Text analysis in Foundry

Teach

This lesson introduces text analysis capabilities in Azure AI Foundry and when to use them in real solutions.

Start free lesson

• This is the free lesson for Module 5 (Foundry text and speech).

Text analysis capabilities in Foundry

• Sentiment analysis estimates whether text is positive, neutral, negative, or mixed.
• Key phrase extraction identifies important terms from longer text.
• Language detection predicts the primary language of input text.
• Text classification maps text to labels (for example support category or intent).
• Entity recognition identifies items such as people, organizations, locations, dates, and more.

Prebuilt language models vs custom models

• Use prebuilt models when your use case matches common NLP tasks and you want fast deployment.
• Use custom models when domain language, labels, or extraction targets are specific to your business.
• A common approach is starting prebuilt for baseline value, then moving to custom if quality gaps remain.

Typical REST and SDK call pattern

1. Create the Foundry language resource and capture endpoint + key.
2. Choose the task endpoint (for example sentiment, entities, or classification).
3. Send one or more text documents in request payload.
4. Parse per-document results and confidence scores.
5. Log outcomes and monitor for drift or misclassification patterns.

Python-style pseudocode (concept)

endpoint = os.getenv("FOUNDRY_LANGUAGE_ENDPOINT")
api_key = os.getenv("FOUNDRY_LANGUAGE_KEY")

documents = [
    {"id": "1", "text": "The support engineer resolved my issue quickly."},
    {"id": "2", "text": "Delivery was late and the packaging was damaged."},
]

result = analyze_text(
    endpoint=endpoint,
    api_key=api_key,
    task="sentiment",
    documents=documents,
)

for doc in result["documents"]:
    print(doc["id"], doc["sentiment"], doc["confidenceScores"])

Optional employer-lab checklist (text analysis)

1. Test one sample for sentiment, key phrases, and language detection.
2. Compare confidence scores across short vs long text inputs.
3. Map entity output to a simple downstream workflow (for example ticket routing).
4. Document where prebuilt output is sufficient and where custom training may help.
5. Confirm secrets are stored outside source code.