Custom Language Models

Custom model terms — defined before code

Understanding what users want to do

Building a CLU model — the process

Python — Calling CLU from Python

Install: pip install azure-ai-language-conversations

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

# ── Create the CLU client ──────────────────────────────────────────────
# CLU is part of Azure AI Language — same cognitiveservices.azure.com endpoint
clu_client = ConversationAnalysisClient(
    endpoint   = os.environ.get("LANGUAGE_ENDPOINT"),  # cognitiveservices.azure.com
    credential = AzureKeyCredential(os.environ.get("LANGUAGE_KEY"))
)

# ── Analyse an utterance ───────────────────────────────────────────────
user_utterance = "I'd like to book a flight to Paris for next Friday"

# analyze_conversation() sends the utterance to your CLU model for classification
with clu_client:  # "with" ensures the client is properly closed
    result = clu_client.analyze_conversation(
        task = {
            "kind": "Conversation",   # Task type for CLU
            "analysisInput": {
                "conversationItem": {
                    "id"          : "1",
                    "participantId": "user1",
                    "text"        : user_utterance   # The user's message
                }
            },
            "parameters": {
                "projectName"   : "FlightBookingCLU",  # Your CLU project name
                "deploymentName": "production",        # Your deployment slot
                "verbose"       : True                 # Return all intent scores
            }
        }
    )

# ── Extract the prediction ─────────────────────────────────────────────
prediction = result["result"]["prediction"]

# Top intent — the intent with the highest confidence score
# This is what you act on in your application
top_intent    = prediction["topIntent"]            # e.g. "BookFlight"
top_score     = prediction["intents"][0]["confidence"] # e.g. 0.97
print(f"Intent: {top_intent} (confidence: {top_score:.2f})")

# Entities — key information extracted from the utterance
for entity in prediction.get("entities", []):
    # entity["category"]: the entity type you defined (e.g. "destination")
    # entity["text"]: the actual text extracted (e.g. "Paris")
    # entity["confidenceScore"]: how confident (0-1)
    print(f"  Entity: {entity['category']} = '{entity['text']}' "
          f"({entity['confidenceScore']:.2f})")

# ── Use the result in your application ────────────────────────────────
# Route to the right action based on the top intent
if top_intent == "BookFlight":
    print("→ Routing to flight booking handler")
elif top_intent == "GetWeather":
    print("→ Routing to weather information handler")
elif top_intent == "None":
    print("→ Utterance not understood — show help menu")

Building knowledge bases for FAQ bots

Custom Question Answering (the replacement for QnA Maker) lets you build a knowledge base from FAQ documents and web pages. Users ask questions in natural language and the service finds the best matching answer.

Building a Custom Q&A knowledge base

Python — Calling Custom Q&A

import os
from azure.ai.language.questionanswering import QuestionAnsweringClient
from azure.core.credentials import AzureKeyCredential

# QuestionAnsweringClient connects to your Custom Q&A knowledge base
qa_client = QuestionAnsweringClient(
    endpoint   = os.environ.get("LANGUAGE_ENDPOINT"),
    credential = AzureKeyCredential(os.environ.get("LANGUAGE_KEY"))
)

user_question = "How do I cancel my subscription?"

with qa_client:
    result = qa_client.get_answers(
        question        = user_question,
        project_name    = "CustomerSupportKB",    # Your Q&A project name
        deployment_name = "production",          # Deployed slot
        top             = 3                       # Return top 3 matching answers
    )

# result.answers is a list sorted by confidence (highest first)
for answer in result.answers:
    print(f"Confidence: {answer.confidence:.2f}")
    print(f"Answer: {answer.answer}")
    print(f"Source: {answer.source}")
    print()

# Use the top answer if confidence is above a threshold
# Confidence below 0.3-0.4 usually means no good match was found
if result.answers[0].confidence > 0.4:
    print(f"Bot response: {result.answers[0].answer}")
else:
    print("Bot response: I'm not sure I have an answer for that. Can I connect you with a human agent?")

Building custom extractors and classifiers

Building conversational bots

Azure Bot Service is the platform for hosting and deploying bots. The Bot Framework SDK is used to write the bot logic. Together they integrate with CLU, Custom Q&A, and messaging channels like Teams, webchat, Slack, and email.

Bot architecture — how the pieces connect

Custom Language Models — what the exam tests

Test your understanding