Generative AI Design Patterns. Solutions to Common Challenges When Building GenAI Agents and Applications Valliappa Lakshmanan, Hannes Hapke

Dostępność produktu

Produkt nie został jeszcze oceniony pod kątem ułatwień dostępu lub nie podano żadnych informacji o ułatwieniach dostępu lub są one niewystarczające. Prawdopodobnie Wydawca/Dostawca jeszcze nie umożliwił dokonania walidacji produktu lub nie przekazał odpowiednich informacji na temat jego dostępności.

Spis treści książki

• Preface
  • Who Is This Book For?
  • Conventions Used in This Book
  • Using Code Examples
  • OReilly Online Learning
  • How to Contact Us
  • Acknowledgments
• 1. Introduction
  • GenAI Design Patterns
  • Building on Foundational Models
    • Prompt and Context
    • Using the Model Providers API
    • Using an LLM-Agnostic Framework
    • Running Your Model Locally
    • How Foundational Models Are Created
    • The Landscape of Foundational Models
  • Agentic AI
    • Autonomy
    • Characteristics of Agents
  • Fine-Grained Control
    • Logits
    • Temperature
    • Top-K Sampling
    • Nucleus Sampling
    • Beam Search
  • In-Context Learning
    • Zero-Shot Learning
    • Few-Shot Learning
  • Post-Training
    • Post-Training Methods
    • Fine-Tuning a Frontier Model
    • Fine-Tuning an Open-Weight Model
    • Considerations
  • The Organization of the Rest of the Book
• 2. Controlling Content Style
  • Pattern 1: Logits Masking
    • Problem
    • Solution
      • Step 1: Intercepting sampling
      • Step 2: Zeroing out invalid sequences
      • Step 5: Backtracking and regenerating sequences
    • Examples
      • Sequence selection
        • Zero-shot learning doesnt work
        • Using Logits Masking
      • Sequence regeneration
        • One-shot examples dont work
        • Initializing the poem
        • Generating a poem
        • Applying Logits Masking
        • Regeneration logic
        • Example output
    • Considerations
      • Alternatives
      • Extension to autocomplete
      • Caveats
    • References
  • Pattern 2: Grammar
    • Problem
    • Solution
      • Option 1: Using the grammar-constrained logits processor
        • Step 1: Create a formal grammar
        • Step 2: Create a logits processor that applies grammar
        • Step 3: Apply logits processing
      • Option 2: Using standard data format
      • Option 3: Using user-specified schema
    • Examples
      • Arithmetic expressions
      • The pipe separator
      • JSON output format
      • Extracting invoice information
    • Considerations
      • Variations
      • Alternatives
      • Caveats
    • References
  • Pattern 3: Style Transfer
    • Problem
    • Solution
      • Option 1: Few-shot learning
      • Option 2: Model fine-tuning
      • Example: Style Transfer in images
    • Considerations
    • References
  • Pattern 4: Reverse Neutralization
    • Problem
    • Solution
    • Example 1: Generating Legal Text
    • Example 2: Personal style
      • Step 1: Neutralization
      • Step 2: Dataset creation
      • Step 3: Fine-tuning
      • Step 4: Inference
    • Considerations
      • Choosing the neutral form
      • Dataset considerations
      • Other uses of neutralization
    • References
  • Pattern 5: Content Optimization
    • Problem
    • Solution
      • Step 1: Generate pairs of content from the same prompt
        • Repeated generation
        • Changing the generation settings
        • Prompt rewriting
      • Step 2: Compare the two content pieces and pick a winner
        • Variant 1: Human labeling
        • Variant 2: Using an evaluator
        • Variant 3: Choosing the item that provides a better outcome
      • Step 3: Create a training dataset
      • Step 4: Do preference tuning
    • Example
      • Zero-shot learning doesnt work
      • Step 1: Generate pairs of ads
      • Step 2: Compare two ads using LLM-as-judge
      • Step 3: Create a training dataset
      • Step 4: Do preference tuning
      • Inference
    • Considerations
      • Choosing between the variants
      • The in-distribution requirement
      • Extension to images
      • Continuous improvement
        • Using a better evaluator
        • Using diverse prompts
        • Iterative training
    • References
  • Summary
• 3. Adding Knowledge: Bass
  • Pattern 6: Basic RAG
    • Problem
    • Solution
      • Grounding
        • Priming
        • Relevance
      • Pipelines
      • Indexing
      • Retrieval
      • Generation
    • Example
    • Considerations
      • Alternatives
      • Limitations
    • References
  • Pattern 7: Semantic Indexing
    • Problem
    • Solution
      • Embeddings
      • Semantic chunking
      • Handling images and videos
      • Handling tables
      • Handling industry jargon
      • Contextual retrieval
      • Hierarchical chunking (alternative chunking techniques)
    • Example
    • Considerations
      • Alternatives
        • Domain-specific embedding models
        • Chunking versus long context windows
      • Limitations
    • References
  • Pattern 8: Indexing at Scale
    • Problem
      • Disambiguation
      • Data freshness
      • Contradictory information
      • Model lifecycle
    • Solution
      • Metadata
      • Detecting contradictory content
      • Detecting outdated content
      • Managing model lifecycle
