Concept Fables

Turning AI / Agent concepts into stories you can remember

This collection contains 684 concepts. Each entry first builds understanding through a fable, then reveals the definition and metaphorical mapping after the story ends.

See the breakdown of easily confused concepts Browse all concepts Keyword search

Enter the concept library from a question

Do you want to understand a word, or do you want to make a judgment?

If you are comparing options, start with the problem navigator; if you have already encountered a specific term, use the concept library below to look up definitions, stories, applicable conditions and adjacent concepts.

Enterprise Problem Navigation →Search 946 concepts and distinctions →Bring the concept back to a real job →

Featured concepts

Read concepts and stories by topic

Browse by topic

Agent runtime Tool calling Tool registry Weights & Biases Weave Synthetic data Low-Rank Adaptation Quantized low-rank adaptation Inference engine Distillation Batch inference Online inference Model routing Semantic cache Inference endpoint Model gateway Model Context Protocol AI job replacement rate Automated deflection rate Autonomous driving Copilot Watermark Service as software Agent operating model Agent Owner Authority boundaries Autonomy level Operational metrics Agent loop Executor Planner Action Computer use GUI agents Web agents Tool selection Run Run steps Real-time agent Voice agent Episodic memory Memory compression Semantic Memory Approval flow State diagram Agent-to-Agent protocol Long-term memory Working memory Draft area LangGraph Alignment Constitutional AI Direct Preference Optimization Preference data Preference optimization Proximal Policy Optimization AI feedback reinforcement learning Reinforcement learning from human feedback Reward model Large model applications Hierarchical Navigable Small World Graph Attention mask ALiBi BERT Causal attention / autoregressive Context length/window Cross-attention Decoder-only Diffusion language models Encoder-decoder Feedforward network GPT Grouped Query Attention Jamba Layer normalization Linear attention Mamba MoD Mixture of Experts Multi-query attention Residual connections RMSNorm Rotary position encoding Sliding window attention State space model SwiGLU T5 Vision Transformer Weight tying Audit evidence Audit trail Auditability Internal Audit Source citation tamper-evident log Agent workflow Workflow Idempotency Task queue Measurement Annual recurring revenue Churn rate Net revenue retention Customer health score Gross profit Unit economics model Enterprise sales Ideal Customer Profile Product-led growth Sales-led growth Customer acquisition cost Customer lifetime value Service Level Agreement Token cost Customer Success Professional services Managed service Hybrid pricing Outcome-based pricing Seat-based pricing Usage-based pricing Software as a Service Productized services Digital employee Skills Compliance mapping Control testing Technical documentation Control Framework Controls ISO 27001 SOC 2 Type I SOC 2 Type II SOC 2 AWQ bitsandbytes Deep pruning Distilled reasoning model GGML GPTQ GGUF Model Pruning Post-training quantization Quantization-aware training Sparsity Structured pruning Unstructured Pruning Width pruning Chain of thought and scratchpad C2PA Human in the loop Conversation thread Handoff Token budget Budget-aware routing Intelligent customer service LlamaIndex Document parsing OCR Data connector Data flywheel Beam search Decode Greedy decoding Logits Sampling Temperature Top-k sampling Top-p sampling Backpropagation Deep learning Neural networks Representation learning Canary release Canary release Private deployment Localized deployment Container Kubernetes Blue-green deployment Canary release Infrastructure as Code GitOps Continuous integration and continuous delivery Slot filling Noise predictor/UNet Pipeline parallelism Tensor parallelism Data parallelism FSDP ZeRO Layout analysis Table extraction Dataset datasheet Model Card System card Chroma Embed Semantic similarity Vector database Word embeddings Platform Engineering Developer experience Knowledge base Q&A Semantic Kernel Co-creation Pilot Evaluation benchmark Benchmark contamination Data leakage Distribution shift Perplexity Step success rate Tool call accuracy Trajectory evaluation LLM evaluation platform Cost Attribution Fine-tuning Instruction fine-tuning Parameter-efficient fine-tuning Supervised Fine-Tuning Full fine-tuning Instruction data Layer-wise learning rates rejection sampling Artificial Intelligence Model Diffusion models Diffusion Transformer Latent Diffusion Model Scenario-based implementation AI impact assessment AI use case checklist Data lineage Governance, Risk & Compliance Policy engine Proof of origin Shadow AI Traceability Transparency Trustworthy AI Agent policy AI disclosure Control everyone Human oversight Bias Explainability Fairness Three lines of defense Content filtering Policy as code AI Operations Delivery acceptance Autoregressive decoding Best-of-N sampling Context pruning Dynamic batching Early departure Flash Attention Inference Cache offloading Quantized KV cache Operator fusion Latency Retrospective decoding Medusa Model serving Paged attention Prefill Prefix caching Quantization Speculative decoding Speculative editing Speculative prefix tree Speculative sampling Test-time compute Throughput Tree search Continuous batching Prompt caching Runtime Retrieval-Augmented Generation Knowledge graph Ontology Triple Graph-enhanced retrieval generation Context window Language model Large Language Model LiteLLM Agent Memory Planning Hallucination Few-shot prompting In-context learning Prompt Prompt engineering System prompt Zero-shot prompting Chain of Thought LangChain TruLens Decoding stage Prefill Paged attention Text generation inference Cache LangSmith Helicone Cost tracking LangChain Expression Language Signature DSPy KV cache Time to First Token Prompt management Evaluation Golden dataset Model file configuration Ollama Feature Store Model version management Model drift Model registry Generalization Gradient descent Hyperparameters Loss function Machine learning Overfitting Parameters Reinforcement Learning Supervised Learning Test set Training set Unsupervised Learning Validation set OpenAI Assistants API Assistant OpenAI Responses API File search