liveBookmanning.com

What you will learn in The Art of AI Product Development

Delivering business value

by Dr. Janna Lipenkova

Read it on liveBook Get the book at manning.com

Chapter by chapter

Concepts explained in this book

• language modelA machine-learning model trained on large text corpora to understand and generate natural language.
• AI agentAn AI agent is an artificial intelligence system that perceives relevant conditions and takes actions toward goals.
• continuous improvementSelf-refinement is an iterative technique in which a model critiques and improves its own generated outputs.
• human-in-the-loopHuman evaluation is the assessment of model outputs by people using qualitative or task-specific judgments.
• fine-tuningFine-tuning is adapting a pretrained model by training it further on task- or domain-specific data.
• guardrailA guardrail is a constraint or control mechanism that limits an AI system's behavior to reduce harmful, unsafe, or noncompliant outputs.
• similarity searchSearch is the process of locating information relevant to a query from one or more sources.
• mental modelA mental model is an internal representation of relevant entities, relationships, and processes that supports reasoning about a domain.
• hallucinationAn AI hallucination is a generated output that presents false, unsupported, or fabricated information as if it were reliable.
• feedbackUser feedback is information provided by users about their needs, experiences, problems, preferences, or reactions to a product.
• onboardingOnboarding is the initial process of introducing users to a product’s capabilities, benefits, limitations, and basic operation.
• trust calibrationTrust calibration is the process of aligning a user's degree of reliance on an AI system with the system's actual capabilities and limitations.
• personalizationPersonalization is the adaptation of product content, recommendations, messages, or experiences to an individual user's behavior, preferences, or characteristics.
• prompt engineeringPrompt engineering is the deliberate design and refinement of instructions provided to a generative model to elicit desired outputs.
• augmentationAugmentation is the use of technology to enhance a person's capabilities or performance without fully replacing their involvement.
• retrieval-augmented generationAn architecture that retrieves relevant external information and supplies it to a generative model as context for producing an output.
• algorithmic aversionAlgorithmic aversion is a tendency to distrust or reject algorithmic or AI-generated judgments, often after observing errors or negative outcomes.
• automationAutomation is the use of software or machines to perform tasks with reduced direct human intervention.
• AI footprintAn AI footprint is an interface representation that lets users trace the steps or intermediate process connecting an AI prompt to its result.
• AI functionality explanationAI functionality explanation is communication about an AI system’s capabilities, processes, limitations, and behavior that helps users form accurate expectations.
• AI signalingAI signaling is the deliberate communication that artificial intelligence is processing information or generating an output.
• AI user experienceAI user experience is the quality and character of interactions with a product whose behavior or outputs depend on artificial intelligence.
• automation biasAutomation bias is the tendency to favor or accept automated outputs without sufficient independent verification.
• citationA source citation is a link or reference identifying information sources used to produce an AI output.
• cognitive forcing functionA cognitive forcing function is an intentional interaction disruption that prompts a person to critically evaluate an automated output before accepting it.
• failure modeA failure mode is a distinct way an AI system can produce an incorrect, harmful, or otherwise undesired result.
• in-context explanationAn in-context explanation is information presented during or immediately before an interaction or output to clarify its cause and meaning.
• uncertainty indicatorAn uncertainty indicator is a visual or linguistic signal that marks content or conclusions whose reliability is limited or unclear.
• training datasetThe collection of examples used to fit the parameters of a machine-learning model.
• embeddingA numeric vector representation that encodes the semantic or structural properties of an item for computational comparison.
• machine learningMachine learning is a field of artificial intelligence in which systems learn patterns from data to make predictions or decisions.
• valueValue is the beneficial outcome a product provides to customers or a business, such as improved access, saved time, or increased quality.
• interpretabilityInterpretability is the degree to which humans can understand how a model produces its outputs or decisions.
• user experienceUser experience is the design of how people access, understand, control, and respond to a product's functionality and outputs.
• feasibilityThe degree to which a task or product capability can be implemented effectively with available technology, expertise, time, and resources.
