Cu Sdk Sample Run

Run a Specific Sample

Run a specific sample from the Azure AI Content Understanding SDK.

[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.

Prerequisites

• Python >= 3.9
• Virtual environment set up with SDK installed (see cu-sdk-setup skill)
• Environment variables configured in .env
• Model deployments configured via sample_update_defaults.py -- required for any sample that uses gpt-5.2 / text-embedding-3-large (prebuilt analyzers like prebuilt-invoice, custom analyzers with extract / generate / classify field methods, and sample_create_analyzer_with_labels)

[ASK USER] Prerequisites check (run BEFORE Step 1 of the Workflow below): Before starting the numbered Workflow, verify the user's environment by asking these three questions in order. Do NOT skip Step 1 until each has a yes/handled answer:

1. "Have you already set up your Python environment and installed the SDK?" -- If no, direct them to the cu-sdk-setup skill first; pause this skill until that finishes.
2. "Have you configured your .env file with your endpoint and credentials?" -- If no, direct them to Step 4 of the cu-sdk-setup skill.
3. "Have you run sample_update_defaults.py to configure model defaults?" -- Defer judgement: ask once you know which sample they want (Step 3 below), then check it against the "Samples that require sample_update_defaults first" list in Step 4.

Package Directory

sdk/contentunderstanding/azure-ai-contentunderstanding

Available Samples

All sync samples have async versions with _async suffix in samples/async_samples/.

Getting Started (Run These First)

sample_update_defaults -- Required First!

One-time setup - Configures model deployment mappings (gpt-5.2, text-embedding-3-large) for your Microsoft Foundry resource. Must run before using prebuilt analyzers.

sample_analyze_url -- Start Here!

Analyzes content from a URL using prebuilt-documentSearch. Works with documents, images, audio, and video.

• Key concepts: URL input, markdown extraction, multi-modal content

sample_analyze_binary

Analyzes local PDF/image files using prebuilt-documentSearch.

• Key concepts: Binary input, local file reading, page properties

Document Analysis

sample_analyze_invoice

Extracts structured fields from invoices using prebuilt-invoice.

• Key concepts: Field extraction (customer name, totals, dates, line items), confidence scores, array fields

sample_analyze_configs

Extracts advanced features: charts, hyperlinks, formulas, annotations.

• Key concepts: Chart.js output, LaTeX formulas, PDF annotations, enhanced analysis options

sample_analyze_return_raw_json

Gets raw JSON response for custom processing.

• Key concepts: Raw response access, saving to file, debugging

sample_to_llm_input

Advanced usage of the to_llm_input helper -- converts structured AnalysisResult into LLM-ready text (YAML front matter + markdown body).

• Key concepts: LLM integration, output options (fields-only, markdown-only, custom metadata), multi-segment video/audio rendering, page markers via content_range
• Python-only helper (not available in other language SDKs)

Custom Analyzers

sample_create_analyzer

Creates custom analyzer with field schema for domain-specific extraction.

• Key concepts: Field types (string, number, date, object, array), extraction methods (extract, generate, classify)
• Next step: to run this workflow on your own folder of documents (validate schema → create → batch test → stdout summary), use the cu-sdk-author-analyzer skill.

sample_create_classifier

Creates classifier to categorize documents (Loan_Application, Invoice, Bank_Statement).

• Key concepts: Content categories, segmentation, document routing
• Next step: to run this workflow on your own mixed-document packets, use the cu-sdk-author-analyzer-classify-route skill.

sample_create_analyzer_with_labels

Builds analyzers with training labels (labeled data from Azure Blob Storage).

• Key concepts: Labeled data, knowledge sources, Blob Storage SAS URIs
• Requires additional env vars: CONTENTUNDERSTANDING_TRAINING_DATA_SAS_URL (Option A) or CONTENTUNDERSTANDING_TRAINING_DATA_STORAGE_ACCOUNT + CONTENTUNDERSTANDING_TRAINING_DATA_CONTAINER (Option B)

Analyzer Management

sample_get_analyzer

Retrieves analyzer details and configuration.

sample_list_analyzers

Lists all available analyzers (prebuilt and custom).

sample_update_analyzer

Updates analyzer description and tags.

sample_delete_analyzer

Deletes a custom analyzer.

sample_copy_analyzer

Copies analyzer within the same resource.

sample_grant_copy_auth

Cross-resource copying between different Azure resources/regions.

