Azure multimodal extraction for Python

Azure AI Content Understanding SDK for Python

Multimodal AI service that extracts semantic content from documents, video, audio, and image files for RAG and automated workflows.

Installation

pip install azure-ai-contentunderstanding

Environment Variables

CONTENTUNDERSTANDING_ENDPOINT=https://<resource>.cognitiveservices.azure.com/  # Required for all auth methods
AZURE_TOKEN_CREDENTIALS=prod # Required only if DefaultAzureCredential is used in production

Authentication & Lifecycle

🔑 Two rules apply to every code sample below: 1. Prefer `DefaultAzureCredential`. It works locally (Azure CLI / VS Code / Developer CLI) and in Azure (managed identity, workload identity) with no code change. Avoid connection strings, account/API keys — they bypass Entra audit and rotation. - Local dev: DefaultAzureCredential works as-is. - Production: set AZURE_TOKEN_CREDENTIALS=prod (or AZURE_TOKEN_CREDENTIALS=<specific_credential>) to constrain the credential chain to production-safe credentials. 2. Wrap every client in a context manager so HTTP transports, sockets, and token caches are released deterministically: - Sync: with <Client>(...) as client: - Async: async with <Client>(...) as client: and async with DefaultAzureCredential() as credential: (from azure.identity.aio) Snippets may abbreviate this setup, but production code should always follow both rules.

import os
from azure.ai.contentunderstanding import ContentUnderstandingClient
from azure.identity import DefaultAzureCredential, ManagedIdentityCredential

endpoint = os.environ["CONTENTUNDERSTANDING_ENDPOINT"]
# Local dev: DefaultAzureCredential. Production: set AZURE_TOKEN_CREDENTIALS=prod or AZURE_TOKEN_CREDENTIALS=<specific_credential>
credential = DefaultAzureCredential(require_envvar=True)
# Or use a specific credential directly in production:
# See https://learn.microsoft.com/python/api/overview/azure/identity-readme?view=azure-python#credential-classes
# credential = ManagedIdentityCredential()
with ContentUnderstandingClient(endpoint=endpoint, credential=credential) as client:
    analyzers = list(client.list_analyzers())

Core Workflow

Content Understanding operations are asynchronous long-running operations:

1. Begin Analysis — Start the analysis operation with begin_analyze() (returns a poller)
2. Poll for Results — Poll until analysis completes (SDK handles this with .result())
3. Process Results — Extract structured results from AnalyzeResult.contents

Prebuilt Analyzers

Analyze Document

import os
from azure.ai.contentunderstanding import ContentUnderstandingClient
from azure.ai.contentunderstanding.models import AnalyzeInput
from azure.identity import DefaultAzureCredential

endpoint = os.environ["CONTENTUNDERSTANDING_ENDPOINT"]
with ContentUnderstandingClient(
    endpoint=endpoint,
    credential=DefaultAzureCredential()
) as client:
    # Analyze document from URL
    poller = client.begin_analyze(
        analyzer_id="prebuilt-documentSearch",
        inputs=[AnalyzeInput(url="https://example.com/document.pdf")]
    )

    result = poller.result()

    # Access markdown content (contents is a list)
    content = result.contents[0]
    print(content.markdown)

Access Document Content Details

from azure.ai.contentunderstanding.models import MediaContentKind, DocumentContent

content = result.contents[0]
if content.kind == MediaContentKind.DOCUMENT:
    document_content: DocumentContent = content  # type: ignore
    print(document_content.start_page_number)

Analyze Image

from azure.ai.contentunderstanding.models import AnalyzeInput

poller = client.begin_analyze(
    analyzer_id="prebuilt-imageSearch",
    inputs=[AnalyzeInput(url="https://example.com/image.jpg")]
)
result = poller.result()
content = result.contents[0]
print(content.markdown)

Analyze Video

from azure.ai.contentunderstanding.models import AnalyzeInput

poller = client.begin_analyze(
    analyzer_id="prebuilt-videoSearch",
    inputs=[AnalyzeInput(url="https://example.com/video.mp4")]
)

result = poller.result()

