Cu Sdk Setup

SDK User Environment Setup for Azure AI Content Understanding

Set up your Python environment to use the Azure AI Content Understanding SDK and run samples.

[COPILOT INTERACTION MODEL]: This skill is designed to be interactive. At each step marked with [ASK USER], pause execution and prompt the user for input or confirmation before proceeding. Do NOT silently skip these prompts. Use the ask_questions tool when available.

[COPILOT] Step numbering is stable contract. Authoring skills (cu-sdk-author-analyzer and cu-sdk-author-analyzer-classify-route) route the user to specific steps here on probe failures — e.g. "endpoint MISSING → Step 4 (env vars)". When renumbering or restructuring this skill, update the routing tables in those skills (or keep step numbers stable).

Prerequisites

Before starting, ensure you have:

• Python 3.9 or later installed
• An Azure subscription (create one for free)
• A Microsoft Foundry resource in a supported region

[COPILOT] Probe Python runtime first (before asking): Do not take the user's word for it — run these checks, then report. This prevents silent failures later in python -m venv.

# POSIX / WSL / macOS
python3 --version 2>/dev/null || python --version 2>/dev/null
python3 -m pip --version 2>/dev/null || python -m pip --version 2>/dev/null
python3 -c "import venv" 2>/dev/null && echo "venv: ok" || echo "venv: MISSING"

# Windows PowerShell
py -3 --version 2>$null; if ($LASTEXITCODE -ne 0) { python --version }
py -3 -m pip --version 2>$null
py -3 -c "import venv" 2>$null; if ($LASTEXITCODE -eq 0) { 'venv: ok' } else { 'venv: MISSING' }

Decision table:

Finding	Action
Python 3.9+ and venv: ok and pip present	✓ Good to go. Proceed to Step 1.
Python missing or < 3.9	Report the finding, then go to the [ASK USER] Python install choice block below.
venv: MISSING (common on Debian/Ubuntu base images)	Report the finding, then go to the [ASK USER] Python install choice block below.
Windows resolves to a WindowsApps\python*.exe stub	Report the finding, then go to the [ASK USER] Python install choice block below (auto-install cannot fix this — user must disable App Execution Aliases).

[ASK USER] Python install choice (only when probe fails): Ask the user: "Python is missing / too old / the venv module is unavailable. How would you like to proceed?"

• Option A: Install it for me — Agent runs the platform-appropriate install command (see below), verifies, and continues. Requires elevated privileges on Linux (sudo). Not available for the Windows App Execution Alias case — that requires manual UI action.
• Option B: I'll install it myself — Agent prints the install command for the user's platform and stops. User runs it, re-opens the terminal, and tells the agent to resume.

Default install commands (Option A):

• macOS → brew install python@3.12 (requires Homebrew; if not installed, fall back to Option B)
• Debian / Ubuntu / WSL → sudo apt update && sudo apt install -y python3.12 python3.12-venv (will prompt for sudo password)
• Debian / Ubuntu — venv: MISSING only → sudo apt install -y python3-venv (matches the existing Python version)
• Windows → winget install Python.Python.3.12 (run in an elevated PowerShell if needed)

Before running Option A, confirm with the user one more time by restating the exact command that will execute, then proceed. After install, re-run the probe to verify Python 3.9+ and venv: ok before continuing.

Windows gotcha — App Execution Aliases: Even after uninstalling the Store Python stub, the python.exe / python3.exe entries under C:\Users\<you>\AppData\Local\Microsoft\WindowsApps\ may persist as App Execution Aliases and shadow a real install. If the probe still reports a stub after the user confirms they installed real Python:

1. Point them to Settings → Apps → Advanced app settings → App execution aliases, and toggle off python.exe and python3.exe.
2. As a fallback, suggest invoking Python via the py launcher (py -3 -m venv .venv) or the explicit install path (e.g. %LOCALAPPDATA%\Programs\Python\Python312\python.exe). The companion setup_user_env.ps1 script already probes well-known install paths automatically.

