To make a model's responses more deterministic and less creative, you must reduce the randomness in its token selection process. The temperature and topp parameters directly control this.

Temperature: A lower value (e.g., 0.1) makes the model's output more deterministic by consistently selecting the tokens with the highest probability. A value of 0 makes it almost completely deterministic.

Top P (Nucleus Sampling): This parameter instructs the model to consider only a subset of tokens whose cumulative probability mass exceeds the topp value. A lower topp (e.g., 0.1) narrows the selection to a small group of the most likely tokens, reducing randomness and creativity.

Configuring lower values for these two parameters will achieve the desired deterministic behavior.

1. Microsoft Azure Official Documentation

"Azure OpenAI Service models": This document explicitly defines the key parameters for controlling model behavior.

temperature: "Controls randomness. Lowering the temperature means that the model produces more repetitive and deterministic responses. Increasing the temperature results in more unexpected or creative responses."

topp: "Similar to temperature

this parameter controls randomness but uses a different method. Lowering topp narrows the model's token selection to a smaller pool of the most likely tokens."

Source: Microsoft Learn

"Azure OpenAI Service models"

Section: "Completions models > Parameters".

2. Holtzman

A.

et al. (2020). "The Curious Case of Neural Text Degeneration". This peer-reviewed paper introduces and analyzes Nucleus Sampling (topp).

Section 3

"NUCLEUS SAMPLING": The paper explains that top-p sampling mitigates the issue of language models generating incoherent text by truncating the probability distribution to the most probable tokens (the nucleus)

thus controlling randomness more effectively than temperature alone.

DOI: https://doi.org/10.48550/arXiv.1904.09751

Custom question answering, a feature of Azure AI Language, is specifically designed to create a conversational question-and-answer layer over existing content, such as product manuals. This service minimizes development effort by allowing you to simply point to your content; it then automatically extracts and indexes information to build a knowledge base. This purpose-built solution is the most direct, cost-effective, and low-effort approach to creating a support chatbot from a single document, directly aligning with the stated requirements.

A. Azure AI Phi-3-medium with fine-tuning requires significant data preparation, training, and deployment effort, which contradicts the goal of minimizing development and cost.

C. This describes a Retrieval-Augmented Generation (RAG) pattern. While powerful, it is more complex and costly to implement, involving multiple services (Azure OpenAI, Azure AI Search), than the purpose-built Custom question answering feature.

D. Azure AI Document Intelligence is used for extracting text and structure from documents. It is a data extraction tool, not a complete chatbot or question-answering solution.

1. Microsoft Documentation, Azure AI Language, "What is question answering?": "Question answering provides cloud-based Natural Language Processing (NLP) that allows you to create a natural conversational layer over your data. It is used to find the most appropriate answer for any input from your custom knowledge base (KB) of information." It lists "Product Manuals" as a common data source.

Source: Microsoft Learn, learn.microsoft.com/en-us/azure/ai-services/language-service/question-answering/overview (Section: "When to use question answering").

2. Microsoft Documentation, Azure AI Language, "Question answering quickstart": The quickstart guide demonstrates that a user can create a fully functional knowledge base simply by providing a URL or uploading a file, highlighting the minimal development effort required.

Source: Microsoft Learn, learn.microsoft.com/en-us/azure/ai-services/language-service/question-answering/quickstart/sdk (Section: "Create a question answering project").

3. Microsoft Documentation, Azure OpenAI, "Azure OpenAI on your data": This document outlines the architecture which requires setting up an Azure AI Search resource, creating an index, and configuring an Azure OpenAI model. This multi-step process is more complex than the single-service approach of Custom question answering.

Source: Microsoft Learn, learn.microsoft.com/en-us/azure/ai-services/openai/concepts/use-your-data (Section: "How it works").

4. Microsoft Documentation, Azure AI Document Intelligence, "What is Azure AI Document Intelligence?": "Azure AI Document Intelligence is a cloud-based Azure AI service that enables you to build intelligent document processing solutions... It's not a standalone conversational agent."

Source: Microsoft Learn, learn.microsoft.com/en-us/azure/ai-services/document-intelligence/overview (Section: "What is Azure AI Document Intelligence?").