• reflectionIntrospection is examination of a system's own outputs, reasoning, or internal assessments.
• natural language interfaceA natural language interface is an interaction boundary through which users or software systems communicate with another system using human language.
• validationData validation is the process of checking data for correctness, consistency, completeness, and conformity to specified conditions.
• toolA tool is a software capability that an AI system can invoke to retrieve information, communicate, execute operations, or affect an external system.
• biasBias is a systematic skew in data, model behavior, or outcomes that produces unequal or inaccurate results.
• predictive analyticsThe analysis of data with statistical or machine-learning methods to estimate future or unknown outcomes.
• convenienceConvenience is a product benefit that reduces friction, effort, or tedious steps in a user's journey.
• emotional benefitAn emotional benefit is value produced by an interaction that creates personal connection, resonance, or positive feeling for the user.
• human-AI collaborationHuman-AI collaboration is a working arrangement in which people and AI systems contribute complementary capabilities to accomplish tasks or make decisions.
• everboardingEverboarding is ongoing product education provided after initial user onboarding as the product and user needs evolve.
• experimentationExperimentation is the deliberate variation and evaluation of product or system conditions to learn what works.
• transparencyTransparency is the clear communication of an AI system's capabilities, limitations, behavior, and relevant decision factors.
• securityCybersecurity is the practice of protecting computer systems, networks, software, and data from unauthorized access, disruption, alteration, or attack.
• parameter-efficient fine-tuningParameter-efficient fine-tuning is a fine-tuning method that adapts a model by updating a small subset or additional parameterization rather than all model parameters.
• trustUser trust is a user's willingness to rely on a product or system based on perceived reliability, integrity, competence, transparency, and responsible behavior.
• pipelineA workflow is an ordered, often iterative sequence of tasks and operations that produces a process outcome.
• sensitive informationInformation owned or controlled by an organization and not generally available for unrestricted public use.
• promptA prompt is an instruction or contextual input supplied to a generative model to elicit an output.
• chain-of-thought promptingChain-of-thought is a prompting technique that elicits an intermediate sequence of reasoning steps before a model produces a final answer.
• vectorizationAn embedding model is a machine-learning model that maps text or other inputs to numeric representations in an embedding space.
• generative modelingArtificial intelligence that generates new content such as text, images, audio, or code from learned patterns.
• demonstrationA demonstration is an input-output example supplied in a prompt to illustrate the task and desired response format.
• recommender systemA recommendation system is a computational system that generates personalized suggestions by estimating which items or content may interest a user.
• structured dataStructured data is information organized according to a defined schema that supports systematic querying and processing.
• privacyPrivacy is the protection of individuals' or organizations' sensitive information from unauthorized collection, disclosure, or use.
• labeled dataLabeled data is data paired with target annotations, classifications, or outcome values for supervised learning or evaluation.
• multimodalA multimodal model processes or generates information across multiple modalities such as text, images, audio, or video.
• regulationA regulation is an authoritative rule or requirement imposed by a governing authority on organizations, products, or activities.
• classifierClassification is a supervised learning technique that assigns inputs to predefined categories or labels.
• confidence intervalA confidence interval is a statistical range constructed from sample data to estimate an unknown population parameter with a stated confidence level.
• clusteringClustering is an unsupervised learning technique that groups data points according to similarities in their characteristics.
• decompositionTask decomposition is the process of breaking a complex objective into smaller subtasks that can be planned or executed separately.
• text chunkSemantic chunking divides text into segments according to meaningful topics or semantic boundaries.
• labelGround truth is a reference set of inputs and desired outputs used to compare and assess actual system outputs.
• callFunction calling is a model interaction capability in which a language model produces a structured request to invoke a specified software function or tool.
• behavioral dataData recording users' actions and interactions with a product, such as listening, skipping, liking, and adding items to playlists.
• prototypingA prototype is an early or simplified version of a product or feature that demonstrates its behavior for learning and evaluation.
• preference alignmentPreference alignment is the adjustment of a model's behavior to better conform to stated preferences about the quality or acceptability of its outputs.
