feat: Add AzureDocumentIntelligenceConverter using the azure-ai-documentintelligence

[vblagoje]

Why

Azure deprecated azure-ai-formrecognizer in favor of azure-ai-documentintelligence (v1.0.0, GA Dec 2024). New package supports markdown output format (GitHub Flavored Markdown) which is better suited for RAG/LLM applications - tables inline with context, preserved document structure (headings, lists), no manual assembly required.

• fixes feat: Add a Azure OCR Converter that uses the azure-ai-documentintelligence library #8404

What

Added AzureDocumentIntelligenceConverter component:

• Uses azure-ai-documentintelligence>=1.0.0 package (2024-11-30 API)
• Markdown output mode (default): single document with inline tables, preserved structure
• Text output mode (backward compat): separate CSV table documents or markdown tables
• Simplified API: removed page_layout, threshold_y, preceding_context_len, following_context_len, merge_multiple_column_headers
• Added output_format (markdown/text), table_format (csv/markdown)

Deprecated AzureOCRDocumentConverter (removal in Haystack 2.25)

How can it be used

  from haystack.components.converters import AzureDocumentIntelligenceConverter
  from haystack.utils import Secret

  # Markdown mode (recommended for RAG)
  converter = AzureDocumentIntelligenceConverter(
      endpoint=os.environ["AZURE_DI_ENDPOINT"],
      api_key=Secret.from_env_var("AZURE_AI_API_KEY"),
      output_format="markdown"
  )
  results = converter.run(sources=["invoice.pdf"])
  # Returns single document with markdown, tables inline

  # Text mode (backward compat)
  converter = AzureDocumentIntelligenceConverter(
      endpoint=os.environ["AZURE_DI_ENDPOINT"],
      api_key=Secret.from_env_var("AZURE_AI_API_KEY"),
      output_format="text",
      table_format="csv"
  )
  # Returns separate CSV table documents + text document

How did you test it

• 3 unit tests (init, to_dict, from_dict)
• 4 integration tests with real Azure API (markdown output, text+CSV tables, metadata handling, multiple files)

Notes for the reviewer

Migration path from old converter:

• page_layout="natural" → output_format="markdown"
• Remove context/layout params (Azure API handles this now)
• Tables inline in markdown mode vs separate CSV docs

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Review	Updated (UTC)
haystack-docs	Ignored	Preview	Jan 8, 2026 1:15pm

[anakin87]

@vblagoje @sjrl could this be a good opportunity to move this component to core-integrations?
Just an idea, curious to hear your opinions.

[sjrl]

Yeah I think this would be a good time to move this to core integrations

See this comment #8404 (comment) from the issue thread

[vblagoje]

Deal, moving to https://github.com/deepset-ai/haystack-core-integrations/

[vblagoje]

Superseded by deepset-ai/haystack-core-integrations#2717 Closing