To integrate an Azure AI Search index with an Azure AI Agent using the SDK, you must follow this dependency order:

1. Initialize Connection: First, instantiate the Azure AI Client (AIProjectClient) and retrieve the unique connection ID for the linked Azure AI Search resource. This ID is required to authorize the agent's access to the service.
2. Configure Tool: Next, configure the Azure AI Search tool (e.g., AzureAISearchTool) by defining the specific index_name and the retrieved connection_id. This tells the tool what data to search.
3. Create Agent: Then, create the agent and enable the tool by passing the fully configured tool definition to the agent's constructor or update method.
4. Validate: Finally, ask questions to verify the agent correctly retrieves and synthesizes data from the vector index.

Microsoft Learn - Azure AI Agent Service: "How to use Azure AI Search tool in Azure AI Agent Service." Reference to the SDK steps for initializing AIProjectClient, retrieving connections via project_client.connections.get, and initializing AzureAISearchTool. [https://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/azure-ai-search-samples]

Microsoft Learn - Azure AI Projects SDK: "Azure AI Projects client library for Python - Connect to resources." Section on retrieving connection strings and IDs for tools. [https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/develop/sdk-overview]

1. Detect the incoming language: Speech to Text The input for this requirement is a live audio stream (a phone call). To process this, you must use the Speech to Text service, which specifically includes a Language Identification feature. This capability analyzes the audio input to determine the spoken language (e.g., distinguishing between French and English) before or during transcription. Services like Text Analytics or Translator require text input, making them unsuitable for the initial detection from raw audio.

2. Respond in the callers' own language: Text to Speech "Responding" in a call handling context implies generating an audible voice reply. The Text to Speech service is designed to convert the system's text-based responses into synthesized, human-like spoken audio in the specific target language identified (French or English).

Microsoft Learn. (n.d.). What is the Speech service? Retrieved from https://learn.microsoft.com/en-us/azure/ai-services/speech-service/overview

Relevance: Confirms that the Speech service provides "Language identification" to "identify languages spoken in audio" and "Text to speech" to "convert input text into human-like synthesized speech."

Microsoft Learn. (n.d.). Implement language identification. Retrieved from https://learn.microsoft.com/en-us/azure/ai-services/speech-service/language-identification

Relevance: Details how "Speech to text recognition" is used when you need to "identify the language in an audio source."

Microsoft Learn. (n.d.). Text to speech overview. Retrieved from https://learn.microsoft.com/en-us/azure/ai-services/speech-service/text-to-speech

Relevance: Validates that Text to speech enables applications to convert text into synthesized speech, supporting the requirement to "respond" vocally.

The Azure AI Search indexer pipeline processes data in a specific sequence:

1. Document cracking: The indexer connects to the data source and extracts text and metadata (cracks the document).
2. Field mappings: Source fields are mapped to destination fields or skill inputs. This happens before skill execution.
3. Skillset execution: The AI enrichment (custom skills, cognitive services) processes the data to create new information.
4. Output field mappings: The enriched data (output from skills) is mapped to the index fields. This must happen after the skills have run.
5. Push to index: The final assembled document is written (pushed) to the search index.

Microsoft Learn - Azure AI Search Indexer Overview

Section: Indexer stages

"Stage 1: Document cracking. Document cracking is the process of opening files and extracting content."

"Stage 2: Field mappings. Field mapping occurs after document cracking, but before transformations [skillsets]..."

"Stage 3: Skillset execution. Skillset execution is an optional step that invokes built-in or custom AI processing."

"Stage 4: Output field mappings. Output field mappings associate the content of an internal enriched document (skill outputs) to destination fields in the index."

URL: https://learn.microsoft.com/en-us/azure/search/search-indexer-overview

Microsoft Learn - Create a skillset in Azure AI Search

Section: Skillset definition

"A skillset executes after raw content is extracted from an external data source, and after field mappings are processed."

URL: https://learn.microsoft.com/en-us/azure/search/cognitive-search-defining-skillset

The task requires detecting the language of an email to route it correctly. This is a core feature of the Azure AI Language service (formerly known as Text Analytics).