Code Interpreter Pinecone Benchmark case Auxiliary loss Expert routing Top-k routing Model inventory Model monitoring Model risk management Model validation Function calling JSON mode Structured output Structured output Supervisor agent Group chat orchestration AutoGen Agent team User Proxy Agent CrewAI CLIP Cross-Modal Alignment Multimodal models Video understanding Contrastive learning Cross-modal embedding Multimodal Vision-language model Entity extraction Relation extraction Haystack Tool call latency Span Tracking records OpenTelemetry Arize Phoenix Langfuse Milvus Qdrant Weaviate Change management Upgrade path Exception handling Hot reload Human-in-the-loop oversight Post-mortem review Quality assurance Rollback Rollback plan Root cause analysis Cloud cost operations Autoscaling Large language model operations Machine Learning Operations Dead letter queue Disaster recovery Observability Recovery point objective Recovery time objective Incident response Sampling review Version control Approval threshold Change Advisory Board AI incidents Business continuity Confidence threshold Fact-checking Incident Management Directed acyclic graph Intent recognition Semantic routing Router Adapter IA3 Merge weights Prefix Fine-Tuning Prompt fine-tuning Target module selection Workflow engine Quota Connector Durable execution Cloud native Internal developer platform Golden path AI policy Acceptable Use Policy pgvector BPE SentencePiece Corpus Foundation model Pretraining Self-supervised learning Anonymization Data Minimization Data protection impact assessment Data retention policy Differential privacy Membership inference attack Model inversion attack Personal data Personally identifiable information Pseudonymization Purpose limitation Secure aggregation Sensitive personal data Cross-border data transfer Data localization Data residency Supplier evaluation Vendor Lock-In Data processing agreement Data usage rights Sub-processor Training data exclusion Exit plan Third-party risk management Supplier due diligence Security questionnaire Product-market fit Time to value Self-service User profile Proof of concept Feature flag Onboarding Activate Minimum viable product Tasks to be completed AI-native workflow Prompt optimizer Semantic function Evaluation framework Double quantization NF4 Multimodal RAG Reranker Answer Relevance Context precision Context recall Faithfulness Retrieval precision Retrieval recall Graph of Thoughts Self-consistency Tree of Thoughts Program-aided language model Reasoning-Action Paradigm EU AI Act General-purpose AI High-risk AI systems Conformity assessment Rollback Circuit breaker Error budget Service Level Objective Degradation Model fallback Fact Anchoring Embed Cross-encoder BM25 Dual encoder Hypothetical document embeddings Maximum Marginal Relevance Multi-hop retrieval Query rewriting Recursive retrieval Approximate nearest neighbor search Document storage Vector database ecosystem Dense retrieval Hybrid search Semantic search Sparse retrieval Chunking Parent-child chunking Semantic chunking Sliding Window Chunking Reranking Query engine AI risk management AI risk management framework Hallucination risk Residual risk Risk acceptance Risk appetite Risk assessment Risk owner Risk Register Multi-tenant Tenant isolation Safety guardrails Sales Agent AI safety Attribute-based access control Backdoor attack Context contamination Data exfiltration Data poisoning Unauthorized agency Model poisoning OWASP LLM Top 10 Permission boundaries Role-based access control Supply chain risk Tool output contamination Sandbox Audit logs Permission model Principle of least privilege Secrets management Information Security Management System Separation of duties Data loss prevention vLLM Streaming responses Tokens per second Chunked Prefill Draft model State machine Speech recognition Speech synthesis Multi-agent orchestration Orchestration Context Engineering Byte Pair Encoding Token Tokenization Tokenizer Vocabulary Browser tools Plugin Activation Checkpoint bf16 Catastrophic forgetting Chain-of-thought training Chinchilla approach Continued pretraining Curriculum learning Data mixing Denoising objective GRPO Masked language modeling Mixed precision training Multi-task learning Next-word prediction Pretraining Reinforcement fine-tuning Verifiable reward reinforcement learning Scaling laws Federated learning Knowledge distillation Data augmentation Attention mechanism Decoder Encoder Multi-head attention Positional encoding Self-attention Transformer Citation AI assistant FAISS Agent negotiation Multi-agent reflection Task decomposition Agent debate Agent swarm Multi-agent consensus Role-playing multi-agent Emergent behavior Agent decision tracing Compliance log User feedback loop Hallucination detection LLM observability Latency monitoring Production drift detection Prompt version management Token usage monitoring Call chain tracing BIG-bench (Beyond the Imitation Game Benchmark)HumanEval (code generation evaluation)MMLU (Massive Multitask Language Understanding)SWE-bench HELM (Holistic Evaluation of Language Models)Model safety evaluation Red team testing Harmful content filtering Safety fine-tuning Evaluation suite Task success rate Adversarial examples Denial-of-service attack Indirect prompt injection Jailbreak Multimodal jailbreak Prompt injection Jailbreak Agent orchestration Blackboard architecture Decentralized multi-agent Hierarchical multi-agent Hybrid multi-agent Multi-agent system AI Governance Access control Agent Identity and Authentication Data privacy and PII masking Model approval process Model risk assessment Rate limiting Responsible AI Right of regret and revocation mechanisms Security incident response Usage policy LLM evaluation Adversarial evaluation Alignment Evaluation Capability assessment Human evaluation Round-by-round evaluation Capability contamination and data leakage Prompt sensitivity Agent routing Agent communication protocol

Domain classification

All concepts

684

Coding agent / developer tool

New

opencode

AI Coding Agent

opencode is an open-source AI coding agent for programming tasks, suited to understanding engineering in a codebase, editing files, running commands, debugging errors, generating tests, and refactoring code. It is a development tool for production business Agents, not a business Agent itself.

