AI-103: Implement Text Analysis Solutions

Topic snapshot

Sample questions

These questions are original IT Mastery practice items aligned to this topic area. They are designed for self-assessment and are not official exam questions.

Question 1

Topic: Implement Text Analysis Solutions

You are building a Microsoft Foundry agent for equipment maintenance technicians. The RAG connection to Azure AI Search already grounds answers in manuals. The app must let technicians ask questions by voice through a headset and hear the agent’s answer aloud, with no keyboard or screen required. Which two capabilities should you add? Select TWO.

Options:

• A. Use Azure AI Speech text-to-speech for agent replies.

• B. Use Azure AI Speech speech-to-text for headset audio.

• C. Use Azure AI Language sentiment analysis on requests.

• D. Use Azure Translator for text-only localization.

• E. Use a text-only chat UI for responses.

• F. Use Azure AI Search semantic ranking for manuals.

Correct answers: A and B

Explanation: The requirement is about the user experience, not grounding quality. A hands-free headset workflow needs speech-to-text for spoken input and text-to-speech for audible output around the Foundry agent and RAG pipeline.

For a speech-enabled agent, the audio modality must be handled before and after the agent turn. Speech-to-text turns the user’s voice into text that can be sent to the agent and its Azure AI Search grounding flow. Text-to-speech then turns the agent’s final text response into audio so the technician can hear it without looking at a screen. Improving retrieval, analyzing sentiment, or translating text can be useful in other scenarios, but they do not satisfy the stated voice input and spoken output requirements by themselves.

• Retrieval improvement does not capture microphone input or produce audible responses.
• Sentiment analysis may classify tone, but it does not enable speech interaction.
• Text localization handles language translation, not the required headset-based speech input and output.
• Text-only chat violates the no-keyboard, no-screen user experience constraint.

Question 2

Topic: Implement Text Analysis Solutions

You are building a Microsoft Foundry voice support agent. Callers may speak English or Spanish, but the agent tools and Azure AI Search index use English text. The solution must detect urgency from the original audio, ground the answer in the English knowledge base, and reply by voice in the caller’s language.

Which two ordered workflow segments should you implement? Select TWO.

Options:

• A. Synthesize an English audio answer before grounding the agent response.

• B. Translate the synthesized audio before applying text safety checks.

• C. Analyze source audio, then produce English transcript before agent grounding.

• D. Send raw audio directly to Azure AI Search for semantic retrieval.

• E. Use text translation on the audio stream before speech-to-text.

• F. Translate finalized answer text, then synthesize speech in the caller’s language.

Correct answers: C and F

Explanation: The inbound flow should preserve the original audio for audio reasoning and produce English text for the agent, tools, and retrieval. The outbound flow should operate on finalized text first, translate it if required, and only then use text-to-speech for the voice reply.

For an agentic voice interaction, the order matters because each component expects a specific modality. Audio reasoning, such as urgency or emotion cues, should use the original audio. Speech-to-text or speech translation then creates text in the language required by the agent and its grounding data. After the agent completes retrieval, tool calls, and response generation, the final text should be translated to the caller’s language if needed and then synthesized with text-to-speech.

A reliable pattern is: source audio → audio reasoning and transcription/translation → agent grounding and tools → final text → translation → text-to-speech. Reversing these steps either sends the wrong modality to a service or produces speech before the agent has a grounded answer.

• Early synthesis fails because TTS should occur after the agent has produced a grounded final response.
• Raw audio retrieval fails because Azure AI Search retrieval needs indexed text/vector content, not unprocessed call audio.
• Text translation first fails because text translation requires text, so speech must be transcribed or speech-translated first.
• Audio translation last fails because safety and localization are normally applied to text before TTS output.

Question 3

Topic: Implement Text Analysis Solutions

You are configuring a Microsoft Foundry customer-support agent. The model deployment already uses content filters for hate, sexual, violence, and self-harm categories. You must add transcript analysis that tags each customer utterance as positive, neutral, or negative with opinion targets, and redacts PII before traces are stored.

Which TWO configurations should you add? Select TWO.

Options:

• A. Configure Azure AI Search semantic ranking

• B. Configure Azure AI Language PII recognition and redaction

• C. Configure Azure AI Language sentiment analysis with opinion mining

• D. Configure Azure Translator language detection

• E. Configure Azure Content Understanding layout extraction

• F. Configure Azure AI Content Safety text moderation

Correct answers: B and C

Explanation: The new requirements are transcript sentiment/tone analysis and PII redaction. Azure AI Language provides sentiment analysis with opinion mining for customer utterances and PII recognition for sensitive personal data before trace storage.