• dataInformation used as input for training, fine-tuning, evaluating, or operating an AI product.
• lexical searchLexical search retrieves and ranks documents by matching words or phrases in their textual forms.
• planningOnline planning is the dynamic generation and revision of an action plan while an agent is executing a task.
• evaluationEvaluation is the systematic assessment of a search system or model against specified criteria, data, or user outcomes.
• AI-specific governanceAI governance is the set of policies, controls, and practices for identifying and managing risks arising from AI systems.
• symbolic AIA rule-based approach to artificial intelligence derives behavior or conclusions from explicitly specified rules rather than patterns learned from data.
• performanceEfficiency is the achievement of an intended result with minimal expenditure of time, effort, or resources.
• supervised learningSupervised learning is machine learning in which a model learns a mapping from inputs to explicitly labeled targets.
• rule of logicA logical rule is an explicit conditional statement that derives or permits an outcome when specified conditions hold.
• collaborative filteringA collaborative-filtering approach that compares users and recommends items based on the preferences of similar users.
• subject matter expertA domain expert is a person with substantial knowledge of a particular problem area, industry, workflow, or user population.
• LangChainLangChain is a software framework for developing applications that compose language models with prompts, retrieved context, memory, tools, and other components.
• key performance indicatorA key performance indicator is a quantitative measure used to assess progress toward a defined business, operational, or product objective.
• meanLatency is the elapsed time between an input or request and the corresponding model response or system result.
• zero-shot promptingZero-shot prompting is prompting a model to perform a task without providing example input-output pairs in the prompt.
• market positioningMarket positioning is the deliberate presentation of a product or business relative to customer needs and competing offerings.
• search spaceA solution space is the set of product components and design choices through which value is created and risks are managed.
• data minimization strategyData minimization is the practice of collecting and using no more personal or sensitive data than is necessary for a defined purpose.
• multi-agent systemA multi-agent system is an AI system in which multiple agents perform specialized tasks and coordinate toward a shared goal.
• context windowThe maximum amount of tokenized input and retained conversational or task information that a language model can process in one operation.
• pretrainingPretraining is the initial training of a model on broad data before adapting it to specific tasks or domains.
• benchmarkA benchmark is a standardized dataset, task, or evaluation procedure used to compare model capabilities or performance.
• few-shot learningFew-shot prompting is prompting a model with a small number of input-output examples to guide its behavior on a new input.
• automation and productivityAutomation and productivity is a benefit category in which AI performs or accelerates work to reduce effort, time, or operating cost.
• inspiration and innovationInspiration and innovation is a benefit category in which AI supports creativity, idea generation, experimentation, and the creation of new products or business models.
• opportunity treeAn opportunity tree is a framework that organizes potential AI opportunities by the types of benefits they could provide.
• data integrationData integration is the combining and coordinating of data from multiple sources, formats, or repositories so it can be processed consistently.
• cold start problemThe cold start problem is the challenge of making useful machine-learning or recommendation decisions when little or no relevant data is initially available.
• constraintA prompt constraint is an instruction that restricts the content, length, vocabulary, or other properties of a model's output.
• open-source modelAn open source model is a model whose source artifacts and usage rights permit users to inspect, modify, deploy, or otherwise use it under applicable licensing terms.
• mean reciprocal rankMean reciprocal rank is the average reciprocal of the rank at which the first relevant result appears in a ranked result list.
• conversational aiA virtual assistant is software that interacts with users conversationally to provide information or perform requested tasks.
• data qualityData quality is the degree to which data is accurate, complete, consistent, timely, and fit for its intended purpose.
• time series analysisTime series analysis is the study of observations ordered over time to identify trends, seasonal patterns, anomalies, and other temporal behavior.
• target audienceUser segmentation is the process of dividing users into groups based on shared characteristics, behavior, or needs.
• user engagementA user engagement metric is a numerical measure of how users interact with a product or its content.
• mlopsMachine learning operations is the set of engineering practices for developing, deploying, monitoring, maintaining, and governing machine-learning systems.