Read the fable →

Agent runtime / harness

New

OpenClaw

Agent Runtime / Harness

OpenClaw is a more general-purpose Agent runtime / harness that organizes channels, sessions, tools, skills, memory, tasks, automation and node capabilities into agent infrastructure that can run continuously.

Read the fable →

AI agent platform

Agent runtime

Agent Runtime

An agent runtime is the execution environment that lets an agent actually do work. It is not just a single large-model call; it manages task state, context, tool calls, permission boundaries, human approval, error recovery, log tracing, and multi-step execution.

Read the fable →

AI agent platform

Tool calling

Tool Calling, Function Calling

Tool calling, also called Function Calling, is a mechanism by which a model requests external functions, APIs, databases, file systems, or business systems in a structured way to complete tasks.

Read the fable →

AI agent platform

Tool registry

Tool Registry

A tool registry is the centralized catalog and control layer for the tools available to an Agent. It records tool names, purposes, parameter schemas, invocation methods, permission requirements, risk levels, approval rules, rate limits, return formats, and owners.

Read the fable →

AI application observability and evaluation tool

Weights & Biases Weave

Weights & Biases Weave is a tracking, evaluation, and debugging tool for AI applications that records model calls, prompts, outputs, and iteration processes.

Read the fable →

AI engineering / data pipeline

Synthetic data

Synthetic Data

Synthetic data is generated by models, rules or simulation systems for training, fine-tuning, evaluation and test coverage; the key lies in quality filtering and distribution boundaries.

Read the fable →

AI engineering / fine-tuning

Low-Rank Adaptation

LoRA, Low-Rank Adaptation

LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning method. It freezes most of the base model's weights and trains only small low-rank matrices to adapt to a specific task, significantly reducing training cost, VRAM requirements, and the complexity of multi-task adaptation.

Read the fable →

AI engineering / fine-tuning

Quantized low-rank adaptation

QLoRA, Quantized LoRA

QLoRA (Quantized LoRA) is a method for training LoRA adapters on a quantized base model. A common approach is to use 4-bit quantization to reduce memory usage while training a small number of adapter parameters.

Read the fable →

AI engineering / infrastructure

Inference engine

Inference Engine

The inference engine is responsible for efficiently executing the model's forward computation and managing GPU memory, KV cache, batching, concurrent scheduling, streaming output, and hardware acceleration.

Read the fable →

AI engineering / model optimization

Autopilot

Autopilot is a product model that lets AI complete tasks autonomously under clear goals, permission boundaries, and monitoring mechanisms. Humans are mainly responsible for setting goals, supervising boundaries, handling exceptions, and accepting results.

Read the fable →

AI product pattern

An Agent owner is the business or technical owner responsible for a certain Agent's goals, permissions, performance, risks, and lifecycle. In an enterprise, Agent responsibility cannot be left hanging between "the model," "the platform," "IT," or "the business team"; it must be clear who can change it, who watches the metrics, who handles incidents, and who decides on upgrades or decommissioning.

Read the fable →

Agent Operating Model

Authority boundaries

Authority Boundary

Authority boundaries are the explicit limits on the decisions, actions, spending, data access and external commitments an Agent is permitted to make. They further constrain "can call tools" into "under what conditions it can call them, how much scope it can affect, whether approval is required, and whether it can make external commitments."

Read the fable →

Agent Operating Model

Autonomy level

Autonomy Level

Autonomy level is the grading of how independently an Agent works: it can only recommend, draft first, execute after approval, execute autonomously within limits, or run fully automatically in specific scenarios. In enterprise governance, autonomy level determines the mode of human intervention, approval thresholds, audit requirements, and risk boundaries.

Read the fable →

Agent Operating Model

Operational metrics

Operating Metrics

Operational metrics are quantitative metrics used to manage Agents in production, such as completion rate, escalation rate, error rate, latency, cost, rework rate, first-pass acceptance rate, and incident rate.

Read the fable →

Agent architecture

Agent loop

Agent Loop

The agent loop is the basic closed loop of Agent operation: observe the state, reason and plan, choose an action, call tools to execute, then write the result back into context for the next round.

Read the fable →

Agent architecture

Executor

The executor is responsible for turning planned steps into actual actions, such as calling tools, running code, accessing external systems, and returning results for evaluation or re-planning.

Read the fable →

Agent architecture

Planner

The planner is responsible for breaking user goals into executable steps, subgoals, constraints, and dependencies, and adjusting the plan when new information appears.

Read the fable →

Agent capability

LangGraph is a graph-based framework for building stateful, multi-step, controllable Agent workflows.

Read the fable →

Alignment

Alignment makes an AI model's or Agent's behavior conform to human intent, organizational policy, safety boundaries and real task objectives.

