expands the range of PDF tables Document Filters can understand, while improving precision and structural reconstruction. Downstream AI, analytics, and automation workflows receive cleaner, more complete tables, without requiring a separate table-recognition service or custom reconstruction logic.

Compare V1 and V2 Table Detection​

This comparison starts with a PDF containing a complex, mostly borderless system risk table. The first conversion uses PDF_TABLE_DETECTION=V1. In the resulting Markdown, the detector separates related headers and values, while portions of the table are reconstructed as unrelated fragments. The same PDF is then converted with the new PDF_TABLE_DETECTION=V2 option. V2 recognizes the complete table and preserves the relationships among its columns, rows, and values, producing substantially cleaner Markdown for downstream processing.

The same PDF converted with PDF_TABLE_DETECTION set first to V1 and then to V2, showing V2 correctly reconstructing the table in Markdown

Document Filters Resources​