OCI Generative AI Professional Study Plan¶

Five weeks at 5-6 hours per week, assuming OCI AI Foundations level knowledge going in.

Week 1: LLM fundamentals¶

• Transformer architecture: attention, encoder and decoder variants
• Tokenization, and how tokens relate to cost and context limits
• Embeddings and what makes two texts similar in vector space
• Decoding parameters: temperature, top-k, top-p, penalties, max tokens, stop sequences
• Greedy decoding versus sampling, and the effect on reproducibility
• Lab: run the same prompt at temperature 0 and 1 in the playground and compare
• Review Notes: notes/01-llm-fundamentals.md

Week 2: Prompting and customization¶

• Zero-shot, few-shot, chain-of-thought, system prompts
• Prompt injection at a conceptual level, and why it matters for a deployed application
• The customization spectrum: prompting, RAG, T-Few and LoRA fine-tuning, full fine-tuning, pre-training
• When fine-tuning is the right answer and when it is not
• Evaluation: loss, perplexity, held-out accuracy, LLM-as-judge
• Lab: compare zero-shot, few-shot, and chain-of-thought on the same reasoning task

Week 3: The OCI Generative AI service¶

• Available models: chat and embedding, and their context limits
• On-demand inference versus dedicated AI clusters, and the billing difference
• Cluster sizing, hosting clusters versus fine-tuning clusters
• Creating a custom model and serving it through an endpoint
• Playground, API, SDK, and CLI
• Security: compartments, IAM policies, private endpoints, encryption, data isolation
• Content moderation, limits, and quotas
• OCI Generative AI Agents
• Lab: call the service from the playground and from the SDK; inspect the request parameters
• Review Notes: notes/02-oci-generative-ai-service.md

Week 4: Building RAG applications¶

• RAG end to end: ingest, chunk, embed, store, retrieve, augment, generate
• Chunking strategy: size, overlap, and structure-aware splitting
• Embedding model choice and dimensionality
• Similarity metrics: cosine, dot product, Euclidean
• Vector stores, including AI Vector Search in Autonomous Database
• LangChain: models, prompts, chains, memory, retrievers, document loaders, output parsers
• Conversation memory and its effect on context and cost
• Lab: build a small RAG pipeline; change the chunk size and observe the effect
• Review Notes: notes/03-building-llm-applications.md

Week 5: Evaluation, deployment, and review¶

• Evaluating retrieval quality separately from answer quality
• Observability, tracing, and cost management for LLM applications
• Deployment patterns on OCI and integration with other services
• Guardrails and content moderation in a production application
• Work every scenario in scenarios.md
• Two timed practice exams

Readiness check¶

• Explain what temperature, top-k, and top-p each do, and which to change for determinism
• Choose between prompting, RAG, and fine-tuning for a given requirement, and justify it
• Explain when a dedicated AI cluster is required rather than on-demand
• Describe the full RAG pipeline in order
• Explain how chunk size affects retrieval quality in both directions
• Name three causes of poor RAG answers and say how you would diagnose each
• Explain how to evaluate a RAG system, distinguishing retrieval from generation