Read the fable →

Alignment

Jamba is a hybrid architecture route combining Transformer attention with Mamba/SSM ideas, used to improve long-context processing efficiency.

Read the fable →

Architecture

Layer normalization

Layer Normalization

Layer Normalization is a technique for normalizing activation values inside a neural network layer, making the feature distribution of each sample more stable. In Transformers, LayerNorm is often used before and after attention and the FFN to improve training stability and gradient propagation.

Read the fable →

Architecture

Linear attention

Linear Attention

Linear Attention is a class of methods that reduce the computational complexity of attention from the quadratic growth of traditional self-attention to approximately linear growth.

Read the fable →

Architecture

Audit Trail

An Audit Trail is a traceable chain recording the entire process of an object, request, data item, or operation from start to finish. It typically includes timestamps, operators, inputs, outputs, system state, approvals, reasons for changes, and results.

Read the fable →

Auditability

Auditability refers to the ability of a system, process or organization to be audited, reviewed and explained. It is not a single logging feature, but an attribute formed jointly by records, permissions, evidence, controls, responsibilities, documentation and reproducible processes.

Read the fable →

Auditability

A workflow is a predefined or semi-structured process made up of steps, decisions, roles, and system integrations, used to reliably produce a business outcome. In AI / Agent systems, workflows coordinate model calls, tools, human approvals, state, and external systems.

Read the fable →

Backend engineering / operations

Read the fable →

Compliance/Security

Controls

Controls are specific measures designed to reduce risk, meet policy, or achieve control objectives. They can be preventive, detective, or corrective, such as access approval, least privilege, log monitoring, manual review, configuration baselines, change approval, backup, alerting, and incident response.

Read the fable →

Compliance/Security

ISO 27001

ISO/IEC 27001

ISO 27001 is the international standard for information security management systems, requiring organizations to continuously improve information security controls through risk management.

Read the fable →

Compliance/Security

SOC 2 Type I

SOC 2 Type I is an independent auditor's assessment, at a single point in time, of whether a service organization's system description and control design are suitable for meeting the relevant Trust Services Criteria, such as security, availability, confidentiality, processing integrity, and privacy.

Read the fable →

Compliance/Security

Model pruning is the removal of unimportant weights, neurons, attention heads, or layers from a trained model to reduce model size, memory usage, and inference latency.

Read the fable →

Compression

Post-training quantization

Post-Training Quantization, PTQ

Post-training quantization is a compression method that reduces weight or activation precision after model training is complete. It is fast but may lose some effectiveness.

Read the fable →

Compression

Quantization-aware training

Quantization-Aware Training, QAT

Quantization-Aware Training (QAT) simulates the effects of low-precision quantization during training or fine-tuning, letting the model adapt in advance to the compressed computation method. It is not compressing after training; rather, it lets the model feel the low-precision constraints during the training process.

Read the fable →

Compression

Sparsity

Sparsity means that many weights in a model are zero, or that many activations are inactive during computation. Methods such as pruning can create sparse structures, thereby reducing theoretical compute and memory requirements.

Read the fable →

Compression

Structured pruning

Structured Pruning

Structured pruning removes entire groups of neurons, attention heads, channels, layers or other regular structures. The pruned model has a tidier structure, making it easier for standard hardware and runtimes to actually run faster.

Read the fable →

Compression

Handoff is the process by which task ownership, context, and next actions are transferred from one Agent, workflow, system, or person to another executor. In enterprise Agents, handoffs often occur in multi-agent collaboration, human escalation, customer service transfer to specialists, and automated processes transferring to human approval.

Read the fable →

Cost / context

Token budget

Token Budget

A token budget is the upper limit of tokens allocated to a task, request, user, or workflow. It controls prompt length, retrieved context, tool output, model reasoning space, and answer length.

Read the fable →

Cost optimization

Budget-aware routing

Budget-Aware Routing

Budget-aware routing makes choices among budget, task value, quality requirements, and latency: not every request needs to go to the most expensive model, nor should you cut costs blindly. The system selects models, tools, retrieval depth, or execution paths based on the scenario.

Read the fable →

Customer support / Contact center AI

Intelligent customer service

AI Customer Service, Intelligent Customer Service

Intelligent customer service uses natural language understanding, knowledge base retrieval, dialogue management and ticket system integration to automatically handle common inquiries and hand off complex issues.

Read the fable →

Data framework for LLM applications

Logits are the raw scores a model outputs before softmax converts them into probabilities. Decoding parameters, constraints, banned words, temperature scaling, and sampling strategies all act on the logits or the distribution after conversion.

Read the fable →

Decoding

Sampling

Sampling is a decoding method that randomly selects tokens according to the model's probability distribution. It can increase diversity and reduce the mechanical feel of greedy output, but it also brings non-reproducibility and low-probability errors.

Read the fable →

Decoding

Read the fable →

Documentation

System card

System Card

The system card describes the capabilities, components, risks, mitigations, evaluation results, applicable boundaries and oversight mechanisms of a deployed AI system.

Read the fable →

Embedding database

Platform engineering improves engineering delivery efficiency and governance consistency through internal developer platforms, self-service capabilities, standard templates, and shared infrastructure.

Read the fable →

Engineering organization / platform

Developer experience

Developer Experience, DevEx

Developer Experience, DevEx measures the efficiency, smoothness, and cognitive load developers face from understanding, developing, testing, deploying, to troubleshooting.

