Text analysis with language models and Foundry Tools – Azure AI applications (AI-103, lesson 22)

What you learn in this lesson

get entities, topics, summary, sentiment and tone of a customer message from the language model as a fixed schema

weigh Language in Foundry Tools, Content Safety and a language model against each other according to the state of Learn

redact personal data with the PII detection of Language before text is logged

read harmful content with Content Safety by category and severity

steer the output with an instruction, enumerations and rules in code, and test it without a model

The travel assistant of Nordlicht Reisen GmbH (R-02) receives messages from the booking portal: Customers write in German, English and Danish (R-01). Before the assistant answers, every message is to be classified: language, summary, topic, sentiment, tone, places named, and whether it contains personal data or harmful content. Three building blocks come into question for this: a language model, Language in Foundry Tools and Content Safety.

Prior knowledge. Structured answers (GetResponseAsync<T>, strict schema, enumerations, checking in code, "not assessed" on broken JSON) are taught by "AI track, lesson T2.3". Putting foreign text in only as fenced data: "AI track, lesson T6.2"; hard rules in code: "AI track, lesson T6.4". The choice between model and Foundry Tool is justified by lesson 1.1 (decision 2). None of that is repeated here, only applied.

New in this lesson. The assistant has an input for messages whose result is a type MailAnalyse. Messages about refunds, legal matters or a blocked account always go to a person under R-02; a rule in the code decides that, never the model. All messages in the exercises are made up (R-12).

Paper, account, costs. Paper: decision and exercises. An Azure account is needed only for the optional run of the services. Cost drivers: tokens per million for the model, text units for Language and Content Safety (pricing pages).

The requirement

All numbers are assumptions of the task.

No.	Requirement	Type	Measure	Weight	Source
4.1-01	Every message delivers language, summary, topic, sentiment, tone, places as fixed fields.	F	100% of the 40 test cases schema-valid or "not assessed"	Must	R-10
4.1-02	Refund, legal matters, blocked account and every unreadable analysis go to a person.	R	0 exceptions in the test set	Must	R-02
4.1-03	Sentiment, tone and topic agree with the editorial team's assignment.	Q	at least 90% of the 40 cases	Should	R-10
4.1-04	Personal data are not in the log.	R	0 hits of the PII detection in the log	Must	R-12
4.1-05	Harmful content is detected with category and severity.	F	all four categories in the result	Should	Operations
4.1-06	No new build depends on a feature that Learn lists as being retired.	Q	0 such calls	Should	Operations

This lesson belongs to course AI-103

You will see the rest of this lesson and its video once you have bought the course. The goals above show what you learn here.

€129.00

incl. VAT, one-time payment

Access to all of the course's lessons with their exercises, solutions and videos on this site, for 24 months from the purchase. A self-study course: no tutoring, no exams or grading by us, and no certificate.

Withdraw from contract here