    • Example
    • Considerations
    • References
  • Summary
• 4. Adding Knowledge: Syncopation
  • Pattern 9: Index-Aware Retrieval
    • Problem
      • Question not present in knowledge base
      • Knowledge base uses technical language that differs from user queries
      • Answer is a fine detail hidden in a chunk
      • Answer involves holistic interpretation
    • Solution
      • Component 1: Hypothetical answers
      • Component 2: Query expansion
      • Component 3: Hybrid search
      • Component 4: GraphRAG
    • Example
      • Limitations of Semantic Indexing
      • HyDE
      • Query expansion
    • Considerations
    • References
  • Pattern 10: Node Postprocessing
    • Problem
      • Similarity is not relevance
      • Irrelevant content
      • Ambiguous entities
      • Conflicting or obsolete content
      • Generic answers
    • Solution
      • Reranking
      • Hybrid search
      • Query expansion and decomposition
      • Filtering for obsolete information
      • Contextual compression
      • Disambiguation
      • Personalization and conversation history
    • Example
      • Limitations of semantic RAG
      • Impact of Node Postprocessing
    • Considerations
    • References
  • Pattern 11: Trustworthy Generation
    • Problem
    • Solution
      • Out-of-domain detection
      • Citations
        • Source-level tracking
        • Classification-based citations
        • Token-level attribution
      • Guardrails
      • Observability
      • Human Feedback
      • Corrective retrieval-augmented generation (CRAG)
      • Self-RAG
      • User interface design
    • Example
      • Technique
        • Step 1: Generate an initial response
        • Step 2: Chunk the response into smaller sections
        • Step 3: Check whether a chunk needs a citation
        • Step 4: Look up the sources
        • Step 5: Incorporate the citations into the response
        • Step 6: Add any necessary warnings or corrections
    • Considerations
      • Limitations
      • Alternatives
    • References
  • Pattern 12: Deep Search
    • Problem
      • Context window constraints
      • Query ambiguity
      • Information staleness and verification
      • Reasoning depth limitations
      • Multihop query challenges
    • Solution
      • Using foundational models
      • Iterative refinement
      • Evaluation metrics
      • Information integration
    • Example
      • Retrieval
      • Generation
      • Thinking
      • Orchestration
    • Considerations
    • References
  • Summary
• 5. Extending Model Capabilities
  • The Limits of LLM Reasoning
    • Known Capabilities
    • Unknown Capabilities
  • Pattern 13: Chain of Thought
    • Problem
      • Training data coverage
      • Multistep reasoning
      • Black-box answers
    • Solution
      • Variant 1: Zero-shot CoT
      • Variant 2: Few-shot CoT
      • Variant 3: Auto-CoT
    • Examples
      • Zero-shot CoT
      • Few-shot CoT
    • Considerations
      • Limitations
        • Data gaps
        • Nonsequential logic
      • Alternatives
        • Test-time compute in models
        • Using an agentic approach
    • References
  • Pattern 14: Tree of Thoughts (ToT)
    • Problem
    • Solution
      • Thought generation
      • Path evaluation
      • Tree search
      • Summary generation
    • Example
    • Considerations
      • Complexity
      • Alternatives
        • Reasoning models
        • Least-to-most prompting
        • Reflection
        • Wait-injection
    • References
  • Pattern 15: Adapter Tuning
    • Problem
    • Solution
      • Architecture
      • Training
      • Inference
    • Example
      • Anatomical descriptions
      • Dataset preparation
      • Adapter Tuning
      • Inference
    • Considerations
      • Variations
        • Merging models
        • Closed-weights models
      • Alternatives
    • References
  • Pattern 16: Evol-Instruct
    • Problem
    • Solution
      • SFT on open-weights models
        • PeFT
        • Inference
      • Preparing an instruction-tuning dataset
        • Step 1: Evolve the instructions
        • Step 2: Generate answers
        • Step 3: Evaluate and filter
    • Example
      • The goal: Business strategy
      • Bootstrapping initial questions
      • Evolving the initial questions
      • Generating answers
      • Evaluating examples
      • Training Gemma
    • Considerations
    • References
  • Summary
• 6. Improving Reliability
  • Pattern 17: LLM-as-Judge
    • Problem
    • Solution
      • Option 1: Prompting approach
      • Option 2: ML
      • Option 3: Fine-tuning
    • Example
      • Choosing an evaluation method
      • Scoring criteria
    • Considerations
      • Inconsistency
      • Leniency
      • Bias
      • Caveats
    • References
  • Pattern 18: Reflection
    • Problem
    • Solution
      • Maximum attempts
      • Conversational state
    • Example
      • Evaluate and critique
      • Logo designer
      • Apply Criticism
      • Regenerate
    • Considerations
      • Tradeoff: Cost versus quality
      • Getting evaluation right
    • References
  • Pattern 19: Dependency Injection
    • Problem
    • Example
      • Step 1: Generate a critique
      • Testing the first step
      • Step 2: Implement a suggestion
      • Testing the second step