Read the fable →

Enterprise AI application / Retrieval

Knowledge base Q&A

Knowledge Base Q&A, KBQA, RAG Q&A

Knowledge base Q&A connects enterprise documents, policies, product materials, and business knowledge to retrieval and generation systems, providing natural-language answers with supporting evidence.

Read the fable →

Enterprise AI orchestration SDK

Perplexity is a language model evaluation metric used to measure how easily a model predicts a text sequence. Lower perplexity usually means the model predicts that text distribution better. It is suitable for evaluating language modeling and how well a model fits similar corpora, but it cannot directly represent an Agent's task completion rate, factual correctness, safety or tool-calling ability.

Read the fable →

Evaluation

Step success rate

Step Success Rate

Step success rate measures the proportion of intermediate steps executed correctly in a multi-step task. For enterprise Agents, it locates quality problems better than final success rate alone: whether the plan is reasonable, whether retrieval hits, whether the right tool is chosen, whether parameters are correct, whether permissions pass, and whether exceptions are handled.

Read the fable →

Evaluation

Tool call accuracy

Tool Call Accuracy

Tool call accuracy measures whether an Agent picks the right tool, fills in the right parameters, calls it compliantly, and reasonably uses the results returned by the tool.

Read the fable →

Evaluation

Trajectory evaluation

Trajectory Evaluation

Trajectory evaluation is the holistic assessment of an Agent's execution process, including plans, reasoning steps, tool calls, observations, retries, exception handling, and final output. It suits multi-step tasks because the final answer may mask overreach, luck, incorrect tool calls, or non-reproducible paths in the process.

Read the fable →

Evaluation infrastructure

LLM evaluation platform

LLM Evaluation Platform

An LLM evaluation platform manages datasets, test cases, scoring functions, human annotation, model judges, regression checks, experiment comparisons, and release gates. It turns evaluation from scattered scripts into a continuous quality system.

Read the fable →

FinOps / operations

Artificial Intelligence, AI

Artificial Intelligence is the broad field of building systems capable of intelligent tasks such as perception, language, reasoning, planning and decision-making. In an enterprise context, AI should not be understood as a single model, but as a combination of capabilities, data, processes, controls and business delivery methods.

Read the fable →

Foundations

Provenance

Provenance is the evidence about the origin, author, authorization and transformation history of data, content or output. Enterprise Agents read documents, retrieve knowledge bases, call external APIs, reference customer data and generate new content, so it must be known where the material comes from, whether it is trustworthy, whether there is a right to use it, and what processing it has undergone.

Read the fable →

Governance

Shadow AI

Shadow AI is the use of AI tools, models or Agents by employees or teams without formal approval, security review, procurement visibility or governance controls. It can bring data leaks, vendor risk, compliance violations, unauditable decisions and duplicate procurement.

Read the fable →

Governance

Traceability

Traceability is the ability to reconstruct why a system produced a certain output or took a certain action. In enterprise Agents, it requires linking the user request, context, retrieval sources, model selection, prompt version, tool calls, permission decisions, human approvals, external writes and final results.

Read the fable →

Governance

Transparency

Transparency requires clearly disclosing the AI system's purpose, capabilities, limitations, data use, decision-making role, responsible parties, and user choices.

Read the fable →

Governance

Bias is a systematic skew in data, models, design, objective functions, or deployment processes that can lead to inaccurate, unfair, or harmful outcomes. Bias in enterprise Agents can come from historical data, labeling bias, insufficient sampling, feedback loops, proxy variables, or business rules.

Read the fable →

Governance/Risk

Explainability

Explainability is the ability of an AI system to explain the key reasons, evidence, factors, and limitations of an output or decision in a way suited to the use case. For enterprise Agents, explainability helps business staff judge whether to adopt a recommendation, helps users appeal, and helps audit and risk teams trace problems.

Read the fable →

Governance/Risk

Early Exiting

Early exit, also called early-exit inference, lets a model stop at an intermediate layer when confidence is high enough, instead of running the entire model every time. Simple samples use less computation, while difficult samples still take the full path.

Read the fable →

Inference

Flash Attention

FlashAttention

FlashAttention is an IO-optimized attention algorithm that reduces memory reads and writes through tiling and operator fusion, lowering the overhead of attention on long sequences.

Read the fable →

Inference

Here, inference refers to model inference or model execution: applying a trained model to new input and generating output. It is not logical reasoning in the human sense, but the model performing computation on the current request.

Read the fable →

Inference

Cache offloading

KV Cache Offloading

KV cache offloading moves less active KV cache blocks from GPU memory to CPU memory or SSD. This allows the system to support longer contexts or more concurrent requests without being entirely limited by GPU memory.

Read the fable →

Inference

Quantized KV cache

KV Cache Quantization

KV cache quantization stores the key-value cache during inference at lower precision, supporting longer context or higher concurrency under limited VRAM.

Read the fable →

Inference

Operator fusion

Kernel Fusion

Operator fusion combines multiple GPU operations into a single kernel, reducing memory round trips and launch overhead, and is a common optimization in high-performance inference runtimes.

Read the fable →

Inference

Latency

Latency is the time a model or Agent system takes from receiving a request to returning a response, or generating tokens. It may include queuing, routing, retrieval, tool calls, prefill, first token, decoding, and post-processing.

