Edit

Data parsing for Agentic Retrieval in Foundry Local

Agentic Retrieval uses advanced data parsing to help you extract more value from your documents by capturing structure, tables, and images for more accurate, context-rich search and chat experiences. This article explains how data parsing works, what it extracts, and how to optimize your content for better retrieval results.

Important

Agentic Retrieval in Foundry Local is currently in PREVIEW. See the Supplemental Terms of Use for Microsoft Azure Previews for legal terms that apply to Azure features that are in beta, preview, or otherwise not yet released into general availability.

Key capabilities

Data parsing in Agentic Retrieval provides the following capabilities:

  • Enhanced document understanding: Extracts headings, tables, images, and formatting for richer context.
  • Semantic chunking: Splits text into meaningful sections, preserving context and document hierarchy.
  • Rich metadata: Captures paragraph headings, page numbers, and other structural details.
  • Intelligent table processing: Detects, merges, and indexes tables—even those spanning multiple pages.
  • Improved retrieval accuracy: Enables more relevant and precise search results.

When data parsing adds the most value

Data parsing is built in and is especially valuable when your documents include:

  • Tables, charts, or images that are important for your use case.
  • Complex structure, such as reports, scientific papers, or financial statements.
  • A need for precise search, filtering, or attribution (such as page numbers or section headings).

How data parsing works

Data parsing analyzes your documents to identify and preserve structure, including:

  • Headings, paragraphs, lists, and hierarchy.
  • Tables and images.
  • Multiple file formats, including PDF, Word, PowerPoint, HTML, Markdown, and common image types.

The process splits text into context-aware chunks that align with natural boundaries, such as sentences or sections. Each chunk includes metadata like headings and page numbers, making it easier to trace information back to its source.

Supported file formats

Data parsing supports the following file types:

  • .txt (text files)
  • .pdf (PDF documents)
  • .docx (Microsoft Word documents)
  • .pptx (Microsoft PowerPoint presentations)
  • .html (HTML documents)
  • .md (Markdown files)
  • .png, .jpg, .jpeg (images)

Text chunking and metadata

Data parsing uses semantic chunking to create meaningful sections. Each chunk includes metadata such as headings, page numbers, and unique chunk IDs. Table chunks include details like table index, shape, and a preview of the content. This approach helps preserve context and improves retrieval relevance.

Table extraction and processing

Data parsing automatically detects tables across all pages, including tables in scanned documents. It merges tables that span multiple pages, restores column headers, and ensures consistent column structure. Each table chunk includes metadata such as table index, shape, page numbers, section headings, and a preview of the table. This information makes tables searchable and retrievable with full context.

Image extraction and processing

Data parsing automatically detects images across all pages. Each image is stored as the original source in full quality with source page number metadata. This information supports full-quality image display and source page context during inference.

How data parsing improves results

Data parsing directly improves the quality of retrieval-augmented generation (RAG) responses. It helps you get better context retrieval, enhanced table understanding, and improved accuracy. For example, heading information helps identify relevant sections quickly, and page numbers allow precise source attribution. Semantic chunks align with query intent, and structured queries about tabular data are more accurate. These improvements help you deliver more accurate and trustworthy results to your users.

Performance considerations

Data parsing can take longer for large, complex documents because it extracts richer structure and metadata. It can also increase storage requirements per chunk. In most workloads, improved retrieval relevance and source attribution outweigh the extra processing time and storage.

Best practices

To get the best results from data parsing, follow these recommendations:

  • Use native digital formats (such as PDFs with selectable text) when possible.
  • Ensure tables are well-formatted and consistent.
  • Use clear, consistent headings to help the parser identify document structure.

Review your documents before ingestion to ensure they meet these guidelines.

Troubleshooting

If you notice missing or incomplete tables, check that your source documents are well-structured and legible. For unexpected results, review the metadata and chunk details to verify that headings and page numbers are captured. If ingestion fails, try simplifying the document or splitting it into smaller sections.

Related content

  • Last updated on