Report the detected version + venv status back to the user in one sentence before the [ASK USER] block below.

[ASK USER] Prerequisites Check: After the probe above, confirm the remaining items:

1. "Do you already have a Microsoft Foundry resource set up in Azure?" — If no, jump to Step 5 (Azure Resource Setup) first, then return here.
2. "Have you already deployed the required AI models (gpt-5.2, text-embedding-3-large) in Microsoft Foundry?" — If no, include Step 5.3 and Step 6 in the workflow.

Package Directory

sdk/contentunderstanding/azure-ai-contentunderstanding

Workflow

Step 1: Navigate to Package Directory

cd sdk/contentunderstanding/azure-ai-contentunderstanding

Step 2: Create and Activate Virtual Environment

[ASK USER] Platform: Ask the user: "Which platform are you on?" with options:

• Linux/macOS
• Windows PowerShell
• Windows Command Prompt

Use their answer to show the correct activation command throughout the rest of the setup.

Check and create virtual environment:

if [ -d ".venv" ]; then
    echo "Virtual environment already exists at .venv"
else
    echo "Creating virtual environment..."
    python -m venv .venv
    echo "Virtual environment created at .venv"
fi

[COPILOT] Existing .venv behavior: If .venv already exists, prefer reusing it. If the existing virtual environment was created with a different Python minor version than the interpreter selected in the prerequisite probe, or if the environment is incomplete/corrupted, recreate .venv before continuing. This avoids subtle failures when the machine later upgrades from one supported Python version to another (for example, 3.9 → 3.12).

Activate virtual environment:

Verify activation:

which python  # Linux/macOS
# where python  # Windows
# Should show path inside .venv

[ASK USER] Confirm venv: After running the activation command, ask: "Does which python (or where python on Windows) show a path inside .venv? Please confirm the virtual environment is active before we install dependencies." If the user reports an error, troubleshoot before continuing.

Step 3: Install SDK and Dependencies

[ASK USER] Installation mode: Ask the user: "How would you like to install the SDK?"

• Option A: PyPI install (recommended) — Installs the latest published version from PyPI. Best for running samples and developing Content Understanding-based solutions using the SDK.
• Option B: Local editable install (for SDK contribution) — Use this only when you are contributing to the Content Understanding SDK. Installs from local source code; changes are reflected immediately without reinstalling.

Option A: PyPI install (recommended):

pip install azure-ai-contentunderstanding
pip install -r dev_requirements.txt

Option B: Local editable install (for Content Understanding SDK contribution):

pip install -e .
pip install -r dev_requirements.txt

[COPILOT] Repeated-run behavior: On repeated runs, if the required SDK/sample dependencies are already importable from the active virtual environment, the setup scripts may skip the pip install step instead of reinstalling everything. Only rerun the install commands when dependencies are missing or the virtual environment was recreated.

This also installs all dependencies needed to run samples:

• aiohttp - Required for async operations
• python-dotenv - For loading .env files
• azure-identity - For DefaultAzureCredential authentication

[ASK USER] Installation check: After running the install commands, ask: "Did the installation complete without errors?" If the user reports errors (e.g., permission issues, missing Python headers), help troubleshoot before continuing.

Step 4: Configure Environment Variables

4.1 Copy env.sample to .env (with safety check)

Important: This step copies the template without overwriting any existing .env file.

if [ -f ".env" ]; then
    echo "WARNING: .env file already exists - NOT overwriting"
    echo "If you want to start fresh, manually delete .env first: rm .env"
else
    cp env.sample .env
    echo "Created .env from env.sample"
    echo "Please edit .env and configure the required variables (see Step 4.2)"
fi

For Windows PowerShell:

if (Test-Path ".env") {
    Write-Host "WARNING: .env file already exists - NOT overwriting"
    Write-Host "If you want to start fresh, manually delete .env first: Remove-Item .env"
} else {
    Copy-Item env.sample .env
    Write-Host "Created .env from env.sample"
    Write-Host "Please edit .env and configure the required variables (see Step 4.2)"
}

[ASK USER] Existing .env: If a .env file already exists, ask: "A .env file already exists. Would you like to keep the existing one, or start fresh by deleting it and copying from env.sample?"

