Welcome to our GenAI: Best Practices!!! For each chapter, we provide detailed Colab notebooks that you can open and run directly in Google Colab. The PDF version can be downloaded from HERE.

Contents

• 1. Preface
  • 1.1. About
    • 1.1.1. About this book
    • 1.1.2. About the authors
  • 1.2. Feedback and suggestions
• 2. Preliminary
  • 2.1. Math Preliminary
    • 2.1.1. Vector
    • 2.1.2. Norm
    • 2.1.3. Distances
  • 2.2. NLP Preliminary
    • 2.2.1. Vocabulary
    • 2.2.2. Tagging
    • 2.2.3. Lemmatization
    • 2.2.4. Tokenization
    • 2.2.5. BERT Tokenization
    • 2.2.6. Modern BERT
  • 2.3. Platform and Packages
    • 2.3.1. Google Colab
    • 2.3.2. HuggingFace
    • 2.3.3. Ollama
    • 2.3.4. langchain
• 3. Word and Sentence Embedding
  • 3.1. Traditional word embeddings
    • 3.1.1. One Hot Encoder
    • 3.1.2. CountVectorizer
    • 3.1.3. TF-IDF
  • 3.2. Static word embeddings
    • 3.2.1. Word2Vec
    • 3.2.2. GloVE
    • 3.2.3. Fast Text
  • 3.3. Contextual word embeddings
    • 3.3.1. BERT
    • 3.3.2. gte-large-en-v1.5
    • 3.3.3. bge-base-en-v1.5
• 4. Prompt Engineering
  • 4.1. Prompt
  • 4.2. Prompt Engineering
    • 4.2.1. What’s Prompt Engineering
    • 4.2.2. Key Elements of a Prompt
  • 4.3. Advanced Prompt Engineering
    • 4.3.1. Role Assignment
    • 4.3.2. Contextual Setup
    • 4.3.3. Explicit Instructions
    • 4.3.4. Chain of Thought (CoT) Prompting
    • 4.3.5. Few-Shot Prompting
    • 4.3.6. Iterative Prompting
    • 4.3.7. Instructional Chaining
    • 4.3.8. Use Constraints
    • 4.3.9. Creative Prompting
    • 4.3.10. Feedback Incorporation
    • 4.3.11. Scenario-Based Prompts
    • 4.3.12. Multimodal Prompting
• 5. Retrieval-Augmented Generation
  • 5.1. Overview
  • 5.2. Naive RAG
    • 5.2.1. Indexing
    • 5.2.2. Retrieval
    • 5.2.3. Generation
  • 5.3. Self-RAG
    • 5.3.1. Load Models
    • 5.3.2. Create Index
    • 5.3.3. Retrieval
    • 5.3.4. Generate
    • 5.3.5. Utilities
    • 5.3.6. Graph
    • 5.3.7. Test
  • 5.4. Corrective RAG
    • 5.4.1. Load Models
    • 5.4.2. Create Index
    • 5.4.3. Retrieval
    • 5.4.4. Generate
    • 5.4.5. Utilities
    • 5.4.6. Graph
    • 5.4.7. Test
  • 5.5. Adaptive RAG
    • 5.5.1. Load Models
    • 5.5.2. Create Index
    • 5.5.3. Retrieval
    • 5.5.4. Generate
    • 5.5.5. Utilities
    • 5.5.6. Graph
    • 5.5.7. Test
  • 5.6. Agentic RAG
    • 5.6.1. Load Models
    • 5.6.2. Create Index
    • 5.6.3. Retrieval Tool
    • 5.6.4. Generate
    • 5.6.5. Agent State
    • 5.6.6. Graph
    • 5.6.7. Test
  • 5.7. Advanced Topics
    • 5.7.1. Multi-agent Systems
    • 5.7.2. Graph RAG
• 6. Fine Tuning
  • 6.1. Cutting-Edge Strategies for LLM Fine-Tuning
    • 6.1.1. LoRA (Low-Rank Adaptation)
    • 6.1.2. QLoRA (Quantized Low-Rank Adaptation)
    • 6.1.3. PEFT (Parameter-Efficient Fine-Tuning)
    • 6.1.4. SFT (Supervised Fine-Tuning)
    • 6.1.5. RLHF (Reinforcement Learning from Human Feedback)
    • 6.1.6. Summary Table
  • 6.2. Key Early Fine-Tuning Methods
    • 6.2.1. Full Fine-Tuning
    • 6.2.2. Feature-Based Approach
    • 6.2.3. Layer-Specific Fine-Tuning
    • 6.2.4. Task-Adaptive Pre-training
  • 6.3. Embedding Model Fine-Tuning
    • 6.3.1. Prepare Dataset
    • 6.3.2. Import and Evaluate Pretrained Baseline Model
    • 6.3.3. Loss Function with Matryoshka Representation
    • 6.3.4. Fine-tune Embedding Model
    • 6.3.5. Evaluate Fine-tuned Model
    • 6.3.6. Results Comparison
  • 6.4. LLM Fine-Tuning
    • 6.4.1. Load Dataset and Pretrained Model
    • 6.4.2. Fine-tuning Configuration
    • 6.4.3. Fine-tune model
• 7. Pre-training
  • 7.1. Transformer Architecture
    • 7.1.1. Attention Is All You Need
    • 7.1.2. Encoder-Decoder
    • 7.1.3. Positional Encoding
    • 7.1.4. Embedding Matrix
    • 7.1.5. Attention Mechanism
    • 7.1.6. Layer Normalization
    • 7.1.7. Residual Connections
    • 7.1.8. Feed-Forward Networks
    • 7.1.9. Label Smoothing
    • 7.1.10. Softmax and Temperature
    • 7.1.11. Unembedding Matrix
    • 7.1.12. Decoding
  • 7.2. Modern Transformer Techniques
    • 7.2.1. KV Cache
    • 7.2.2. Multi-Query Attention
    • 7.2.3. Grouped-Query Attention
    • 7.2.4. Flash Attention
    • 7.2.5. Mixture of Experts (MoE)
  • 7.3. Case Studies: DeepSeek-V3
    • 7.3.1. Architecture:
• 8. LLM Evaluation Metrics
  • 8.1. Statistical Scorers (Traditional Metrics)
  • 8.2. Model-Based Scorers (Learned Metrics)
  • 8.3. Human-Centric Evaluations (Augmenting Metrics)
  • 8.4. GEval with DeepEval
    • 8.4.1. G-Eval Algorithm
    • 8.4.2. G-Eval with DeepEval
• 9. LLM Guardrails
  • 9.1. LLM Risks
  • 9.2. Overview of Jailbreak Techniques
  • 9.3. Introduction to Guardrails
  • 9.4. Overview of Guardrails Tools
  • 9.5. NeMo Guardrails
    • 9.5.1. Yaml Config
    • 9.5.2. Rails
    • 9.5.3. Test
  • 9.6. Llama Guard3
    • 9.6.1. Load Models
    • 9.6.2. Define Safety Check
    • 9.6.3. Tests
• 10. LLM Deployments
  • 10.1. Prototype Deployment
    • 10.1.1. Gradio vs. Streamlit
    • 10.1.2. Deployment with Gradio
  • 10.2. Production Deployment
• 11. Main Reference