• small language modelA small language model is a compact language model with relatively few parameters, designed to provide efficient task-specific language processing and generation.
• testTesting is the systematic examination of a product, feature, or model to assess its behavior and quality.
• evaluatorLLM-as-a-Judge is an evaluation technique in which a large language model assesses generated outputs against specified criteria.
• uncertaintyUncertainty is the lack of complete confidence about the value, state, or outcome represented by an AI system.
• catastrophic forgettingCatastrophic forgetting is the loss of previously learned knowledge when a model is trained on new tasks or data.
• inputA feature is an input variable representing a measurable property of an example supplied to a machine-learning model.
• inference-time compute scalingInference-time scaling is the allocation of additional computation or generated reasoning during inference to improve a model's output.
• model theftThe unauthorized replication or extraction of a model's capabilities through repeated queries and analysis of its outputs.
• co-creation mindsetCo-creation is a collaborative pattern in which users and a product team jointly contribute effort, knowledge, or feedback to create or improve a product.
• assisted intelligenceAssisted intelligence is an AI application mode in which AI supports and enhances human decision making without making decisions independently.
• augmented intelligenceAugmented intelligence is an AI application mode in which AI combines with human expertise by offering suggestions and actively participating in decision making.
• automation levelAn automation level is a classification of how much of a task an AI application performs independently versus with human participation.
• common senseCommon-sense reasoning is the application of ordinary background knowledge and practical assumptions to interpret situations and make judgments.
• intuitionIntuition is the ability to reach a judgment or sense that something is wrong through experience-based impressions rather than explicit analytical reasoning.
• moral judgmentMoral and ethical judgment is the evaluation of choices according to values, principles, obligations, and potential effects on others.
• desirabilityDesirability is the degree to which intended users want a product capability and perceive it as beneficial, acceptable, and appropriate for their needs.
• ai opportunity assessmentAn AI opportunity assessment is a structured evaluation of candidate tasks for potential AI value, feasibility, and user desirability.
• ai projectA project that develops a product or solution incorporating artificial intelligence.
• click-through rateClick-through rate is the proportion of users or impressions that result in a click on a presented link, recommendation, or control.
• generative interfaceA generative interface dynamically creates or adapts its content, design, and interactions to a user's current needs and context.
• learning signalA learning signal is information that indicates to a model what output or behavior it should learn.
• opportunity spaceThe set of potential opportunities in which artificial intelligence could address problems or create value.
• ai add-onAn AI add-on is an artificial-intelligence capability added to an established product rather than serving as its primary foundation.
• AI integrationAI integration is the incorporation of artificial intelligence capabilities into an existing or newly developed digital product.
• brownfield environmentA brownfield scenario is a product-development situation in which new capabilities are integrated into an established product, system, or user base.
• core value driverA core value driver is the primary product capability or mechanism responsible for creating the product's customer value.
• greenfield environmentA greenfield scenario is a product-development situation in which a new product is built without depending substantially on an existing product foundation.
• internal enablerAn internal enabler is a capability that improves an organization's internal product processes or operations without necessarily adding a visible user-facing feature.
• feature engineeringFeature engineering is the process of creating, selecting, transforming, and refining input features to improve a machine-learning model's results.
• training data biasData bias is systematic skew or incompleteness in data that can produce inaccurate or unequal model outputs.
• recallRecall is the proportion of actual positive cases that a classifier correctly identifies.
• dogfoodingDogfooding is the practice of building or using a product to address the creators' own needs as an initial form of product validation.
• hypothesisA hypothesis is a provisional, testable proposition about a product change, user behavior, or expected outcome.
• internal crowdsourcingInternal crowdsourcing is the collection of ideas or contributions from a broad group of people within an organization.
• minimal productA minimum viable product is a product version containing enough functionality to deliver value and test assumptions with real users.
• product senseProduct sense is the ability to recognize valuable product opportunities and make sound product decisions based on user, market, and product understanding.
• prompt suggestionA prompt suggestion is a proposed example or partial instruction that helps a user formulate an input for an AI system.