4.2 Configure Required Variables

[ASK USER] Authentication method: Ask the user: "How would you like to authenticate with Azure?"

• Option A: API Key — Configure CONTENTUNDERSTANDING_KEY locally without sharing it in chat.
• Option B: DefaultAzureCredential (recommended) — Uses az login or managed identity. No API key needed.

Based on their choice, guide accordingly below.

[ASK USER] Provide endpoint: Ask the user: "Please provide your Microsoft Foundry endpoint URL."

• It should look like: https://<your-resource-name>.services.ai.azure.com/
• Validate: it should NOT include api-version or other query parameters.
• If the user doesn't know where to find it: direct them to Azure Portal → Their Foundry resource → Keys and Endpoint.

[COPILOT] Configure API key locally (if Option A): Never ask the user to paste, send, or otherwise expose an API key in chat or a tool argument. Prefer DefaultAzureCredential. If the user chooses key authentication, tell them to get the key from Azure Portal → Their Foundry resource → Keys and Endpoint → Key1 or Key2, then either:

• edit CONTENTUNDERSTANDING_KEY in .env directly, or
• run the setup helper themselves in an interactive terminal and enter the key at its hidden prompt.

Ask only for confirmation that local configuration is complete. Do not read, print, summarize, or validate the secret value through the model.

If the user chose DefaultAzureCredential, remind them: "Make sure you've run az login to authenticate."

Open .env in your editor and set the following required variables:

For running sample_update_defaults.py (one-time model configuration):

[COPILOT] Probe existing model defaults on the Foundry resource: Before asking the user for deployment names, probe what the resource already has configured. The venv is active and the SDK is installed, so call get_defaults() via a short inline Python snippet. Export CONTENTUNDERSTANDING_ENDPOINT first. For Option A, have the user configure CONTENTUNDERSTANDING_KEY locally as described above; never pass it in a command or tool argument. The snippet inherits it from the environment.

python - <<'PY'
import os, sys
from azure.ai.contentunderstanding import ContentUnderstandingClient
from azure.core.credentials import AzureKeyCredential
from azure.identity import DefaultAzureCredential
from azure.core.exceptions import HttpResponseError, ClientAuthenticationError

ep = os.environ["CONTENTUNDERSTANDING_ENDPOINT"]
key = os.environ.get("CONTENTUNDERSTANDING_KEY") or None
cred = AzureKeyCredential(key) if key else DefaultAzureCredential()
try:
    d = ContentUnderstandingClient(ep, cred).get_defaults().model_deployments or {}
except ClientAuthenticationError:
    sys.exit(3)
except HttpResponseError as e:
    sys.exit(3 if e.status_code in (401, 403) else 1)
except Exception:
    sys.exit(1)

keys = ["gpt-5.2", "text-embedding-3-large"]
vals = [d.get(k, "") for k in keys]
print(";".join(f"{k}={v}" for k, v in zip(keys, vals)))
sys.exit(0 if all(vals) else (2 if not any(vals) else 10))
PY

Branch on the exit code:

Exit	Meaning	Action
0	ALL_SET — all 3 deployments already mapped on the resource	Show the detected values and ask "Detected existing defaults: gpt-5.2=<A>, text-embedding-3-large=<C>. Use these? (Y/n)". On Y, prefill the 3 env vars and skip Step 6 (defaults already configured). On n, fall through to the per-model prompts below.
10	PARTIAL — some mapped, some missing	Prefill the ones that are set. For missing models, ask per-item with the default shown below. After Step 4 completes, run Step 6 to fill the gaps.
2	NONE — resource has no defaults yet	Fall through to the per-model prompts below. Step 6 will configure them.
3	AUTH_ERROR (401/403)	Print a one-line warning: "Probe unavailable (auth failed). If you're using DefaultAzureCredential, run az login and ensure the Cognitive Services User role is assigned. Continuing with manual entry." Fall through to per-model prompts.
other	Unexpected error	Print "Probe failed: <error>. Continuing with manual entry." Fall through.

