Certified Generative AI Engineer AssociateDemo

A safety filter provides a guardrail that identifies harmful or malicious inputs and prevents unsafe assistance by returning a refusal. Configurable safety settings support blocking content based on defined harm categories and thresholds.

A RAG workflow ingests and indexes source documents in a vector search system before users query the application. Each user query retrieves relevant context, which the LLM uses to generate a grounded response. The application is evaluated during development to assess quality before it is deployed through Model Serving.

Llama Guard is a specialized safety-classification model designed to identify unsafe LLM content, including generated responses. Deploying it through a provisioned-throughput Foundation Model API endpoint enables an automated output-moderation layer without designing a separate general-LLM moderation chain or brittle regex-based rules.

The unstructured package provides PDF partitioning through partition_pdf, producing document elements suitable for RAG pipelines. It supports extractable PDF text and can use OCR or high-resolution processing when content is image-based, avoiding the need to build custom PDF and image-processing logic.

Section headers supply semantic metadata and document hierarchy to each chunk, helping retrieval match a query to the appropriate policy section. Larger chunks retain more surrounding, related policy content, which can prevent context from being split away and support more complete answers. Official RAG guidance identifies section headers as useful semantic metadata and notes that larger chunks preserve more context.

A RAG system embeds the incoming query so it can perform vector similarity search for relevant content. The retrieved content is added to the user’s query to form a context-augmented prompt, which the response-generating LLM uses to produce the final answer.

Providing a clear JSON-output instruction together with an example of the expected fields and values gives the model a concrete output pattern to follow. This few-shot example improves extraction consistency and schema adherence compared with an instruction that only requests JSON.

ReAct interleaves reasoning traces with task-specific actions, allowing an LLM to form, maintain, and revise plans while invoking external tools and using their results in subsequent reasoning.

An allowlist of approved social-media and news accounts constrains the retrieval corpus to trusted sources, reducing the chance that toxic or otherwise untrusted content is supplied to the LLM and influences its output.

A profanity filter screens for offensive language but does not determine whether generated content is pertinent to financial news or current stock-news articles.

A correct classification prompt must explicitly state both the relevant condition—product availability—and the exact label assigned to each outcome. This ensures that available products are labeled “Success” and unavailable products are labeled “Fail.”

A Delta Sync Vector Search index uses a Delta source table with a primary-key column, so every chunk must be represented as an individual row with its own unique identifier. For standard endpoints, the source Delta table must also have Change Data Feed enabled to support incremental synchronization. Databricks: Create AI Search endpoints and indexes

A model card documents a model’s intended uses, limitations, training details, and the datasets used for training. Checking whether the model and its training data include medical-domain material provides evidence of relevance to medical documents; popularity, upload date, and model family alone do not establish domain suitability.

Provisioned Throughput reserves a specified amount of model throughput, providing predictable capacity and performance for production workloads with consistently high request volume. Pay-as-you-go requests use shared capacity and can be affected by contention.

Whisper large-v3 is an open-source automatic speech-recognition model built for multilingual speech transcription. The other models are text-focused large language models and do not perform speech-to-text transcription directly.

A contextual guardrail restricts interactions to the application’s intended domain and handles disallowed topics with a configured response. Politics is an out-of-scope topic for an insurance-only chatbot, so the guardrail should detect that context and return the standard refusal message. Amazon Bedrock Guardrails similarly supports denied topics that are undesirable in an application’s context and can return a blocked message when detected.

Model inversion can expose private training information by exploiting access to a model and its outputs. Restricting model access with ACLs and enforcing fine-grained attribute-based authorization reduce the ability of unauthorized parties to access or repeatedly invoke the proprietary model. Databricks ACLs control permissions on workspace objects, including MLflow models, and Unity Catalog ABAC centralizes scalable access policies based on governed tags.

Diversity of responses helps keep chatbot interactions fresh and engaging across repeated use, which can encourage users to remain active on the gaming platform. Repetitive or irrelevant responses reduce the quality of the experience, and randomness alone does not measure meaningful engagement.

A REST call made with the requests library requires the caller to explicitly provide authentication, such as an authorization token/header; requests has no Databricks session-aware credential provider. The Databricks Python SDK handles authentication details, and the databricks-openai client is automatically configured with authorization for generative AI model queries.

Properly curating upstream data, including manual review before it enters the RAG system, prevents offensive or inflammatory material from being retrieved and incorporated into generated responses.

Translation output should be accurate to the intended meaning and relevant to the requested content. These qualities help reveal unsafe or unsuitable transformations, such as meaning changes, omissions, or irrelevant generated text.

Code Llama is a code-specialized model family designed for code synthesis and understanding. The 34B Instruct variant is instruction-tuned for code-assistant and code-generation applications, making it the strongest out-of-the-box choice among the listed general-purpose alternatives.

Databricks Model Serving deploys MLflow-formatted models that have been registered in Unity Catalog as serving endpoints. MLflow logging captures the model and its dependencies during training, and the MLflow API registers that logged model to Unity Catalog for managed serving.

An agent architecture can use tools to retrieve structured data through SQL queries against Delta tables and to retrieve real-time unstructured information through web search, then supply the retrieved results to the LLM for response generation. This keeps stock prices and news current without requiring them to be copied into and continuously refreshed in a vector store.