azure di

Azure Document Intelligence Adapter

Code reference: backend/app/adapters/ocr/azure_di.py Config model: backend/app/config/settings.py → AzureDIConfig

What Azure DI Provides

Azure Document Intelligence (formerly Form Recognizer) is the default OCR engine in azure_di pipeline mode. It handles the full extraction pipeline — PDF to text with per-word metadata. Key capabilities:

Capability	Detail
Per-word handwriting detection	Every word carries an is_handwritten boolean
Per-word confidence scores	Float 0.0–1.0 indicating OCR certainty
Barcode / QR-code reading	17+ symbologies decoded per page
Selection mark detection	Checkbox and radio-button state with confidence

---

Deployment Modes

The same AzureDIOCRAdapter class is used in all environments. Only the endpoint URL changes:

Mode	Environment	Endpoint example
Cloud API (Azure AI Foundry)	Dev / staging	https://<resource>.cognitiveservices.azure.com
Disconnected container	Production on-prem	http://localhost:5000

Configuration

class AzureDIConfig(BaseModel):
    endpoint: str = "https://<resource>.cognitiveservices.azure.com"
    api_key: str = ""
    features: list[str] = Field(
        default_factory=lambda: ["barcodes", "keyValuePairs"]
    )

The features list controls which add-on capabilities are requested. barcodes enables barcode extraction and keyValuePairs enables key-value pair detection (useful for forms).

---

Client Initialisation

Like Marker, the Azure DI client is lazily initialised on first use:

def _get_client(self):
    if self._client is not None:
        return self._client

    from azure.ai.documentintelligence import DocumentIntelligenceClient
    from azure.core.credentials import AzureKeyCredential

    self._client = DocumentIntelligenceClient(
        endpoint=self._config.endpoint,
        credential=AzureKeyCredential(self._config.api_key),
    )
    return self._client

The SDK import is deferred into the method body so the module can be loaded even if the azure-ai-documentintelligence package is not installed (useful for local dev with Marker-only runs).

---

Analysis Flow

The adapter calls the prebuilt-layout model with the configured feature flags:

poller = client.begin_analyze_document(
    "prebuilt-layout",
    analyze_request=AnalyzeDocumentRequest(bytes_source=pdf_bytes),
    features=self._config.features,
)
result = poller.result()

The call is wrapped in run_in_executor because the SDK's begin_analyze_document / poller.result() cycle is synchronous and involves network I/O:

result = await loop.run_in_executor(None, _analyze)

---

Extracted Data

Per-Word Data

For every page, each word is mapped to an OCRWord:

OCRWord(
    text=word.content,
    confidence=word.confidence,          # 0.0–1.0
    is_handwritten=word.is_handwritten,  # bool
    bounding_region=BoundingRegion(
        page_num=page_num,
        x=min(x_coords),
        y=min(y_coords),
        width=max(x_coords) - min(x_coords),
        height=max(y_coords) - min(y_coords),
    ),
)

The BoundingRegion is computed from the Azure polygon (list of x,y coordinate pairs) by taking the min/max extents.

Barcodes

Each detected barcode is mapped to a BarcodeResult:

BarcodeResult(
    barcode_type=bc.kind,   # e.g. "Code128", "QRCode", "EAN13"
    value=bc.value,         # decoded string
    page_num=page_num,
)

Azure DI supports 17+ barcode symbologies including Code 39, Code 93, Code 128, UPC-A, UPC-E, EAN-8, EAN-13, ITF, Codabar, Data Matrix, QR Code, PDF417, and more.

Selection Marks

Checkboxes and radio buttons are extracted as SelectionMark:

SelectionMark(
    state=sm.state,            # "selected" or "unselected"
    confidence=sm.confidence,  # 0.0–1.0
    page_num=page_num,
)

---

Cross-Page Support

Azure DI's bounding_regions can reference multiple pages, enabling native cross-page table and field detection. The adapter currently extracts the first bounding region per word:

def _to_bounding_region(regions, page_num):
    if not regions:
        return None
    r = regions[0]
    # ... compute from polygon ...

For tables, Azure DI natively tracks cell spans across pages via bounding_regions on each cell.

---

Confidence Scoring (azure_di mode)

In azure_di pipeline mode, confidence comes directly from DI's per-word scores:

Page confidence = 0.50 × avg_word_confidence + 0.20 × min_word_confidence + 0.30 × validation_pass_rate

Pages below the HITL threshold are routed for human review.

---

Capability Flags

supports_handwriting()    → True
supports_barcodes()       → True
supports_selection_marks() → True

---

Port Model Reference

The adapter uses these port models from backend/app/core/ports/ocr.py:

Model	Purpose
OCRWord	Per-word text, confidence, handwriting flag, bounding box
BarcodeResult	Decoded barcode type + value per page
SelectionMark	Checkbox/radio state + confidence per page
BoundingRegion	Rectangular bounding box (page, x, y, width, height)
OCRPageResult	Aggregation of words, barcodes, selection marks per page
OCRResult	Collection of page results + full Markdown content

---

Related Pages

• OCR engine overview
• Marker adapter
• Docling quality adapter