# Access video content (AudioVisualContent)
content = result.contents[0]

# Get transcript phrases with timing
for phrase in content.transcript_phrases:
    print(f"[{phrase.start_time} - {phrase.end_time}]: {phrase.text}")

# Get key frames (for video)
for frame in content.key_frames:
    print(f"Frame at {frame.time}: {frame.description}")

Analyze Audio

from azure.ai.contentunderstanding.models import AnalyzeInput

poller = client.begin_analyze(
    analyzer_id="prebuilt-audioSearch",
    inputs=[AnalyzeInput(url="https://example.com/audio.mp3")]
)

result = poller.result()

# Access audio transcript
content = result.contents[0]
for phrase in content.transcript_phrases:
    print(f"[{phrase.start_time}] {phrase.text}")

Custom Analyzers

Create custom analyzers with field schemas for specialized extraction:

# Create custom analyzer
analyzer = client.create_analyzer(
    analyzer_id="my-invoice-analyzer",
    analyzer={
        "description": "Custom invoice analyzer",
        "base_analyzer_id": "prebuilt-documentSearch",
        "field_schema": {
            "fields": {
                "vendor_name": {"type": "string"},
                "invoice_total": {"type": "number"},
                "line_items": {
                    "type": "array",
                    "items": {
                        "type": "object",
                        "properties": {
                            "description": {"type": "string"},
                            "amount": {"type": "number"}
                        }
                    }
                }
            }
        }
    }
)

# Use custom analyzer
from azure.ai.contentunderstanding.models import AnalyzeInput

poller = client.begin_analyze(
    analyzer_id="my-invoice-analyzer",
    inputs=[AnalyzeInput(url="https://example.com/invoice.pdf")]
)

result = poller.result()

# Access extracted fields
print(result.fields["vendor_name"])
print(result.fields["invoice_total"])

Analyzer Management

# List all analyzers
analyzers = client.list_analyzers()
for analyzer in analyzers:
    print(f"{analyzer.analyzer_id}: {analyzer.description}")

# Get specific analyzer
analyzer = client.get_analyzer("prebuilt-documentSearch")

# Delete custom analyzer
client.delete_analyzer("my-custom-analyzer")

Async Client

import asyncio
import os
from azure.ai.contentunderstanding.aio import ContentUnderstandingClient
from azure.ai.contentunderstanding.models import AnalyzeInput
from azure.identity.aio import DefaultAzureCredential

async def analyze_document():
    endpoint = os.environ["CONTENTUNDERSTANDING_ENDPOINT"]
    async with DefaultAzureCredential() as credential:
        async with ContentUnderstandingClient(
            endpoint=endpoint,
            credential=credential
        ) as client:
            poller = await client.begin_analyze(
                analyzer_id="prebuilt-documentSearch",
                inputs=[AnalyzeInput(url="https://example.com/doc.pdf")]
            )
            result = await poller.result()
            content = result.contents[0]
            return content.markdown

asyncio.run(analyze_document())

Content Types

Both derive from MediaContent which provides basic info and markdown representation.

Model Imports

from azure.ai.contentunderstanding.models import (
    AnalyzeInput,
    AnalyzeResult,
    MediaContentKind,
    DocumentContent,
    AudioVisualContent,
)

Client Types

Best Practices

1. Pick sync OR async and stay consistent. Do not mix azure.ai.contentunderstanding sync clients with azure.ai.contentunderstanding.aio async clients in the same call path. Choose one mode per module.
2. Always use context managers for clients and async credentials. Wrap every client in with ContentUnderstandingClient(...) as client: (sync) or async with ContentUnderstandingClient(...) as client: (async). For async DefaultAzureCredential from azure.identity.aio, also use async with credential: so tokens and transports are cleaned up.
3. Use `begin_analyze` with `AnalyzeInput` — this is the correct method signature
4. Access results via `result.contents[0]` — results are returned as a list
5. Use prebuilt analyzers for common scenarios (document/image/audio/video search)
6. Create custom analyzers only for domain-specific field extraction
7. Use async client for high-throughput scenarios with azure.identity.aio credentials
8. Handle long-running operations — video/audio analysis can take minutes
9. Use URL sources when possible to avoid upload overhead

Reference Files