• Requires additional env vars: CONTENTUNDERSTANDING_TARGET_ENDPOINT, CONTENTUNDERSTANDING_TARGET_RESOURCE_ID

Result Management

sample_get_result_file

Retrieves keyframe images from video analysis.

• Key concepts: Operation IDs, extracting generated files

sample_delete_result

Deletes analysis results for data cleanup.

• Key concepts: Result retention (24-hour auto-deletion), compliance

Workflow

Step 1: Navigate to Package Directory

cd sdk/contentunderstanding/azure-ai-contentunderstanding

Step 2: Activate Virtual Environment

source .venv/bin/activate  # Linux/macOS
# .venv\Scripts\activate   # Windows

[ASK USER] Confirm venv active: Ask: "Is your virtual environment active? Run which python (or where python on Windows) and confirm it points to a path inside .venv." If the user reports it is not active or does not exist, direct them to the cu-sdk-setup skill.

Step 3: Choose the Sample

Pick the sample (and sync/async variant) before configuring sample-specific settings -- subsequent [ASK USER] blocks in Step 4 are gated on the chosen sample.

[ASK USER] Which sample?: If the user has already named a specific sample (e.g., "run sample_analyze_invoice"), confirm rather than re-prompt: "You want to run <sample_name>. Continuing with that -- correct? (Reply "no" to see the full list.)"

If the user has not yet specified a sample, ask: "Which sample would you like to run?" with options:

• sample_analyze_url -- Analyze content from a URL (recommended for first-time users)
• sample_analyze_binary -- Analyze a local PDF/image file
• sample_analyze_invoice -- Extract structured fields from an invoice
• sample_create_analyzer -- Create a custom analyzer
• sample_create_analyzer_with_labels -- Create a labeled-data analyzer (auto-uploads training data)
• sample_update_defaults -- Configure model defaults (one-time setup)
• Other -- Let me see the full list

If the user picks "Other" (or replies "no" to the confirmation), show the full Available Samples list above or run the --list command (path note: from package root run .github/skills/cu-sdk-sample-run/scripts/run_sample.sh --list; from the workspace root prepend sdk/contentunderstanding/azure-ai-contentunderstanding/).

[ASK USER] Sync or async?: Ask: "Would you like to run the sync or async version of this sample?"

• Sync (default) -- Runs in samples/
• Async -- Runs in samples/async_samples/ with _async suffix

Step 4: Configure Sample-Specific Settings (if needed)

Now that the chosen sample is known, run only the subsections below that apply to it. Each subsection's [ASK USER] block is already gated on the relevant sample(s); skip a subsection if it doesn't match. For full environment setup (creating .venv, installing the SDK, writing .env), use the cu-sdk-setup skill. The subsections below cover settings that are sample-specific and may not have been configured during the initial setup.

Samples that require sample_update_defaults first

These samples invoke an analyzer that depends on gpt-5.2 / gpt-5.2 / text-embedding-3-large model deployments. They will fail with a model deployment not found error unless sample_update_defaults.py has been run for the resource:

Samples that do not require sample_update_defaults (safe to run first): sample_analyze_url, sample_analyze_binary, sample_analyze_return_raw_json, sample_to_llm_input, all analyzer-management samples (sample_get_*, sample_list_*, sample_update_analyzer, sample_delete_*, sample_copy_analyzer, sample_grant_copy_auth).

[ASK USER] Model defaults check (only if chosen sample is in the table above): Ask: "This sample requires model deployments to be configured. Have you already run sample_update_defaults.py for this Microsoft Foundry resource?"

• Yes -- proceed.
• No -- run it first: cd samples && python sample_update_defaults.py (only needed once per resource).
• Not sure -- it's safe to re-run; the call is idempotent.

Settings by sample

Setting up sample_create_analyzer_with_labels training data

sample_create_analyzer_with_labels needs labeled training data in an Azure Blob container. Two configurations are supported -- the sample auto-detects which one to use based on which env vars are set.

