Oracle Cloud Infrastructure Generative AI Professional — Lesson 2: Prompt Design, Embeddings, Vector Search, Grounding, and RAG Workflows

Learning Objectives

• Trace how prompt design, embeddings, vector search, grounding, and rag workflows changes data, application behavior, accounting, performance, or user experience.
• Configure a repeatable workflow with defined inputs, checkpoints, outputs, and exception paths.
• Defend functional correctness, access control, data quality, recoverability, and support readiness.

Detailed Topics

• Prompt Design
• Embeddings
• Vector Search
• Grounding
• RAG Workflows
• requirements, ownership, prerequisites, data, and acceptance criteria for prompt design, embeddings, vector search, grounding, and rag workflows
• security, controls, integration, performance, recovery, and support implications for prompt design, embeddings, vector search, grounding, and rag workflows

Instructor Demonstration

Demonstrate prompt design, embeddings, vector search, grounding, and rag workflows within an OCI generative-AI assistant grounded in enterprise content, evaluated for answer quality, protected by access controls, and operated within cost limits. Use representative data and capture model and endpoint settings, prompt tests, embedding and retrieval results, grounded responses, evaluation scores, guardrail output, access logs, latency, and cost metrics so another practitioner can independently verify the result.

Hands-On Exercise

Complete a controlled Oracle Cloud Infrastructure Generative AI Professional exercise focused on prompt design, embeddings, vector search, grounding, and rag workflows. Introduce one realistic failure involving hallucination, poor retrieval, sensitive-data exposure, prompt injection, unsafe output, uncontrolled tool use, high latency, and unexpected consumption cost; isolate the cause, apply the smallest reversible correction, and rerun the acceptance checks.

1. Record the requirement, environment, identities, dependencies, data, controls, and expected result.
2. Perform the lesson workflow and capture evidence before and after each material change.
3. Test normal, boundary, failure, security, reconciliation, recovery, and operational conditions relevant to the topic.
4. Document the root cause, correction, rollback, owner, support action, and next improvement priority.

Scenario Analysis

• Separate mandatory business requirements from convenience requests and local workarounds.
• Evaluate data, security, control, usability, performance, integration, and audit trade-offs.
• Choose the smallest supportable change and define evidence for normal, boundary, and failure conditions.

Troubleshooting Guidance

When prompt design, embeddings, vector search, grounding, and rag workflows does not produce the expected result, confirm scope, effective dates, identities, data context, dependencies, and recent changes. Review the Oracle-specific logs, reports, traces, calculations, workflow history, transaction status, or diagnostic output relevant to the lesson. Test one reversible hypothesis at a time and verify that the correction does not create hallucination, poor retrieval, sensitive-data exposure, prompt injection, unsafe output, uncontrolled tool use, high latency, and unexpected consumption cost elsewhere in the process.

Validation Checklist

• Prerequisites, ownership, data context, settings, and expected outcomes are documented.
• The workflow is reproducible and supported by model and endpoint settings, prompt tests, embedding and retrieval results, grounded responses, evaluation scores, guardrail output, access logs, latency, and cost metrics.
• Security, controls, data quality, integration, performance, recovery, and support conditions are reviewed.
• Exceptions, corrective action, rollback, approvals, monitoring, and handover are recorded.

Lesson Outcome

Learners can apply Prompt Design, Embeddings, Vector Search, Grounding, and RAG Workflows in a realistic Oracle Cloud Infrastructure Generative AI Professional situation and defend the result with course-specific evidence.