6.3.1.3. CLU Training and Deployment

6.3.1.3. CLU Training and Deployment

Train your model with labeled utterances, then deploy to production for real-time predictions.

Building on the capability spectrum from Section 1.1, CLU represents the "customizable" level for language understanding—you define intents and entities specific to your domain.

💡 First Principle: Users don't speak in commands—they say "Book me a flight to Seattle next Tuesday" not "BOOK_FLIGHT destination=Seattle date=Tuesday." CLU bridges this gap by learning to extract intent (what they want: BookFlight) and entities (the parameters: Seattle, Tuesday) from natural language. The exam tests whether you know CLU is for custom intent recognition—when pre-built capabilities don't cover your domain's specific intents.

🔧 Implementation Reference: CLU

Request Body:

{
    "kind": "Conversation",
    "analysisInput": {
        "conversationItem": {"id": "1", "participantId": "user", "text": "Book a flight to Paris"}
    },
    "parameters": {"projectName": "FlightBooking", "deploymentName": "production"}
}

Response Structure:

{
    "result": {
        "prediction": {
            "topIntent": "BookFlight",
            "entities": [{"category": "destination", "text": "Paris"}]
        }
    }
}

Testable Pattern:

from azure.ai.language.conversations import ConversationAnalysisClient
from azure.core.credentials import AzureKeyCredential

client = ConversationAnalysisClient(endpoint, AzureKeyCredential(key))

result = client.analyze_conversation(
    task={
        "kind": "Conversation",
        "analysisInput": {
            "conversationItem": {
                "id": "1",
                "participantId": "user",
                "text": "Book a flight to Paris tomorrow"
            }
        },
        "parameters": {
            "projectName": "FlightBooking",
            "deploymentName": "production"
        }
    }
)

top_intent = result["result"]["prediction"]["topIntent"]
confidence = result["result"]["prediction"]["intents"][0]["confidenceScore"]
entities = result["result"]["prediction"]["entities"]

Error Handling Pattern:

from azure.ai.language.conversations import ConversationAnalysisClient
from azure.core.exceptions import HttpResponseError, ResourceNotFoundError

try:
    result = client.analyze_conversation(task=conversation_task)
    
    # Check confidence threshold
    confidence = result["result"]["prediction"]["intents"][0]["confidenceScore"]
    if confidence < 0.7:
        # Low confidence - may need clarification from user
        logging.warning(f"Low intent confidence: {confidence}")
        
except ResourceNotFoundError:
    # Project or deployment not found
    logging.error("CLU project or deployment not found. Verify names are correct.")
except HttpResponseError as e:
    if e.status_code == 400:
        # Invalid request - check conversation item format
        logging.error("Invalid conversation request format")
    elif e.status_code == 429:
        # Rate limited
        time.sleep(int(e.response.headers.get("Retry-After", 60)))

CLI Equivalent (REST):

curl -X POST "https://{endpoint}/language/:analyze-conversations?api-version=2023-04-01" \
  -H "Ocp-Apim-Subscription-Key: {key}" \
  -H "Content-Type: application/json" \
  -d '{
    "kind": "Conversation",
    "analysisInput": {
      "conversationItem": {
        "id": "1",
        "participantId": "user",
        "text": "Book a flight to Paris"
      }
    },
    "parameters": {
      "projectName": "FlightBooking",
      "deploymentName": "production"
    }
  }'