Only proceed to the per-model prompts below when the probe outcome requires it.

[ASK USER] Model deployment names (only when probe did not yield all values): Ask for the completion model (default: gpt-5.2), mini completion model (default: the completion model), and embedding model (default: text-embedding-3-large). For each model not already prefilled from the probe, ask for its deployment name. The mini deployment defaults to the completion deployment; the other deployment names default to their model names.

If the user prefers to configure these later, let them know they can run sample_update_defaults.py (Step 6) anytime before using prebuilt analyzers.

4.3 Validate Your Configuration

[ASK USER] Validate configuration: After the user has provided all values, summarize the configuration and ask them to confirm:

Here's your configuration:
  CONTENTUNDERSTANDING_ENDPOINT = <value>
  Authentication: API Key / DefaultAzureCredential
  CU_COMPLETION_MODEL = <value>
  CU_COMPLETION_MODEL_MINI = <value>
  CU_EMBEDDING_MODEL = <value>
  CU_COMPLETION_MODEL_DEPLOYMENT = <value>
  CU_COMPLETION_MINI_DEPLOYMENT = <value>
  CU_EMBEDDING_DEPLOYMENT = <value>

Does this look correct? (Yes / No — let me fix something)

Only write to .env after the user confirms.

Example .env configuration:

# Required for all samples
CONTENTUNDERSTANDING_ENDPOINT=https://my-foundry-resource.services.ai.azure.com/

# Optional: API key (if not set, DefaultAzureCredential is used)
CONTENTUNDERSTANDING_KEY=

# Required for sample_update_defaults.py (model configuration)
CU_COMPLETION_MODEL=gpt-5.2
CU_COMPLETION_MODEL_MINI=gpt-5.2
CU_EMBEDDING_MODEL=text-embedding-3-large
CU_COMPLETION_MODEL_DEPLOYMENT=gpt-5.2
CU_COMPLETION_MINI_DEPLOYMENT=gpt-5.2
CU_EMBEDDING_DEPLOYMENT=text-embedding-3-large

Step 5: Azure Resource Setup (if not done)

[NOTE]: Only guide the user through this step if they indicated during the prerequisites check that they do NOT yet have a Microsoft Foundry resource. Otherwise, skip to Step 6.

If you haven't set up your Microsoft Foundry resource yet:

5.1 Create Microsoft Foundry Resource

1. Go to Azure Portal
2. Create a Microsoft Foundry resource in a supported region
3. Navigate to Resource Management > Keys and Endpoint
4. Copy the Endpoint URL and optionally a Key

[ASK USER] Resource created: After guiding the user to create the resource, ask: "Have you created the Microsoft Foundry resource? Please share the endpoint URL so we can continue with configuration."

5.2 Grant Cognitive Services User Role

This role is required even if you own the resource:

1. In your Foundry resource, go to Access Control (IAM)
2. Click Add > Add role assignment
3. Select Cognitive Services User role
4. Assign it to yourself

[ASK USER] Role assigned: Ask: "Have you assigned the Cognitive Services User role to yourself? This is required even if you own the resource."

5.3 Deploy Required Models

To deploy a model:

1. In Microsoft Foundry → Deployments → Deploy model → Deploy base model
2. Search and deploy: gpt-5.2, text-embedding-3-large
3. Note deployment names (recommendation: use model name as deployment name)

[ASK USER] Models deployed: Ask: "Have you deployed the required models? Please provide the deployment names you used for each:"

• gpt-5.2 deployment name
• gpt-5.2 deployment name
• text-embedding-3-large deployment name

Use these names to populate the .env file.