[ASK USER] Training data option (sample_create_analyzer_with_labels only): If the user chose sample_create_analyzer_with_labels, ask: "How would you like to provide labeled training data?"

• Option A -- Pre-generated SAS URL (recommended for production): You already have labeled files in a Blob container and a SAS URL with Read + List permissions.

  • Set CONTENTUNDERSTANDING_TRAINING_DATA_SAS_URL in .env to the full URL (https://<account>.blob.core.windows.net/<container>?sv=...&se=...).

• Option B -- Auto-upload from local files (recommended for first-time users): The sample will upload the bundled samples/sample_files/training_samples/ (receipts + label JSON) to your container and generate a User Delegation SAS automatically.

  • Set CONTENTUNDERSTANDING_TRAINING_DATA_STORAGE_ACCOUNT and CONTENTUNDERSTANDING_TRAINING_DATA_CONTAINER in .env.
  • Optionally set CONTENTUNDERSTANDING_TRAINING_DATA_PREFIX (e.g., training_samples/) to scope uploads to a virtual folder.
  • Requires DefaultAzureCredential to have Storage Blob Data Contributor on the storage account (run az login first).

If the user is unsure, recommend Option B. The next subsection ("Samples that need a local file") only applies if Option B is chosen.

Do not set both options. If both are configured, Option A (the SAS URL) takes precedence.

Samples that need a local file

The sample_analyze_binary and sample_analyze_configs samples load a local file from samples/sample_files/. The default file paths are built into the samples (sample_files/sample_invoice.pdf and sample_files/sample_document_features.pdf respectively). To use your own file, update the file_path variable in the sample code.

sample_create_analyzer_with_labels (Option B only) auto-uploads bundled labeled receipts from samples/sample_files/training_samples/ (receipt1.jpg, receipt1.jpg.labels.json, receipt2.jpg, receipt2.jpg.labels.json, etc.) to the configured Blob container. To use your own labeled data, replace the files under samples/sample_files/training_samples/ (keeping the <image>.labels.json naming convention) before running.

[ASK USER] Local file (if applicable): If the user chose a sample that requires a local file (sample_analyze_binary, sample_analyze_configs, or sample_create_analyzer_with_labels and chose Option B in the previous subsection), ask: "This sample uses local files from samples/sample_files/. Would you like to:"

• Use the default test files -- The sample has built-in paths under samples/sample_files/.
• Provide your own files -- For sample_analyze_binary / sample_analyze_configs, update the file_path variable in the sample code. For sample_create_analyzer_with_labels (Option B), replace the files under samples/sample_files/training_samples/.

Setting up sample_grant_copy_auth cross-resource environment

The sample_grant_copy_auth sample requires two separate Microsoft Foundry resources (source and target).

Add the following to your .env:

# Source is your CONTENTUNDERSTANDING_ENDPOINT (already configured above)
CONTENTUNDERSTANDING_SOURCE_RESOURCE_ID="/subscriptions/{subscriptionId}/resourceGroups/{resourceGroup}/providers/Microsoft.CognitiveServices/accounts/{sourceAccountName}"
CONTENTUNDERSTANDING_SOURCE_REGION="eastus"
CONTENTUNDERSTANDING_TARGET_ENDPOINT="https://your-target-foundry.services.ai.azure.com/"
CONTENTUNDERSTANDING_TARGET_RESOURCE_ID="/subscriptions/{subscriptionId}/resourceGroups/{resourceGroup}/providers/Microsoft.CognitiveServices/accounts/{targetAccountName}"
CONTENTUNDERSTANDING_TARGET_REGION="swedencentral"

[ASK USER] Cross-resource setup (sample_grant_copy_auth only): If the user chose sample_grant_copy_auth, ask:

1. "Do you have two separate Microsoft Foundry resources (source and target) set up?" -- If no, guide them to create a second resource.
2. "Your CONTENTUNDERSTANDING_ENDPOINT will be used as the source endpoint. Please provide the following for your source resource:" -- Source ARM Resource ID, Source region.
3. "Please provide the following for your target resource:" -- Target endpoint URL, Target ARM Resource ID, Target region.
4. Confirm: "Cross-resource copy works with both DefaultAzureCredential and API keys. Both resources must have the Cognitive Services User role assigned if using DefaultAzureCredential. Is this configured?"