• technology-first thinkingTechnology-first thinking is an opportunity-discovery approach that begins by considering what newly available technologies make possible.
• BM25BM25 is a lexical ranking function that scores documents using term frequency, document frequency, and document-length normalization.
• contextual retrievalContextual retrieval enriches retrieved text segments with surrounding document information before embedding and indexing.
• hierarchical chunkingHierarchical chunking divides a document into nested structural levels such as chapters, sections, and smaller chunks.
• metadataMetadata is structured descriptive information about a document, such as its author, date, type, or keywords.
• rerankingReranking is the reevaluation and reordering of an initially retrieved set of documents using a more detailed relevance model.
• MiniLMMiniLM is a compact transformer-based language-model family designed to provide efficient language understanding and representation.
• false positiveA false positive is an instance incorrectly assigned to a positive class or category.
• dashboardA dashboard is a user interface that presents data through organized tables, visualizations, and controls.
• relevanceA metric measuring how directly a generated answer addresses the user's query or intent.
• privacy-by-designPrivacy-by-design is an approach that embeds privacy protections into the architecture and operation of systems from the outset.
• ai engineerAn AI engineer is a technical practitioner who selects, integrates, operates, and evolves artificial-intelligence models and supporting systems.
• insecure output handlingThe unsafe acceptance, interpretation, or execution of AI-generated outputs by people or other software.
• prompt injectionAn attack that manipulates a generative model by inserting instructions intended to override or redirect its intended behavior.
• versioningVersion control is a system for recording changes to artifacts so that versions can be compared, managed, and restored.
• data modalityA modality is a distinct type or form of data that an AI system can process or generate.
• poisoning attackAn attack that deliberately corrupts or manipulates data used to train or operate a machine-learning system in order to compromise its behavior or outputs.
• personal dataPersonal information is information that relates to or can identify an individual.
• AI frameworkA software framework that provides reusable abstractions and components for developing artificial intelligence applications.
• domain knowledgeDomain knowledge is specialized understanding of the concepts, practices, workflows, and constraints of a particular field or industry.
• predictive analytics platformA predictive analytics platform is a software system that analyzes data to generate forecasts or other predictions for decision-making.
• flow engineeringFlow engineering is the deliberate design of an explicit sequence of operations for an AI workflow rather than relying entirely on autonomous planning.
• chunkingFixed-size chunking is the division of documents into segments containing a predetermined number of tokens or other units.
• cosine distanceCosine distance measures the difference in orientation between two vectors and is commonly used to compare embeddings.
• distributional hypothesisDistributional similarity is the principle that words or texts occurring in similar contexts tend to have related meanings.
• vector databaseA vector database is a database optimized for storing numeric vector representations and retrieving items by vector similarity.
• schemaA specification of the entity types, relationship types, and structural rules in a graph.
• autoencodingAutoencoding is a training objective in which a model reconstructs corrupted inputs from surrounding context.
• autoregressive modelAn autoregressive language model generates each token by predicting it from the preceding context.
• language modelingLanguage modeling is the task of predicting or generating text units from surrounding or preceding textual context.
• sequence-to-sequence transductionSequence-to-sequence transduction maps an input sequence to an output sequence in the same or a different representation framework.
• templateA prompt template is a reusable structured text pattern containing placeholders or guidance for constructing model prompts.
• unstructured dataUnstructured data is information without a consistent predefined tabular or schema-based organization, such as free text or images.
• ambiguityAmbiguity or misinterpretation is an AI failure caused by unclear inputs or edge cases being understood incorrectly.
• disruptionAn operational disruption is a temporary interruption, slowdown, or added workload in a business process caused by system behavior.
• error rateError rate is the proportion or frequency of system outputs that are incorrect within a specified set of cases.
• escalation pathAn escalation pathway is a fallback process that transfers unresolved cases from an AI system to human personnel.
• false negativeA false negative is an instance that should be assigned to a positive class or category but is not identified as such.
• financial costA financial cost is a monetary loss or expense resulting from an AI system's incorrect output or behavior.
• failureAn AI mistake is an incorrect, misleading, unsupported, or otherwise undesirable output or behavior produced by an AI system.