Step 6: Configure Model Defaults (One-Time Setup)

[COPILOT] Skip condition: If the Step 4.2 probe returned ALL_SET and the user accepted the detected values, defaults are already configured on the Foundry resource — skip this step and tell the user "Your Foundry resource already has model defaults configured; skipping Step 6." Otherwise continue below.

[ASK USER] Run model defaults?: Ask: "Would you like to run sample_update_defaults.py now to configure model defaults? This is a one-time setup per Microsoft Foundry resource. (Yes / Skip for now)"

• If yes, ensure deployment name env vars are set, then run the script.
• If no, let them know they'll need to run it before using prebuilt analyzers.

Run the configuration script to map deployed models to prebuilt analyzers:

cd samples
python sample_update_defaults.py

This is a one-time setup per Microsoft Foundry resource.

Step 7: Run Samples

[ASK USER] Which samples?: Ask: "Which sample would you like to run first?" with options:

• sample_analyze_binary.py — Analyze a local PDF (quickest; completes in under a minute)
• sample_analyze_url.py — Full demo: document + video + audio + image from URLs (runs several analyses; takes a few minutes, please be patient)
• sample_analyze_invoice.py — Extract invoice fields
• Other — Let me see the full list
• Skip — I'll run samples on my own later

If the user picks "Other", list available samples from the samples/ directory.

[COPILOT] Timing note (do not parrot verbatim to user): sample_analyze_url.py runs 14 sequential long-running operations (LROs) (document + video + audio + image, with multiple content-range variants). Video/audio chapter generation is slow on the service side, so total runtime can be on the order of 15+ minutes today. Do not interpret quiet periods (no stdout for several minutes during a video/audio LRO) as a hang. Only consider killing if there is no new stdout for 5+ minutes AND no active HTTP traffic. When talking to the user, prefer phrasing like "takes a few minutes" or "please be patient" rather than citing exact large minute counts.

Sync Samples

# From the package directory with venv activated
cd samples

# Run sync samples
python sample_analyze_url.py
python sample_analyze_binary.py

Async Samples

cd samples/async_samples

# Run async samples
python sample_analyze_url_async.py
python sample_analyze_binary_async.py

[ASK USER] Sample result: After running a sample, ask: "Did the sample run successfully? Would you like to run another sample or are you all set?"

Automated Setup Script (Linux/macOS)

Run the interactive setup script that handles all steps automatically:

# From the package directory
cd sdk/contentunderstanding/azure-ai-contentunderstanding
.github/skills/cu-sdk-setup/scripts/setup_user_env.sh

The script will:

1. Create and activate a virtual environment
2. Install the SDK and dependencies
3. Copy env.sample to .env (without overwriting existing)
4. Interactively prompt you to configure required environment variables

Manual Quick Setup

If you prefer to run steps manually:

cd sdk/contentunderstanding/azure-ai-contentunderstanding

# Create and activate venv
[ ! -d ".venv" ] && python -m venv .venv
source .venv/bin/activate

# Install SDK and dependencies
pip install azure-ai-contentunderstanding
pip install -r dev_requirements.txt

# Copy env.sample if .env doesn't exist
if [ ! -f ".env" ]; then
    cp env.sample .env
    echo "Created .env - Please edit and configure required variables"
else
    echo "WARNING: .env already exists - skipping copy"
fi

Troubleshooting

Related Skills

• cu-sdk-sample-run - Run individual samples (including sample_update_defaults for model deployment setup)
• cu-sdk-common-knowledge - Domain knowledge for Content Understanding concepts

Additional Resources

• SDK README - Full documentation
• Samples README - Sample descriptions
• Product Documentation
• Prebuilt Analyzers