[REQUIRED GATE] sample_create_analyzer_with_labels training data (sample_create_analyzer_with_labels only): sample_create_analyzer_with_labels silently falls back to "create analyzer without labeled data" when no training-data source is configured. That fall-back path completes end-to-end and prints ✓ Sample completed: sample_create_analyzer_with_labels even though the labeled-data API surface is not actually exercised, AND the analyze-test step (begin_analyze against a sample invoice) is also skipped. Before invoking python sample_create_analyzer_with_labels.py (or run_sample.sh sample_create_analyzer_with_labels), the agent must ask the user the questions below and act on the answer:

1. "Do you want to train with labeled data (recommended), or create the analyzer without training data (demo mode)?"
   • If demo mode: confirm explicitly — "I will run sample_create_analyzer_with_labels without training data. The output will say Knowledge sources: 0, the Testing analyzer with sample document... step will be skipped, and you will see a DEMO MODE banner. The labeled-data API path will not be exercised. OK to proceed?" Only continue after the user says yes; leave both Option A and Option B env vars empty/unset.
   • If with training data: continue with one of the next two questions.
2. "Will you use Option A (pre-generated SAS URL) or Option B (auto-upload via DefaultAzureCredential)?"
   • Option A: ask for the SAS URL and (optionally) prefix; walk through the manual-upload steps above if not yet done.
   • Option B: ask for the storage account name and container name; remind them about the Storage Blob Data Contributor role and az login.
3. "Did you upload the files at the container root or inside a subfolder?"
   • If root: leave CONTENTUNDERSTANDING_TRAINING_DATA_PREFIX unset.
   • If subfolder: ask for the prefix path (e.g., training_samples/).
4. Confirm: "For Option A, the SAS token must have at least List and Read permissions and must not be expired. For Option B, the signed-in identity must have Storage Blob Data Contributor on the storage account."

Belt-and-suspenders: run_sample.sh itself emits a loud DEMO MODE banner before the python step when neither CONTENTUNDERSTANDING_TRAINING_DATA_SAS_URL (Option A) nor both CONTENTUNDERSTANDING_TRAINING_DATA_STORAGE_ACCOUNT + CONTENTUNDERSTANDING_TRAINING_DATA_CONTAINER (Option B) are set, so the fall-back is unmissable in the captured run output. Treat that banner as a signal that validation is incomplete unless the user explicitly opted into demo mode in step 1.

Step 5: Run the Sample

Run manually (recommended):

# For sync samples
cd samples
python sample_analyze_url.py

# For async samples
cd samples/async_samples
python sample_analyze_url_async.py

Or use the script:

Path note: All run_sample.sh invocations below assume the package root (sdk/contentunderstanding/azure-ai-contentunderstanding) as the current directory. From the workspace root, prepend sdk/contentunderstanding/azure-ai-contentunderstanding/. The script itself resolves paths relative to its own location, so the working directory only affects how you type the command.

.github/skills/cu-sdk-sample-run/scripts/run_sample.sh <sample_name>

Examples:

# Run sync sample
.github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_analyze_url

# Run async sample
.github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_analyze_url_async

# With .py extension (also works)
.github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_analyze_invoice.py

After the Sample Runs -- Review Results and Explain the Sample

After the sample completes, the skill must do the following for the user (do not skip):

1. Show the terminal command to re-run this sample directly, so the user can iterate without the skill. For example:

   cd samples && python sample_analyze_url.py
   # or for async samples:
   cd samples/async_samples && python sample_analyze_url_async.py
   

   Substitute sample_analyze_url with the sample the user just ran.