Read the fable →

Inference

Retrospective decoding

Lookahead Decoding

Lookahead decoding is a class of methods for accelerating autoregressive generation: it constructs candidate n-gram sequences in parallel, then the main model verifies the candidates in parallel. It is related to speculative decoding but does not rely on a separate draft model.

Read the fable →

Inference

A runtime is the environment that executes an Agent, responsible for tool calls, permissions, memory, state, and lifecycle control, determining what the system can do and how it converges.

Read the fable →

Knowledge access

Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG)

Retrieval-Augmented Generation (RAG) retrieves relevant material from external knowledge bases, document stores, databases or search systems before generating an answer, then passes that material to the model as context for generating the answer.

Read the fable →

Knowledge representation

Knowledge graph

Knowledge Graph

A knowledge graph uses entities, relationships, and attributes to represent knowledge in a structured way, suitable for entity queries, relationship analysis, compliance tracing, and recommendations.

Read the fable →

Knowledge representation

Ontology

Ontology is a structured definition of the concepts, categories, properties, relationships, and constraints in a domain. It helps systems unify "what is what, how they relate to each other, and which inferences are allowed." In enterprise Agents and knowledge graphs, ontologies are used to reduce naming confusion and support relational reasoning, permission boundaries, data integration, and explainable retrieval.

Read the fable →

Knowledge representation

LiteLLM provides an OpenAI-style unified interface, SDK and proxy layer across multiple model providers. Developers can integrate different models in a more consistent way without having to write a completely different calling logic for each provider.

Read the fable →

LLM Agents

Agent

An agent is an AI work unit that, under a given goal, perceives context, plans steps, calls tools, executes actions, and continues advancing based on results.

Read the fable →

LLM Agents

Memory

Memory is the ability of an AI system to save and reuse information, enabling an Agent to maintain continuity across steps, sessions or user interactions. Enterprise Agent memory typically includes working memory, long-term memory, episodic memory and semantic memory.

Read the fable →

LLM Agents

Hallucination is information generated by a model that sounds plausible but is actually wrong, unsupported by evidence, or fabricated out of thin air. It commonly occurs when a model tries to fill in missing facts, answers with overconfidence, confuses similar information, or lacks reliable context.

Read the fable →

LLM Interaction

Few-shot prompting

Few-Shot Prompting

Few-Shot Prompting provides a small number of input-output examples in the prompt, letting the model learn the task format, judgment scale, and output style from the examples. It does not change the model weights, but guides model behavior within the current context.

Read the fable →

LLM Interaction

In-context learning

In-Context Learning

In-Context Learning is the model's ability to temporarily adjust its behavior based on instructions, examples and context in the prompt, without updating model weights. Few-shot, format examples, counterexamples and task descriptions all make use of this ability.

Read the fable →

LLM Interaction

LangChain is a framework for building LLM applications, using composable components to connect prompts, models, retrievers, tools, chains, and Agent workflows.

Read the fable →

LLM evaluation and observability library

TruLens

TruLens is an LLM application evaluation and observability library that can add feedback functions to RAG and Agents to measure relevance, trustworthiness, and groundedness.

Read the fable →

LLM inference

Decoding stage

Decode Phase

The decoding phase is the autoregressive generation phase after prefill: the model generates one token at a time, reads the KV cache at each step, computes the next token, appends the new cache, and may stream tokens back to the user.

Read the fable →

LLM inference

LangSmith is an observability, debugging, evaluation, monitoring and testing platform for LLM applications in the LangChain ecosystem. It helps teams record traces, manage datasets, run evals, compare experiments, debug prompts and observe Agent workflows.

Read the fable →

LLM observability and gateway tool

Signature here refers to a task signature in DSPy, not a cryptographic signature. It is used to declare the input fields, output fields and semantic description of an LLM task.

Read the fable →

LLM programming and optimization framework

Ollama is a local model running and management tool that can pull, package, run and serve LLMs, suitable for deployment on personal machines or in private environments.

Read the fable →

MLOps / data platform

Unsupervised Learning

Unsupervised Learning finds structure in data without explicit labels, such as clustering, dimensionality reduction, anomaly detection, and topic discovery. It is suited to exploring unknown patterns, organizing large amounts of data, and discovering similar groups. When enterprises use it, they should avoid treating automatically discovered clusters directly as business truth; domain expert interpretation and validation are usually needed.

Read the fable →

Machine Learning

Validation set

Validation Set

A Validation Set is a dataset held out during development for tuning parameters and selecting approaches. It helps judge whether a model generalizes, but because it gets reviewed repeatedly, it cannot be treated as a final objective score.

Read the fable →

Managed agent API

OpenAI Assistants API

The OpenAI Assistants API is a managed assistant interface that maintains persistent conversation state around threads, messages, runs, tools and file management.

Read the fable →

Managed agent object

Assistant

An assistant is a preconfigured AI entity in a hosted Agent platform, usually bound to instructions, models, tools, and files—not just a one-off temporary chat.

Read the fable →

Managed model and agent API

OpenAI Responses API

The OpenAI Responses API is a unified interface for models and Agent-style workflows, supporting capabilities such as model responses, tool calls, multimodal input, structured output, and built-in tools.

Read the fable →

Managed retrieval tool

File search

File Search

File Search is a managed retrieval capability: it indexes uploaded files and retrieves relevant content when the model generates an answer. It typically handles the file ingestion, chunking, indexing, and recall stages of RAG, and is suitable for contracts, manuals, knowledge bases, and similar scenarios, but it still requires handling permission isolation, file updates, citation quality, and answer evaluation.