For text-based support transcripts, use Azure AI Language capabilities that directly analyze the text content. Sentiment analysis classifies text as positive, neutral, or negative, and opinion mining can identify targets of opinions, such as a product feature or service issue. PII recognition detects sensitive personal information such as names, phone numbers, email addresses, and identifiers, and can be used to redact those values before logs or traces are persisted.

Content Safety remains important for harmful content categories, but the stem states those filters are already enabled. The added controls should target sentiment/opinion extraction and PII handling.

• Safety moderation is useful for harmful content, but the scenario says those categories are already handled by deployment content filters.
• Semantic ranking improves retrieval relevance; it does not classify sentiment or redact sensitive entities.
• Language detection identifies or routes by language; it does not provide opinion targets or PII redaction.
• Layout extraction is for document structure and fields, not chat transcript sentiment analysis.

Question 4

Topic: Implement Text Analysis Solutions

A Foundry project includes a voice-enabled support agent. Customer audio must be converted to text before the agent chooses tools. Security policy requires keyless access to the speech service, private network traffic, moderation before tool calls, and auditability of each transcript. The solution must still allow normal account-support requests that may contain customer identifiers. Which workflow should you implement?

Options:

• A. Use Speech with an API key stored in Key Vault and public endpoint access.

• B. Block any utterance containing identifiers before creating a transcript.

• C. Use Speech with managed identity, private endpoint, transcript moderation, and trace provenance.

• D. Send audio directly to the agent and rely on the system prompt for safety.

Best answer: C

Explanation: The workflow needs secure speech-to-text before the agent can act. Managed identity with a private endpoint satisfies the access requirements, and transcript moderation plus trace provenance adds responsible AI controls without blocking valid support interactions.

For agentic speech workflows, audio should be transcribed by Azure AI Speech before the transcript is passed to the agent for reasoning and tool selection. A managed identity avoids embedded keys, private networking keeps traffic off the public internet path, and moderation or safety filtering should evaluate the transcript before tool invocation. Trace logging and provenance metadata connect the audio input, generated transcript, safety decision, and downstream agent action for audit review. The key is to evaluate and govern the transcript, not to reject broad categories of normal support speech.

• Key Vault key still uses a shared secret and does not satisfy the keyless-access requirement.
• Prompt-only safety does not provide controlled speech-to-text moderation or auditable enforcement before tool calls.
• Identifier blocking would reject legitimate account-support requests instead of applying appropriate moderation and controls.

Question 5

Topic: Implement Text Analysis Solutions

You are building a Foundry agent for compliance officers. For each regulatory notice, the agent must produce a 150-word summary and extract jurisdiction, obligation, deadline, and riskCategory using the company’s approved taxonomy. Test runs show inconsistent field names and occasional recommendations that are not present in the notice. You need to govern the agent output before it enters an approval workflow. What should you implement?

Options:

• A. Constrain the agent with domain instructions and a typed structured-output schema.

• B. Replace the agent with generic sentiment and key phrase extraction.

• C. Store prior reviewer comments in long-term memory for future notices.

• D. Increase the model context window and summarize the conversation.

Best answer: A

Explanation: The scenario requires consistent compliance summaries and domain-specific extracted fields. In a Foundry agent workflow, the best control is to combine domain instructions with a structured-output contract, such as a JSON schema or function schema, constrained to the approved taxonomy.

For compliance summarization and domain extraction, the agent should be given explicit role and output instructions and required to return a structured response with typed fields, allowed values, and missing-value behavior. This lets downstream approval workflows validate the output before use and reduces free-form variation such as renamed fields or unsupported recommendations. The same instructions can require the summary to be grounded only in the provided notice and to avoid adding advice not present in the source. Memory or larger context can provide more text, but they do not enforce a stable domain schema.

• Reviewer memory can help continuity, but it may introduce prior-case bias and does not enforce the required fields.
• Larger context may reduce truncation, but it does not solve inconsistent labels or unsupported recommendations.
• Generic text analytics can extract broad signals, but it is not tailored to the compliance schema and summary requirements.

Question 6

Topic: Implement Text Analysis Solutions

You are implementing a translation workflow in a Microsoft Foundry project for a medical-device support agent. English support articles are translated to Spanish before indexing. An approved glossary is stored in Azure AI Search and updated weekly: device names must remain in English, and terms such as ’lead’ must use the cardiac-specific Spanish term only when the surrounding context is cardiac. The team cannot fine-tune or retrain models for each update. Which implementation should you use?

