Prompt Engineering for LLMs: The Art and Science of Building Large Language Model–Based Applications

CopyrightTable of ContentsPrefaceWho Is This Book For?What You Will LearnConventions Used in This BookO’Reilly Online LearningHow to Contact UsAcknowledgmentsFrom JohnFrom AlbertPart I. FoundationsChapter 1. Introduction to Prompt EngineeringLLMs Are MagicLanguage Models: How Did We Get Here?Early Language ModelsGPT Enters the ScenePrompt EngineeringConclusionChapter 2. Understanding LLMsWhat Are LLMs?Completing a DocumentHuman Thought Versus LLM ProcessingHallucinationsHow LLMs See the WorldDifference 1: LLMs Use Deterministic TokenizersDifference 2: LLMs Can’t Slow Down and Examine LettersDifference 3: LLMs See Text DifferentlyCounting TokensOne Token at a TimeAuto-Regressive ModelsPatterns and RepetitionsTemperature and ProbabilitiesThe Transformer ArchitectureConclusionChapter 3. Moving to ChatReinforcement Learning from Human FeedbackThe Process of Building an RLHF ModelKeeping LLMs HonestAvoiding Idiosyncratic BehaviorRLHF Packs a Lot of Bang for the BuckBeware of the Alignment TaxMoving from Instruct to ChatInstruct ModelsChat ModelsThe Changing APIChat Completion APIComparing Chat with CompletionMoving Beyond Chat to ToolsPrompt Engineering as PlaywritingConclusionChapter 4. Designing LLM ApplicationsThe Anatomy of the LoopThe User’s ProblemConverting the User’s Problem to the Model DomainUsing the LLM to Complete the PromptTransforming Back to User DomainZooming In to the Feedforward PassBuilding the Basic Feedforward PassExploring the Complexity of the LoopEvaluating LLM Application QualityOffline EvaluationOnline EvaluationConclusionPart II. Core TechniquesChapter 5. Prompt ContentSources of ContentStatic ContentClarifying Your QuestionFew-Shot PromptingDynamic ContentFinding Dynamic ContextRetrieval-Augmented GenerationSummarizationConclusionChapter 6. Assembling the PromptAnatomy of the Ideal PromptWhat Kind of Document?The Advice ConversationThe Analytic ReportThe Structured DocumentFormatting SnippetsMore on InertnessFormatting Few-Shot ExamplesElastic SnippetsRelationships Among Prompt ElementsPositionImportanceDependencyPutting It All TogetherConclusionChapter 7. Taming the ModelAnatomy of the Ideal CompletionThe PreambleRecognizable Start and EndPostscriptBeyond the Text: LogprobsHow Good Is the Completion?LLMs for ClassificationCritical Points in the PromptChoosing the ModelConclusionPart III. An Expert of the CraftChapter 8. Conversational AgencyTool UsageLLMs Trained for Tool UsageGuidelines for Tool DefinitionsReasoningChain of ThoughtReAct: Iterative Reasoning and ActionBeyond ReActContext for Task-Based InteractionsSources for ContextSelecting and Organizing ContextBuilding a Conversational AgentManaging ConversationsUser ExperienceConclusionChapter 9. LLM WorkflowsWould a Conversational Agent Suffice?Basic LLM WorkflowsTasksAssembling the WorkflowExample Workflow: Shopify Plug-in MarketingAdvanced LLM WorkflowsAllowing an LLM Agent to Drive the WorkflowStateful Task AgentsRoles and DelegationConclusionChapter 10. Evaluating LLM ApplicationsWhat Are We Even Testing?Offline EvaluationExample SuitesFinding SamplesEvaluating SolutionsSOMA AssessmentOnline EvaluationA/B TestingMetricsConclusionChapter 11. Looking AheadMultimodalityUser Experience and User InterfaceIntelligenceConclusionIndexAbout the AuthorsColophon