• reputation riskReputational risk is the possibility that visible AI errors will damage an organization's trustworthiness or credibility.
• overrideA manual override is a user-provided change that supersedes or adjusts an automated system output.
• precisionPrecision is the proportion of retrieved search results that are relevant to the query.
• data annotationData annotation is the process of adding labels or other structured annotations to raw data.
• fusion processA process that merges overlapping retrieved information into a coherent representation while removing redundancy.
• GraphRAGA retrieval-augmented generation approach that uses graph-structured knowledge and entity relationships to retrieve contextual information.
• knowledge graphAn organized graph of entities and explicitly represented relationships that can be queried for facts or context.
• multi-turn retrievalA retrieval technique that performs successive search rounds to obtain additional context or fill gaps in earlier results.
• query expansionA retrieval technique that adds terms, variants, or related subqueries to an original search query to improve retrieval results.
• intent classificationThe categorization of a search query into a type or intent that determines how it is processed or retrieved.
• query rewritingA retrieval technique that rewrites an input query into a form better suited to searching a target information collection.
• business-to-businessA business-to-business context is one in which products or services are provided by one business to another business.
• business-to-consumerA business-to-consumer context is one in which products or services are provided directly to individual consumers.
• opportunityAn opportunity is a customer-oriented need, pain point, desire, or strategic possibility that can create business value.
• preprocessingPreprocessing is the transformation and preparation of raw data before it is supplied to a machine-learning model.
• mathematical representationA numerical modality is data represented as numeric values that computational models can accept as input.
• categorical encodingOne-hot encoding represents each item in a categorical set as a vector with one active element and all other elements zero.
• artificial intelligenceArtificial intelligence is the field of computational methods that perform tasks associated with human intelligence, such as reasoning, prediction, and language understanding.
• product management agentA product management agent is an AI agent configured to perform product-management tasks such as discovery, prioritization, or roadmap planning.
• encryptionThe transformation of data into a protected representation that requires authorized decryption to read.
• LangGraphA framework for building stateful AI agents and graph-structured workflows with support for complex interactions.
• LlamaIndexAn open-source software framework for connecting language-model applications with external data and computational capabilities.
• MemGPTAn operating-system-inspired architecture and system for managing language-model memory across context and external storage.
• working memoryAgent memory held within the model's current prompt context and available for immediate processing.
• stateless systemA system property in which each interaction is processed without retaining state from previous interactions.
• user interfaceA graphical interface is a user interface composed of visual controls and elements through which users perform structured interactions.
• data exfiltrationExfiltration is the unauthorized extraction and transfer of data from a system.
• data-driven organizationA data-driven organization bases decisions and actions substantially on large-scale evidence from its operations, stakeholders, and market context.
• data driftDomain shift is a change between the data distribution encountered during model operation and the distribution represented by its training data.
• data aggregationData aggregation is the process of combining data from multiple sources or records into summarized or unified forms.
• data cleaningData cleaning is the detection and correction or removal of inaccurate, inconsistent, duplicate, missing, or otherwise problematic data.
• contextPrompt context is external information supplied to a model to influence its response, including prior conversation or a specified role.
• instructionA prompt instruction is the task specification that tells a model what operation to perform.
• output formatAn output format is a specification of the structure or representation that a model-generated result must follow.
• prompt componentA prompt component is a modular element of a prompt that provides information, instructions, examples, variables, formatting requirements, or output restrictions.
• job lossJob displacement is the loss or substantial reduction of a person's work because tasks or roles are replaced or reduced by another process or technology.
• LLMOpsLLMOps is the set of operational practices and automation specialized for deploying, monitoring, maintaining, and updating large language model systems.
• adaptabilityAdaptability is the capability of a system to detect changing conditions and adjust its behavior in response to them.
• autonomous enterpriseAn autonomous enterprise is a business organization whose functions and operations are coordinated and performed with minimal human intervention through integrated AI and other technologies.
• compound AI systemA compound AI system is an application architecture that combines multiple models with tools, databases, or other computational components.