Read the fable →

Managed tool execution

Code Interpreter

Code Interpreter is a hosted tool that lets models write and run code in a controlled sandbox, commonly used for data analysis, file processing, computation, visualization, and temporary script execution.

Read the fable →

Managed vector database

AutoGen is a framework for building multi-agent applications. It lets multiple conversational agents collaborate through message passing, and can call tools, execute code, involve humans, or orchestrate group chats.

Read the fable →

Multi-agent organization unit

Agent team

Crew

A Crew is a team organizational unit in a multi-Agent system, typically composed of multiple Agents with roles, goals, tools, and task boundaries. It emphasizes organizing complex work into a collaborative team rather than having a single Agent handle all capabilities.

Read the fable →

Multi-agent role

User Proxy Agent

User Proxy Agent is a role representing the user in multi-Agent systems, commonly seen in AutoGen. It can provide human input, approve actions, execute controlled code on the user's behalf, or hand questions back to a real person.

Read the fable →

Multi-agent workflow framework

CrewAI

CrewAI is a multi-Agent workflow framework that organizes collaboration through role-based Agents, tasks, tools, and processes, and is suited to building team-style automation.

Read the fable →

Multimodal

CLIP

Contrastive Language-Image Pre-training

CLIP (Contrastive Language-Image Pre-training) aligns image and text representations through contrastive learning: matching image-text pairs are closer in vector space, while mismatched ones are farther apart. It is an important foundation for multimodal retrieval, zero-shot classification, and image-text understanding.

Read the fable →

Multimodal

Haystack is an open-source framework for search, question answering, RAG, and LLM pipelines, commonly used for retrieval-augmented applications over enterprise documents and knowledge bases.

Read the fable →

Observability

Tool call latency

Tool Call Latency

Tool call latency is the time an Agent waits for an external tool, API, database, browser or service to return a result. Even if the model itself is fast, tool waiting can become the main bottleneck in user experience.

Read the fable →

Observability artifact

Arize Phoenix is an open-source AI observability and evaluation tool used to trace, evaluate, and analyze the behavior of LLM, RAG, and machine learning applications.

Read the fable →

Open-source LLM observability platform

Langfuse

Langfuse is an open-source LLM observability platform for tracing calls, prompts, generated outputs, scores, costs, and user feedback.

Read the fable →

Open-source vector database

Milvus

Milvus is an open-source distributed vector database designed for large-scale vector similarity search. It uses a storage-compute separation architecture and supports multiple index types (HNSW, IVF_FLAT, IVF_PQ, DiskANN, etc.), horizontal scaling, data persistence and hybrid queries.

Read the fable →

Open-source vector database

Qdrant

Qdrant is an open-source vector database and search engine that supports vector similarity search, structured metadata filtering, HNSW indexing and quantization.

Read the fable →

Open-source vector database

Rollback is a mechanism for quickly restoring a previous stable state when a new version, model, prompt, tool, configuration or process runs into problems. It is not a temporary fix after failure, but part of release and change design.

Read the fable →

Operations

Rollback plan

Rollback Plan

A Rollback Plan is a recovery plan defined in advance before a change goes live, including trigger conditions, owners, recovery steps, time limits, verification methods and communication methods. It is more complete than a simple rollback action, emphasizing when to roll back, who decides, how to execute, and how to confirm successful recovery.

Read the fable →

Operations

Root cause analysis

Root Cause Analysis, RCA

Root Cause Analysis (RCA) is an analytical method for finding the true cause behind a problem, rather than only treating surface symptoms. It is often used in postmortems of incidents, quality issues, system failures, and business anomalies.

Read the fable →

Operations / finance

Cloud cost operations

FinOps

FinOps is a practice framework spanning engineering, finance and business teams for managing cloud resources, model calls, budgets, attribution and cost optimization.

Read the fable →

Operations / infrastructure

A Router is a component that dispatches requests, tasks or intermediate steps to the appropriate model, tool, workflow, skill or Agent. It can make decisions based on intent, capability, cost, latency, risk, permissions, load and policy.

Read the fable →

PEFT

Quota defines the upper limit of resources a user, tenant, or plan can use within a certain period, such as tokens, request counts, storage, or concurrency. For LLM/Agent platforms, quota is used to protect shared models, GPUs, tool executors, retrieval services, and budgets, preventing a single tenant or abnormal task from exhausting resources.

Read the fable →

Platform / integration

pgvector is a PostgreSQL extension that adds vector storage, similarity search, and metadata filtering to relational databases, and is often used in lightweight RAG systems.

Read the fable →

Preprocessing

BPE

Byte-Pair Encoding

Anonymization

Anonymization is processing personal data to a state where specific individuals cannot be identified, and re-identification is difficult under reasonably available information and means.

Read the fable →

Privacy

Read the fable →

RAG / search

Reranker

A reranker is a model or component that rescores and reorders candidate documents or chunks after initial retrieval.

Read the fable →

RAG evaluation

Faithfulness measures whether a model's answer stays true to the given evidence or context. An answer may be fluent, logical, and even broadly correct in direction, but if it adds content not supported by the evidence, rewrites factual relationships or exaggerates conclusions, it is not faithful enough.

Read the fable →

RAG evaluation

Retrieval precision

Retrieval Precision

Retrieval Precision measures the proportion of truly relevant results among those returned by retrieval. In RAG, it answers the question: of the chunks that enter the model's context, how many actually support the current question.