Options:

• A. Retrieve relevant glossary entries as context, provide examples, and validate terms.

• B. Fine-tune a new translation model after each glossary update.

• C. Replace every matching English glossary word after translation.

• D. Call translation with automatic language detection and no glossary context.

Best answer: A

Explanation: Domain terminology should be preserved when the business provides approved terms and context-dependent meanings. A Foundry workflow can retrieve the relevant glossary entries, pass them as translation context or examples, and validate the translated output before indexing.

The core concept is terminology-controlled translation. Because the glossary changes weekly and some words depend on context, the workflow should use the glossary at runtime instead of retraining. In Foundry, the translation step can retrieve matching glossary entries from Azure AI Search, include them as constraints or few-shot examples for the translation model, and run post-translation checks that confirm required terms were preserved or translated as specified. This handles device names that must remain unchanged and terms whose translation depends on the surrounding domain context. Blind search-and-replace is unsafe because it ignores meaning and grammar, while generic translation does not enforce approved terminology.

• Generic translation fails because automatic language detection does not enforce business-approved medical terminology.
• Weekly fine-tuning fails because the stem explicitly rules out retraining for each glossary update.
• Blind replacement fails because context-dependent terms such as ’lead’ can be translated incorrectly or corrupt the sentence.

Question 7

Topic: Implement Text Analysis Solutions

A company builds a Microsoft Foundry agent that answers HR policy questions by using Azure AI Search as a grounding source. During ingestion, Azure AI Content Safety labels each chunk with safety category, severity, and a sensitive-content flag. Policy requires the agent to never use chunks flagged as prohibited content when composing grounded answers, and auditors must be able to trace which source chunks were used.

What should you implement?

Options:

• A. Store safety labels as filterable metadata and enforce filters in the retrieval tool.

• B. Retrieve all chunks, then block only the final answer if it violates policy.

• C. Add a system prompt that tells the agent to ignore prohibited retrieved chunks.

• D. Reduce the retriever top-k value and rely on semantic ranking.

Best answer: A

Explanation: The policy must be enforced before the agent uses retrieved content for grounding. Persisting text safety outputs as filterable index metadata lets the retrieval tool exclude prohibited chunks and keep source traceability for allowed chunks.

For an agentic RAG workflow, safety classifications should become enforceable retrieval metadata, not just advisory text. The ingestion pipeline can write category, severity, and sensitive-content flags to the Azure AI Search index as filterable fields. The agent’s grounding tool then applies a metadata filter that excludes prohibited chunks before they are passed to the model. This also supports auditability because the remaining retrieved chunks can retain source IDs, chunk IDs, and safety labels in traces or citations.

A prompt-only control is weaker because unsafe chunks are still provided as grounding context. The key takeaway is to enforce policy at the retrieval boundary and record provenance for what was actually grounded.

• Prompt-only control fails because prohibited chunks can still enter the model context before the agent decides how to respond.
• Final-answer blocking misses the requirement to prevent unsafe chunks from being used as grounding material.
• Semantic ranking tuning affects relevance, not policy enforcement based on text safety outputs.

Question 8

Topic: Implement Text Analysis Solutions

You integrate Azure AI Speech as the input modality for a Foundry customer-support agent. Typed questions about the same devices are grounded correctly, but spoken questions from field technicians often produce irrelevant answers. Trace logging shows:

User audio: "Show reset steps for the XR-17 valve kit."
Speech transcript: "Show reset steps for the extra seventeen value kit."
Retrieval query: "reset steps extra seventeen value kit"
Top result: general pricing FAQ

The audio is clear, and failures occur mainly for proprietary product names and alphanumeric part codes. What should you do next?

Options:

• A. Add retries around the manual lookup tool.

• B. Increase Azure AI Search topK for the retrieval tool.

• C. Use a custom speech recognition deployment for the domain vocabulary.

• D. Fine-tune the agent language model on support conversations.

Best answer: C

Explanation: The failure happens before retrieval or agent reasoning: the speech transcript changes product identifiers into common words. A custom speech recognition model or deployment configured with domain vocabulary helps the speech input preserve product names and part codes.