• intelligence componentThe intelligence component is the part of an AI product that transforms inputs into meaningful outputs through models and related computational mechanisms.
• self-optimizationSelf-optimization is the autonomous adjustment of a system's processes or components over time to improve defined performance objectives.
• confidenceConfidence is a degree of belief or certainty in the validity of a conclusion or expected outcome.
• differential privacyA privacy-preserving framework that limits how much the inclusion or removal of one individual's data affects an analysis or model output.
• HIPAAThe Health Insurance Portability and Accountability Act is a United States law establishing privacy and security requirements for protected health information.
• accountabilityAccountability is the assignment of responsibility and answerability for an AI system's development, operation, decisions, and impacts.
• qualitative assessmentQualitative evaluation assesses a system through descriptive observations, judgments, and feedback rather than primarily numerical measurements.
• quantitative evaluationQuantitative evaluation assesses a system using numerical metrics computed against defined inputs and desired outputs.
• real-world monitoringReal-world monitoring is the ongoing observation of a deployed system's behavior and outcomes under actual user and data conditions.
• common token biasCommon token bias is a tendency for a language model to produce more probable, frequently observed tokens or labels more often than rare ones.
• majority label biasMajority label bias is a tendency for a model to favor labels that occur more frequently among the examples in a prompt.
• recency biasRecency bias is a tendency for a model to replicate outputs associated with examples appearing later in a prompt.
• business valueBusiness impact is the effect that a product, feature, or initiative has on an organization's outcomes, such as revenue, cost, productivity, or risk.
• return on investmentReturn on investment is the financial or economic gain produced by an investment relative to its cost.
• cross-functional collaborationCross-functional collaboration is coordinated work among people from different professional or functional roles toward shared product goals.
• intelligenceIntelligence is the capability of an AI system to infer, generate, decide, or perform tasks that require adaptive reasoning.
• anomalyAn outlier is a data point whose behavior or value is anomalous relative to the dominant patterns in a dataset.
• ready, fire, aimReady, fire, aim is a fast development pattern that favors early implementation and learning through iteration over extensive upfront planning.
• topic classificationTopic classification is assigning documents or other inputs to predefined topical categories.
• llm-as-a-serviceLLM-as-a-Service is a hosted service model that provides access to large language models through provider-managed interfaces or APIs.
• unsupervised learningUnsupervised learning is machine learning in which a model learns patterns from data without explicitly provided target labels.
• fairnessFairness is the property of treating relevant people or groups equitably in an AI system's behavior and outcomes.
• probability theoryA probability is a numerical value between zero and one representing the likelihood of an event or outcome.
• labor distributionLabor distribution is the allocation of task work and responsibility between human participants and AI systems.
• ai interfaceAn agent interface is a user-facing interaction pattern through which people observe, direct, or collaborate with an AI agent.
• knowledge cutoffA knowledge cutoff is the latest point in time represented by a model's training data and learned knowledge.
• adoption rateAdoption is the extent to which intended users begin and continue using a product or feature.
• collaborative agentA collaborative agent is an AI agent that works interactively with a human by requesting or incorporating human assistance during task execution.
• variable rewardA variable reward is a reinforcement whose timing or form changes across interactions, encouraging repeated engagement.
• auditabilityAn audit log is a chronological record of system events, actions, or changes that supports inspection and accountability.
• accuracyAccuracy is the degree to which a model's outputs match the correct or intended results.
• ai teamAn AI team is a cross-functional group that combines the artificial-intelligence, data, and domain capabilities needed to develop an AI product.
• toleranceError tolerance is the degree of incorrectness users are willing to accept in a system's outputs while still considering it useful.
• data classificationData classification is the assignment of data to categories based on characteristics such as confidentiality, sensitivity, or handling requirements.
• intellectual property exposureIntellectual property exposure is the unauthorized inclusion, disclosure, or use of protected creative or proprietary material by an AI system.
• hybrid interfaceA hybrid interface combines graphical interaction with conversational or other interaction modes in one software experience.