Read the fable →

RAG evaluation

Retrieval recall

Retrieval Recall

Retrieval Recall measures the proportion of all relevant results that the system successfully retrieves. In RAG, it answers: was the key evidence recalled? Low recall causes the model to answer based on incomplete context, which is especially dangerous in compliance, contracts, after-sales, and medical scenarios.

Read the fable →

Reasoning

Graph of Thoughts

A thought graph organizes intermediate ideas with a graph structure, letting reasoning nodes branch, merge, backtrack, and be reused—well suited to compositional reasoning tasks.

Read the fable →

Reasoning

Self-consistency

Self-Consistency

Self-Consistency is a method that samples multiple reasoning paths for the same question, then selects the final answer through voting or consistency. It is often used together with Chain-of-Thought to improve the stability of complex reasoning.

Read the fable →

Reasoning

A Cross-Encoder feeds the query and candidate text together into the same model, letting the model judge relevance in joint context. It is usually more precise than encoding queries and documents separately, and is suitable for RAG reranking, search result fine ranking, QA evidence selection, and compliance material matching.

Read the fable →

Retrieval

Chunking is the process of cutting long documents into small passages suitable for retrieval and for placing into the model context. RAG systems usually cannot directly stuff an entire document into the model; instead, they first chunk, vectorize, and index the document, then recall relevant chunks based on the question. Chunking quality directly affects recall precision, context completeness, citation accuracy, and answer quality.

Read the fable →

Retrieval preparation

Parent-child chunking

Parent-Child Chunking

Parent-child chunking uses smaller child chunks for vector retrieval, then returns larger parent chunks as context, balancing hit precision and semantic completeness.

Read the fable →

Retrieval preparation

Semantic chunking

Semantic Chunking

Semantic Chunking splits documents according to semantic boundaries, rather than mechanically by fixed character count, token count, or line count. It tries to make each chunk a relatively complete unit of meaning, such as a rule, a step, a definition, a set of exceptions, or a case.

Read the fable →

Retrieval preparation

Sliding Window Chunking

Sliding Window Chunking splits long documents using fixed-size windows and preserves a certain overlap between adjacent chunks. Its goal is to reduce the risk of semantics or facts being cut off right at chunk boundaries.

Read the fable →

Retrieval quality

Multi-agent orchestration assigns tasks to multiple specialized Agents and coordinates roles, dependencies, messages, outputs, conflicts, and final results.

Read the fable →

System architecture

Orchestration

Orchestration coordinates models, tools, workflows, Agents, human approvals, state, and error handling so that a complete task can advance reliably across multiple components. It focuses on global dependencies, routing, permissions, retries, escalation, parallelism, and result merging.

Read the fable →

System design

Context Engineering

Context Engineering is the engineering practice of selecting, organizing, compressing, updating and isolating context for a model. It is broader than traditional Prompt Engineering: it not only writes instructions, but also handles memory, retrieval, tool results, task state, permission boundaries, source trustworthiness and token constraints.

Read the fable →

Tokenization

Byte Pair Encoding

Byte Pair Encoding, BPE

Byte Pair Encoding, BPE is a common subword tokenization method. It starts from basic symbols and repeatedly merges the most frequently adjacent symbol pairs in the training corpus to obtain a fixed vocabulary.

Read the fable →

Tokenization

Token

A token is the basic unit of text or data that a model processes; it may be a word, word piece, character fragment, punctuation mark or symbol. A model's context window, API billing, latency, throughput and output length are usually calculated in tokens, not characters or words.

Read the fable →

Tokenization

Tokenization is the process of converting raw text into a sequence of tokens that a model can process. Common methods include subword tokenization approaches such as BPE and SentencePiece. Tokenization affects token count, context usage, cost, speed, multilingual performance, code handling, and special character behavior.

Read the fable →

Tokenization

Tokenizer

A Tokenizer is a software component that maps text to token IDs and restores generated token IDs back to text. It usually includes a vocabulary, a tokenization algorithm, and special token rules.

Read the fable →

Tokenization

Vocabulary

A vocabulary is the set of tokens a tokenizer or model can recognize, usually including subword pieces, symbols, special tokens, and their token IDs. It determines how text is encoded, which pieces can be represented efficiently, and how much context space the same text occupies.

Read the fable →

Tool

Browser tools

Browser Tool

Browser Tool is a tool that lets an Agent open web pages, read pages, click controls, fill out forms, or execute web tasks.

Read the fable →

Tool interface

Attention Mechanism

It lets the model, when processing a string of information, not treat all positions equally, but dynamically judge which positions are more important based on what is currently being generated or understood.

Read the fable →

Transformers

Decoder

The Decoder is the part of the Transformer used to generate the output sequence. A typical Decoder uses a causal mask during generation, can only attend to already-generated tokens, and predicts the next token step by step.

Read the fable →

Transformers

Jailbreaking is an attempt to bypass the safety, policy, and permission boundaries of a model or Agent through role-playing, restating the question, obfuscating expression, multi-turn inducement, encoded language, and other methods.

Agent Communication Protocol

Agent Communication Protocol is a protocol or language that defines the message format, semantics and interaction rules between Agents. It not only specifies fields, but also expresses communication intent, such as request, notification, commitment, refusal, query, delegation and confirmation.

Read the fable →