OCI GenAI Pro study plan

Explain how transformer-based large language models generate text, including tokenization, embeddings, attention, and next-token prediction, and describe the practical consequences of context windows and token limits.

Distinguish foundation model capabilities from fine-tuning and retrieval approaches, and classify at least ten concrete use cases (summarization, extraction, classification, chat, code assistance) by the approach they require.

Identify the main OCI Generative AI service building blocks (on-demand versus dedicated AI clusters, playground, endpoints, model catalog) and describe what each is used for in a solution.

Describe the shared responsibility boundary for generative AI on OCI and list which controls belong to the customer versus the cloud provider.

Compare the foundation model families available in the OCI Generative AI model catalog by modality, typical context length, and intended task, without relying on memorized version numbers.

Explain inference parameters (temperature, top-p, top-k, max tokens, stop sequences, frequency and presence penalties) and predict how changing each one alters output for a given prompt.

Describe how the OCI Generative AI playground and API/SDK/CLI access paths differ, and outline the request lifecycle from authentication to response.

Explain pricing and capacity concepts, including the difference between on-demand inference and provisioned dedicated capacity, and when each is appropriate.

Build a decision checklist that selects a model for a scenario based on task type, latency, context length, cost, and data residency, and apply it to five written scenarios.

Compare prompting techniques (zero-shot, few-shot, chain-of-thought, structured output instructions) and choose the right one for extraction, classification, and multi-step reasoning tasks.

Explain when fine-tuning is preferable to prompting or retrieval, and describe the fine-tuning workflow on OCI including data format requirements, training versus validation splits, and evaluation of results.

Describe how dedicated AI clusters and model endpoints are provisioned and scaled, and identify the operational limits that affect throughput planning.

Implement a working inference call against an OCI Generative AI endpoint using the Python SDK, including error handling for throttling, timeouts, and invalid input.

Explain embeddings conceptually (vector representation, dimensionality, semantic similarity) and generate embeddings for a small text corpus through the service.

Describe the retrieval-augmented generation pipeline end to end: ingestion, chunking, embedding, indexing, retrieval, prompt assembly, and grounded generation.

Compare chunking strategies (fixed size, semantic, overlap) and justify a choice for documents with tables, long policy text, and short FAQ entries.

Explain vector similarity metrics (cosine, dot product, Euclidean) and how normalization and embedding dimensionality affect retrieval results.

Describe OCI Vector Store and Oracle Database vector search capabilities, including index types, filtering with metadata, and the tradeoffs between approximate and exact search.

Design a retrieval evaluation that measures recall of relevant chunks and answer faithfulness, and define numeric acceptance thresholds before testing.

Diagnose common RAG failure modes (irrelevant chunks, context overflow, hallucinated citations, stale index) and prescribe a specific fix for each.

Apply OCI Identity and Access Management concepts to generative AI: compartments, policies, dynamic groups, and least-privilege policy statements for endpoints, models, and vector stores.

Describe data protection controls for generative AI workloads, including encryption at rest and in transit, private networking options, and how to keep sensitive data out of prompts and logs.

Explain content moderation, prompt injection, and data exfiltration risks in RAG systems and implement at least two mitigations in the existing pipeline.

Map governance requirements (audit logging, usage tracking, model versioning, human review) to concrete OCI services and configuration settings.

Design an end-to-end generative AI solution architecture on OCI that names each component, its compartment, its network path, and its scaling approach.

Justify build versus buy and managed versus self-managed choices using cost, compliance, and operational burden as explicit criteria.

Describe observability for generative AI applications, including latency, token consumption, error rates, and quality drift, and specify which metrics trigger which alerts.

Explain cost governance levers (token budgets, caching, prompt compression, right-sizing capacity) and estimate monthly cost for a stated workload.

Consolidate weak areas by re-testing every missed concept from weeks 1-7 and writing a corrected explanation for each.

Practice scenario-style questions that combine model selection, RAG design, and IAM policy under time pressure, and review each rationale.

Re-derive the key decision rules (model choice, chunking, fine-tuning versus RAG, capacity type) from memory and verify them against your notes.

Walk through the reference architecture aloud as if presenting it, covering security, cost, and failure handling.