Generative AI - Architectures and applications in depth by K, Mavrodimitraki & D. Papageorgiou

Foundation models are large deep learning neural networks trained on massive datasets, forming the core of generative AI applications

Retrieval Augmented Generation (RAG) helps overcome limitations of foundation models by:

• Retrieving relevant context from knowledge bases
• Augmenting prompts with domain-specific information
• Providing source attribution and traceability
• Being more cost-effective than fine-tuning

Three main approaches to customize foundation models:

• Prompt engineering
• Fine-tuning
• Continuous pre-training

Amazon Bedrock provides:

• Managed access to multiple foundation models
• Built-in knowledge base functionality
• Agent orchestration capabilities
• Vector database integration

Agents work through four key phases:

• Pre-processing
• Orchestration
• Knowledge base resource retrieval
• Post-processing

Word embeddings are crucial for LLMs by:

• Representing words as multidimensional vectors
• Capturing contextual relationships
• Enabling semantic similarity matching

Model parameters have grown significantly:

• BERT (2019): 340M parameters
• GPT-2 (2019): 1.5B parameters
• GPT-3 (2022): 175B parameters

Knowledge bases should be used when:

• Domain-specific accuracy is required
• Real-time data access is needed
• Source attribution is important
• Cost optimization is a priority