1. Endpoint: The correct endpoint for the Language service is a regional one, following the format https://<region>.api.cognitive.microsoft.com. Therefore, eastus.api.cognitive.microsoft.com is the appropriate choice. The api.cognitive.microsofttranslator.com endpoint is specific to the Translator service, and portal.azure.com is the management portal, not an API endpoint.
2. API Path: The specific REST API operation for language detection within the Language service is located at the /text/analytics/v3.1/languages path. The other paths are for different functionalities: /entities/recognition/general is for Named Entity Recognition, and /translator/... is for performing translation, not just detection.

Microsoft Azure AI Language Documentation, "How to call the Language detection REST API": This official document specifies the endpoint structure and the request path. It shows a sample POST request to {Endpoint}/text/analytics/v3.1/languages, where {Endpoint} is the resource endpoint (e.g., https://my-resource.cognitiveservices.azure.com). The format eastus.api.cognitive.microsoft.com represents a valid regional endpoint for the service.

Source: Microsoft Docs, "Quickstart: Use the Language service for language detection". Section: "Send a REST API request".

Microsoft Azure AI Language Documentation, "Language detection feature": This documentation explicitly states that the language detection feature of the Azure AI Language service can detect the language in which a text is written. This directly matches the requirement to identify whether an email is in French or English.

Source: Microsoft Docs, "What is language detection in Azure AI Language?". Section: "Features".

To keep data on-premises, you must use Azure Cognitive Services containers deployed on local infrastructure (on-premises Kubernetes). Standard containers require internet connectivity to pull images and report usage metrics to the Azure billing endpoint; they do not send customer data (payloads) to the cloud. Azure container images are stored in the Microsoft Container Registry (MCR). Finally, when running the container, you must configure the Billing (Endpoint URL) and ApiKey arguments to validate the instance against your Azure subscription.

Microsoft Learn: Install and run Docker containers for the Language service

Section: "Prerequisites" / "The host computer"

Quote: "The host computer is the hardware that runs the Docker container... [It] must have an internet connection to send usage arguments to Azure."

Quote regarding Registry: "Container images for the Language service are available on the Microsoft Container Registry."

Link: https://learn.microsoft.com/en-us/azure/ai-services/language-service/custom-text-analytics-for-health/how-to/install-container

Microsoft Learn: Azure AI Services Containers - Frequently Asked Questions (FAQ)

Section: "How does the container communicate with Azure?"

Quote: "Containers generally require an internet connection to output usage information for billing... The container does not send customer data, such as the image or text that is being analyzed, to Microsoft."

Link: https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-container-support#billing

Microsoft Learn: Configure Azure AI services containers

Section: "Billing configuration setting"

Quote: "The Billing setting specifies the endpoint URI of the Azure AI services resource... The ApiKey setting specifies the Azure resource key used to track billing information."

Link: https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-container-support#billing-configuration-setting

To train a custom language model in Azure Video Indexer with industry-specific terms, you must provide the data as a text file (.txt). The service expects a simple list where each term or phrase is on a new line. This file is then used as training data to improve the accuracy of the speech-to-text transcription for your specific domain's vocabulary. The text file must be encoded as UTF-8 BOM.

A. PDF: This is a document format not supported for uploading vocabulary lists for language model customization in Video Indexer.

B. XML: This structured data format is not the specified format for providing a simple list of terms for language model training.

D. XLS: This spreadsheet format is not supported for this purpose; Video Indexer requires a plain text file for vocabulary lists.

1. Microsoft Learn. (2023). Customize a language model in Azure AI Video Indexer. In the "Upload training data" section, step 2 states, "Browse for a text file. You can upload a text file with a list of domain-specific words. Each word should be in its own line."

Reference: https://learn.microsoft.com/en-us/azure/azure-video-indexer/customize-language-model-with-api#upload-training-data

2. Microsoft Learn. (2023). Create a training data set to customize a language model. This document further clarifies the requirements: "Gather all the text files that contain domain-specific vocabulary... The files should be in UTF-8 BOM format."

Reference: https://learn.microsoft.com/en-us/azure/azure-video-indexer/create-training-data-set-for-language-model

The Text Analytics service, now part of the Azure AI Language service, is designed for natural language processing tasks on raw text. Its Sentiment Analysis and Opinion Mining features are specifically built to evaluate text and determine if the sentiment is positive, negative, neutral, or mixed. This directly addresses the requirement of measuring public perception of a brand by analyzing social media messages for opinions and feelings.

