### **Section 1: Analyze & Design Generative AI Solutions ** 1. **Understand the five capabilities of GenAI/LLMs** - LLMs excel at **text generation, classification, summarization, question answering, and translation**. 2. **Articulate the components in Gen AI Patterns** - Core components: **model architecture, data input/output, prompt design, and post-processing**. 3. **Understand the limitations of GenAI/LLMs** - Key limitations: **hallucinations, bias, context window limits, and lack of real-time data**. 4. **Understand use cases and identify Gen AI application opportunities** - Focus on **customer service, content creation, code generation, and data analysis**. 5. **Understand how to choose the appropriate model for a use case** - Consider **model size, fine-tuning needs, cost, and domain expertise**. 6. **Articulate the optimal model architecture based on a use case** - Select **Transformer-based models (e.g., LLMs) for language tasks, diffusion models for images**. 7. **Identify and apply various tools and techniques** - Tools: **AI agents, RAG (Retrieval-Augmented Generation), LangChain, and prompt engineering frameworks**. 8. **Understand security risks associated with LLMs, prompt engineering, and data** - Risks: **prompt injection, data leakage, adversarial attacks, and model poisoning**. --- ### **Section 2: Prompt Engineering ** 1. **Differentiate between zero-shot and few-shot prompting** - **Zero-shot**: No examples; **Few-shot**: Provides examples to guide responses. 2. **Design prompts based on use case** - Tailor prompts for **specific tasks (e.g., summarization, Q&A, code generation)**. 3. **Generate prompt templates** - Reusable templates for **consistency and efficiency** in prompt design. 4. **Determine the best model parameters for each GenAI prompt** - Adjust **temperature, top-k, top-p, and max tokens** for desired output quality. 5. **Describe the benefits of using prompt variables** - Variables enable **dynamic, reusable prompts** for scalability and adaptability. 6. **Describe the benefits of Prompt Lab** - **Iterative testing, collaboration, and version control** for prompts. 7. **Articulate hyperparameter tuning** - Optimize **parameters** to balance creativity, coherence, and relevance. 8. **Articulate model risks** - Risks: **overfitting, bias amplification, and unintended outputs**. --- ### **Section 3: Fine-Tuning ** 1. **Understand the difference between hard and soft prompts** - **Hard prompts**: Fixed text inputs; **Soft prompts**: Learnable embeddings for adaptability. 2. **Reconstruct prompts to reduce the cost of using GenAI models** - Use **few-shot, structured prompts** to minimize computational overhead. 3. **Plan for Data elements for application usage** - Curate **high-quality, domain-specific datasets** for fine-tuning. 4. **Articulate model quantization techniques** - Reduce model size and inference time using **quantization (e.g., 8-bit, 4-bit)**. 5. **LoRA (Low-Rank Adaptation)** - Efficient fine-tuning by **updating only a small subset of model weights**. 6. **Prepare the dataset for training** - Clean, annotate, and **split data** for training/validation/testing. 7. **Customize LLMs with InstructLab** - Use **InstructLab** to align models with human preferences. 8. **Generate synthetic data using the User Interface** - Create **synthetic datasets** for training and testing via UI tools. --- ### **Section 4: Retrieval-Augmented Generation (RAG) ** 1. **Describe embeddings in the context of GenAI** - Convert text/data into **dense vector representations** for similarity search. 2. **Generate vector embeddings utilizing models** - Use models like **BERT, Sentence-BERT, or proprietary embeddings** (e.g., `text-embedding-3-large`). 3. **Describe when to use a vector database** - Ideal for **semantic search, recommendation systems, and RAG pipelines**. 4. **Develop using libraries** - Libraries: **FAISS, Chroma, Weaviate, Pinecone** for vector storage and retrieval. --- ### **Section 5: Deployment ** 1. **Plan for a deployment based on client needs** - Align deployment with **scalability, latency, and cost requirements**. 2. **Deploy AI Assets** - Deploy **models, prompts, and datasets** as APIs or microservices. 3. **Deploy a custom model** - Use **Docker, Kubernetes, or serverless platforms** for custom model deployment. 4. **Plan out deployment of prompts for versioning** - Track and manage **prompt versions** for iterative improvements. 5. **High-level architecture for deployment options** - Options: **Cloud (AWS SageMaker, GCP Vertex), on-prem, or hybrid**. --- ### **Section 6: Integration with Model Orchestration (8%)** 1. **Integrate watsonx.ai with Other Services/Manage APIs and SDKs** - Use **REST APIs, SDKs, and webhooks** for seamless integration. 2. **Orchestrate AI Workflows** - Automate **end-to-end AI pipelines** using workflow tools (e.g., Airflow, Kubeflow). 3. **Understand real-world Integration Scenarios** - Examples: **Chatbots, recommendation engines, and automated report generation**. 4. **Develop LLM-based applications with LangChain** - Use **LangChain** for **chaining prompts, tools, and models** into cohesive applications.