Document Handling And Text Extraction

Document handling lets an agent receive attached content and project extracted text into the model turn.

Register Document Handling

using HPD.Agent;
using HPD.Agent.Middleware.Document;
using Microsoft.Extensions.AI;

var agent = await new AgentBuilder()
    .WithChatClient(chatClient)
    .WithDocumentHandling()
    .BuildAsync();

Attach content with DataContent:

byte[] pdfBytes = File.ReadAllBytes("report.pdf");

var message = new ChatMessage(ChatRole.User,
[
    new TextContent("Summarize this report."),
    new DataContent(pdfBytes, "application/pdf") { Name = "report.pdf" },
]);

Registered Decoder Families

The default text extraction registration includes decoder families for:

Content	Decoder family
Plain text	TXT and unknown-extension text fallback
Markdown	MD decoder
JSON/XML	structured text decoder
PDF	PDF decoder
Word	DOC/DOCX decoder
Excel	XLS/XLSX decoder
PowerPoint	PPT/PPTX decoder
Images	image decoder
Web/HTML	HTML and URL-like text decoder

Image decoding requires an OCR engine for actual text extraction. No built-in OCR engine is bundled. Without an injected IOcrEngine, image extraction should be treated as no text found.

Applications that need image OCR should register their own OCR implementation through the text extraction service registration path, such as AddTextExtractionWithOcr<T>().

Boundaries

Document handling registration and DataContent attachment are straightforward. The table above describes registered decoder families, not a guarantee that every PDF, Office file, HTML page, or image variant extracts useful text. Validate edge-case extraction quality with fixtures that match your document types.