    • Solution
      • Mockable steps
      • Mocking Step 1
    • Considerations
    • References
  • Pattern 20: Prompt Optimization
    • Problem
    • Solution
    • Example
      • Creating a pipeline
      • Results
      • Evaluating the blurb
      • The generated prompts
      • Optimizing on one blurb
      • The few-shot optimizer
    • Considerations
    • References
  • Summary
• 7. Enabling Agents to Take Action
  • Pattern 21: Tool Calling
    • Problem
    • Solution
      • How Tool Calling works
      • What Tool Calling opens up
      • Function calling in OpenAI
        • Step 1: Implement the flight-booking function
        • Step 2: Call the model with a tool definition
        • Step 3: Process the output and invoke functions on the client side
        • Step 4: Supply the result and call the model again
        • Step 5: Final model response
      • Tool Calling with LangGraph and MCP
        • MCP server
        • MCP client
    • Example
      • Weather tool
      • Geocoding tool
      • MCP server
      • MCP client
    • Considerations
      • Improving reliability
      • MCP limitations
      • Prompt injection
    • References
  • Pattern 22: Code Execution
    • Problem
    • Solution
    • Example
    • Considerations
    • References
  • Pattern 23: Multiagent Collaboration
    • Problem
    • Solution
      • Multiple specialized agents
      • Multiagent architectures
        • Hierarchical structures
        • Peer-to-peer networks
        • Market-based systems
        • Human-in-the-loop
      • Use cases
    • Example
      • Step 1: Setting up the workflow
      • Step 2: Assigning the topic to a writer
      • Step 3: Creating the initial draft
      • Step 4: Setting up a review panel
      • Step 5: Having the review panel discussion
      • Step 6: Summarizing the discussion and converting it into feedback
      • Step 7: Supplying feedback to the history-writing agent
      • Step 8: Generating the final content
    • Considerations
    • References
  • Summary
• 8. Addressing Constraints
  • Pattern 24: Small Language Model
    • Problem
      • GPU requirements
      • Slowness
    • Solution
      • Option 1: Distillation
        • The concept
        • The training process
        • Measuring similarity
        • Meta distillation loop
      • Option 2: Model quantization
        • The concept
        • The method
      • Option 3: Speculative decoding
        • The concept
        • Why it works
        • The method
    • Example
      • Distillation
        • Creating training data
        • Distilling Gemma
      • Quantization
    • Considerations
      • Limitations
      • Alternatives
    • References
  • Pattern 25: Prompt Caching
    • Problem
    • Solution
      • Client-side prompt caching (memoization)
        • Python library caches
        • Model provider client-side caching
        • Semantic caching
      • Server-Side prompt caching
    • Example
    • Considerations
    • References
  • Pattern 26: Inference Optimization
    • Problem
    • Solution
      • Option 1: Continuous batching
      • Option 2: Speculative Decoding
      • Option 3: Prompt compression
        • Hard prompt compression
        • Soft prompt compression
    • Example
    • References
  • Pattern 27: Degradation Testing
    • Problem
    • Solution
      • Core metrics
        • Time to first token (TTFT)
        • End-to-end request latency (EERL)
        • Tokens per second (TPS)
        • Requests per second (RPS)
      • Scalability and resilience
        • Scalability
        • Stress analysis
        • Load testing
    • Example
      • Performance-monitoring tools
    • References
  • Pattern 28: Long-Term Memory
    • Problem
    • Solution
      • Working memory
      • Episodic memory
      • Procedural memory
      • Semantic memory
    • Example
      • Mem0
      • How adding memories works
      • How searching memories works
    • Considerations
    • References
  • Summary
• 9. Setting Safeguards
  • Pattern 29: Template Generation
    • Problem
    • Solution
    • Example
      • Pregeneration loop
      • Generating a template
      • Inference
    • Considerations
    • References
  • Pattern 30: Assembled Reformat
    • Problem
    • Solution
    • Example
    • Considerations
    • References
  • Pattern 31: Self-Check
    • Problem
    • Solution
      • Requesting and processing logprobs
      • How logprobs behave
      • Low-confidence answers
      • Identifying hallucinations
    • Example
    • Considerations
    • References
  • Pattern 32: Guardrails
    • Problem
    • Solution
      • Prebuilt guardrails
      • Custom guardrails
      • Applying a set of guardrails
    • Example
      • Implementing guardrails
      • Wrapping the query engine
    • Considerations
    • References
  • Summary
• 10. Composable Agentic Workflows
  • Agentic Workflow
    • Running the Application
      • Setup
      • Copilot mode
      • Agent mode
    • What the Application Does
      • The first step of the workflow
      • Human feedback
    • System Architecture
      • Agent patterns
      • Multiagent architecture
      • Governance, monitoring, and security
      • Learning pipeline
      • Data program
    • Deployment
  • Summary
• Index

pokaż cały spis treści

ukryj recenzje