2. Briefly explain the key code concepts demonstrated in the sample. Tailor the explanation to the specific sample; common concepts include:

   • Client creation -- how the ContentUnderstandingClient (or aio.ContentUnderstandingClient for async) is constructed (endpoint + DefaultAzureCredential or AzureKeyCredential)
   • Analyzer selection -- which prebuilt (prebuilt-documentSearch, prebuilt-invoice, etc.) or custom analyzer is used and why
   • Input type -- URL vs. binary stream (bytes) vs. local file
   • Result processing -- how the returned AnalyzeResult is traversed (pages, fields, contents)
   • Content type narrowing -- e.g., narrowing AnalyzedContent to AnalyzedDocumentContent / AnalyzedImageContent / AnalyzedAudioContent / AnalyzedVideoContent (using isinstance or content kind discrimination)
   • Long-running operations -- if the sample uses begin_analyze and the returned LROPoller (or AsyncLROPoller)

   In addition to the generic concepts above, prefer the sample-specific highlights below when they apply:

   Sample	Sample-specific concepts to call out
   sample_analyze_url	URL-based analysis, multi-modal content (document / image / audio / video), markdown output
   sample_analyze_binary	Reading local files as bytes, page properties, content type narrowing
   sample_analyze_invoice	prebuilt-invoice field schema, Field confidence scores, array fields (line items)
   sample_analyze_configs	AnalysisConfig options (charts, formulas, hyperlinks, annotations), Chart.js / LaTeX outputs
   sample_analyze_return_raw_json	Accessing the raw JSON via cls=lambda pipeline_response, deserialized, _: pipeline_response.http_response.text() -- bypassing the typed model
   sample_to_llm_input	to_llm_input() helper, output options (fields-only / markdown-only / custom_metadata), content_range page markers
   sample_update_defaults	update_defaults() -- mapping model aliases to your deployment names; idempotent one-time setup
   sample_create_analyzer	Analyzer schema, field types, extract / generate / classify field methods
   sample_create_classifier	ContentCategory definitions, document segmentation, classification routing
   sample_create_analyzer_with_labels	LabeledDataKnowledgeSource, User Delegation SAS auto-generation (Option B), labeled receipts upload pattern, combining labeled data with extract / generate field methods
   sample_get_analyzer / sample_list_analyzers	Listing prebuilt + custom analyzers, paging
   sample_update_analyzer	Patching analyzer description / tags
   sample_delete_analyzer / sample_delete_result	Cleanup patterns; result auto-deletion after 24h
   sample_copy_analyzer	Same-resource copy via begin_copy_analyzer
   sample_grant_copy_auth	Cross-resource copy: grant_copy_auth on source + begin_copy_analyzer on target, DefaultAzureCredential with Cognitive Services User role
   sample_get_result_file	Operation IDs, retrieving generated keyframes from video analysis

[ASK USER] Sample result: Ask: "Did the sample run successfully?"

• If yes: present the re-run command and the key-code explanation (above), then ask: "Would you like to run another sample, or are you all set?"
• If no: help troubleshoot using the Troubleshooting section below. Common issues include missing .env configuration, inactive venv, or model defaults not configured.

[ASK USER] Run another?: If the user wants to run another sample, loop back to the "Which sample?" prompt above.

Quick Reference

Most Common Samples for New Users

1. First-time setup (run once per Foundry resource):

   .github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_update_defaults
   

2. Analyze a document from URL:

   .github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_analyze_url
   

3. Analyze a local PDF file:

   .github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_analyze_binary
   

4. Extract invoice fields:

   .github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_analyze_invoice
   

List Available Samples

.github/skills/cu-sdk-sample-run/scripts/run_sample.sh --list

Scripts

This skill includes a single helper script in the scripts/ directory.

run_sample.sh — Run a sample (activates .venv and loads .env automatically)

A convenience wrapper that activates the virtual environment, sources .env, and runs the sample. Detects sync and async variants automatically.

# Run a sync sample by name (with or without .py extension)
.github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_analyze_url
.github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_analyze_invoice.py

# Run an async sample
.github/skills/cu-sdk-sample-run/scripts/run_sample.sh sample_analyze_url_async

# List all available samples (sync and async)
.github/skills/cu-sdk-sample-run/scripts/run_sample.sh --list

For environment setup (creating .venv, installing the SDK, writing .env), use the cu-sdk-setup skill's setup_user_env.sh / setup_user_env.ps1 script.

Troubleshooting

Related Skills

• cu-sdk-setup - Set up environment for running samples
• cu-sdk-common-knowledge - Domain knowledge for Content Understanding concepts

Additional Resources

• Samples README - Detailed sample descriptions with key concepts
• SDK README - Full SDK documentation
• Product Documentation