B. Content Moderator: This service is used to detect potentially offensive, risky, or undesirable content (text, images, video), not to analyze the positive or negative sentiment for brand perception.

C. Computer Vision: This service analyzes the content of images and videos. It is not designed to process and understand the sentiment within text-based social media messages.

D. Form Recognizer: This service, now Azure AI Document Intelligence, is used to extract text, key-value pairs, and tables from documents and forms, not for analyzing sentiment in unstructured text.

---

1. Microsoft Azure AI Language Documentation. "What is sentiment analysis and opinion mining in Azure AI Language?". This document explicitly states, "These features help you find out what people think of your brand or topic by mining text for clues about positive or negative sentiment...". This directly supports the use of Text Analytics (Language service) for the described scenario.

Source: Microsoft Learn, Azure AI Language documentation, "Concepts" section.

2. Microsoft Azure AI Content Safety Documentation. "What is Azure AI Content Safety?". The documentation defines the service's purpose: "Azure AI Content Safety is a content moderation service that helps you detect user-generated and AI-generated content... that is harmful or otherwise objectionable." This confirms its focus is on moderation, not sentiment analysis.

Source: Microsoft Learn, Azure AI Content Safety documentation, "Overview" section.

3. Microsoft Azure AI Vision Documentation. "What is Computer Vision?". The official documentation states, "The Azure AI Vision service gives you access to advanced algorithms that process images and return information based on the visual features you're interested in." This clarifies its purpose is for visual, not textual, analysis.

Source: Microsoft Learn, Azure AI Vision documentation, "Overview" section.

4. Microsoft Azure AI Document Intelligence Documentation. "What is Azure AI Document Intelligence?". The service is described as one that lets you "build automated data processing software... to extract text, key-value pairs, tables, and structures from your documents." This highlights its function for structured data extraction from documents, not sentiment analysis of social media posts.

Source: Microsoft Learn, Azure AI Document Intelligence documentation, "Overview" section.

Stream processing has access to the most recent data received or data within a rolling time window. ( Yes ) This is a fundamental characteristic of stream processing. It is designed to process unbounded streams of data in near real-time. To manage this continuous flow, computations are often performed over "windows" of time, such as a 5-minute rolling (or sliding) window, which provides access to the most recent events for analysis.

Batch processing must occur immediately and have latency in the order of seconds or milliseconds. ( No ) This statement describes stream processing, not batch processing. Batch processing is characterized by high latency, as it collects and processes large, finite volumes of data over a period (e.g., hours or days). In contrast, stream processing aims for low latency (milliseconds to seconds) to provide immediate insights.

Stream processing is used for simple response functions, aggregates, or calculations such as rolling averages. ( Yes ) This is a primary use case for stream processing. It excels at performing continuous calculations on data in motion. Common operations include stateful aggregations (e.g., counting events over a time period), transformations, and calculating metrics like rolling averages, which are essential for real-time monitoring and analytics dashboards.

Microsoft Azure Documentation. (2024). Introduction to Azure Stream Analytics. Microsoft Learn. In the section "How does Stream Analytics work?", it states, "Azure Stream Analytics makes it easy to ingest, process, and analyze real-time streaming data... You can also perform calculations such as running averages over a time window."

Microsoft Azure Documentation. (2023). Windowing functions (Azure Stream Analytics). Microsoft Learn. This document details how stream processing uses various windows (Tumbling, Hopping, Sliding) to operate on temporal blocks of data, confirming that it works on data within a time window.

Kleppmann, M. (2017). Designing Data-Intensive Applications. O'Reilly Media. In Chapter 11, "Stream Processing," Section: "Comparing Batch and Stream Processing," Kleppmann contrasts the two paradigms, noting that batch processing operates on bounded datasets with high latency, while stream processing operates on unbounded data with low latency. He explicitly mentions rolling averages as a typical stream processing task.

Akidau, T., Chernyak, S., & Lax, R. (2018). Streaming Systems: The What, Where, When, and How of Large-Scale Data Processing. O'Reilly Media. Chapter 2, "The What: Core Concepts," describes how stream processing systems are designed for low-latency, continuous computations and frequently use windowing to scope operations on an otherwise infinite data stream.