For a speech-enabled agent, troubleshooting should start at the earliest stage where evidence shows corruption. Here, typed questions work, but spoken questions create bad transcripts such as XR-17 becoming extra seventeen. That bad transcript is then used as the retrieval query, so the agent grounds on the wrong content. The best fix is to adapt the speech recognition component by using a custom speech deployment with the organization’s product names, part codes, and representative labeled audio or text data, then configure the agent’s speech modality to use that deployment.

Changing retrieval or the LLM would not reliably fix missing terms that were lost before the agent received the user intent.

• More search results can still rank irrelevant documents if the query contains the wrong product identifier.
• LLM fine-tuning does not repair speech-to-text errors that occur before the model receives the prompt.
• Tool retries address transient tool failures, not systematic transcript substitution of domain terms.

Question 9

Topic: Implement Text Analysis Solutions

A Foundry project builds a RAG compliance assistant over multilingual policy PDFs. Users ask in English, and the Azure AI Search configuration, embedding deployment, field extractor, safety classifier, and summarizer have been validated only on English text. Answers must cite the original PDF page and paragraph, not a derived translation file. Which ingestion and grounding pipeline should you implement?

Options:

• A. Extract and summarize source-language chunks, then translate only final summaries.

• B. Translate full PDFs first, then run OCR and cite the translated file.

• C. OCR/chunk with page metadata, translate chunks, then run extraction, safety detection, indexing, and English summarization.

• D. Index original-language chunks, then translate only the snippets returned by retrieval.

Best answer: C

Explanation: Translation should occur after OCR/chunking captures provenance, but before the English-only text analysis and retrieval steps. This lets Azure AI Search ground English prompts against English indexed content while preserving page and paragraph citations to the original PDFs.

The core concept is choosing the translation boundary without breaking grounding. OCR and chunking should first capture stable source metadata such as original PDF page and paragraph. After that, translate the chunk text to English, because the extractor, safety classifier, summarizer, and search configuration are all validated only for English. Indexing English chunks with original provenance metadata gives the retriever usable English terms/vectors and still allows citations to point back to the original document. Translating only after retrieval risks poor recall, while translating the whole PDF before OCR makes citations refer to a derived artifact rather than the source.

• Full-file translation fails because citations would be tied to a translated derivative instead of the original PDF location.
• Late translation fails because English-only retrieval may not find the best original-language chunks.
• Source-language processing fails because extraction, safety detection, and summarization are validated only on English text.

Question 10

Topic: Implement Text Analysis Solutions

A team uses a Microsoft Foundry app to customize a language model for compliance summarization. The output must include a short summary plus extracted obligations, deadlines, and citations from policy documents. Reviewers report that some outputs are fluent but miss obligations or invent deadlines. Which monitoring approach best validates the domain-output quality before each prompt or model update?

Options:

• A. Run regression evaluations against labeled documents with field accuracy and grounding checks.

• B. Measure sentiment polarity of each generated compliance summary.

• C. Track average token count and completion latency for each summary.

• D. Review only safety-filter events for harmful content categories.

Best answer: A

Explanation: The issue is domain accuracy, not fluency or operational performance. For customized compliance summarization and extraction, the best validation is a repeatable evaluation set with expected outputs, field-level comparisons, and grounding checks against source documents.

Domain-specific language model outputs should be evaluated against the business meaning they must preserve. In this case, the app must extract obligations and deadlines and cite supporting text, so a regression evaluation should use representative policy documents with expected fields and source evidence. Field-level accuracy or precision/recall helps detect omissions and wrong deadlines, while grounding or citation checks help detect unsupported claims. Running this before prompt or model updates catches quality regressions in the customized output format. Latency and token analytics are useful operational signals, but they do not validate compliance extraction correctness.

• Operational metrics are useful for cost and performance, but token count and latency do not show whether obligations are missing or fabricated.
• Sentiment analysis does not match the compliance extraction task or validate deadlines and citations.
• Safety events can reveal harmful content issues, but they do not measure domain accuracy or grounded extraction quality.

Continue with full practice

Use the AI-103 Practice Test page for the full IT Mastery route, mixed-topic practice, timed mock exams, explanations, and web/mobile app access.

Try AI-103 on Web View AI-103 Practice Test

Related focused pages

• Free AI-103 Full-Length Practice Exam
• Plan and Manage an Azure AI Solution
• Implement Generative AI and Agentic Solutions
• Implement Computer Vision Solutions
• Implement Information Extraction Solutions

Free review resource

Read the AI-103 Cheat Sheet on Tech Exam Lexicon, then return to IT Mastery for timed practice.