• use caseA use case is a defined situation describing how a person or organization applies a system to accomplish a goal.
• instruction dataAn instruction dataset is a collection of examples pairing user instructions with source content and desired responses or revisions.
• stakeholderA person or group within or connected to an organization that can influence, support, constrain, or be affected by an initiative.
• quick winA quick win is a relatively small, rapidly deliverable improvement that provides near-term value or learning.
• content-based recommenderContent-based filtering recommends items whose attributes match the characteristics of items a user has previously preferred or interacted with.
• risk-based approachRisk-aware decision making is the practice of evaluating potential consequences and applying stricter controls to high-stakes actions.
• brand alignmentBrand alignment is the degree to which generated content conforms to an organization's brand voice, values, terminology, and communication standards.
• code readabilityReadability is the degree to which written content can be understood easily by its intended audience.
• intentionA user’s underlying goal or requested task expressed through an interaction.
• self-consistencySelf-consistency is a prompting technique that samples multiple reasoning or response paths and selects or synthesizes the answer supported by the most consistent results.
• complianceCompliance is adherence to applicable laws, regulations, policies, standards, and organizational requirements.
• tool resultTool output is the data returned by an external tool after it processes an agent's request.
• tool selectionTool selection is the process by which an agent chooses an appropriate external tool for a task based on tool descriptions and the current request.
• custom dataData tailored to a particular application, user population, or domain for adapting or evaluating a model.
• balanced datasetA balanced dataset is a dataset whose classes or labels are represented in relatively comparable proportions.
• ranking scoreA relative score is a numerical value that ranks or compares outcomes without necessarily representing an absolute probability or certainty.
• fidelityGroundedness is the degree to which generated content is supported by the information retrieved or provided as evidence.
• compactionPrompt compression is the reduction of prompt context and examples while preserving information needed for high-quality model outputs.
• data sourceA data source is a repository, system, or stream from which information can be retrieved for analysis or processing.
• read-only toolA software tool that can retrieve or inspect information without modifying external state.
• optimizationOptimization is the systematic adjustment of model, data, or algorithm choices to improve an objective measure of performance.
• logistic regressionLogistic regression is a statistical classification model that estimates class probabilities using a logistic function of input features.
• data agentA data agent is an AI agent that accesses, integrates, queries, and reasons over multiple organizational data sources.
• api integrationAPI integration is the incorporation of functionality or data from a software service through its application programming interface.
• foundational modelA foundational model is a broadly trained model that can be adapted to support many downstream tasks and applications.
• gradient boostingA gradient boosting algorithm builds an ensemble model by sequentially adding weak predictive models that reduce errors made by the existing ensemble.
• data engineerA data engineer is a practitioner who builds and operates systems and processes for collecting, transforming, governing, and providing access to data.
• data scientistA data scientist is a practitioner who applies statistics, mathematics, programming, and machine learning to analyze data and develop data-driven models.
• demand forecasting modelA demand forecasting model predicts future demand for a product, service, or resource from historical and contextual data.
• machine learning engineerA machine learning engineer is a practitioner who designs, trains, optimizes, deploys, and monitors machine-learning models and systems.
• margin of errorAn error margin is an allowed range of deviation between a model's forecast and the corresponding actual or intended value.
• mlops engineerAn MLOps engineer is a practitioner who operates and automates the infrastructure and workflows supporting machine-learning systems across their lifecycle.
• prompt engineerA prompt engineer is a practitioner who designs, evaluates, and manages prompts for applications powered by generative models.
• competitive advantageA competitive advantage is a capability, asset, or position that enables a product or organization to outperform competitors.
• memory fine-tuningMemory fine-tuning is a fine-tuning approach intended to encode persistent factual information into a model's learned parameters.
• machine learning workflowA machine learning lifecycle is an iterative process for preparing data, developing models, evaluating results, and applying machine-learning outputs.
• design thinkingA human-centered, iterative approach to problem framing and solution development that typically includes empathy, definition, ideation, prototyping, and testing.
• thought generationThought generation is the production of a partial solution for each component of a decomposed problem.

These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.