Enterprise Artificial Intelligence Glossary

A knowledge graph is a structured data network composed of interconnected entities, properties, and typed relationships represented as semantic nodes and edges. By mapping domain knowledge into explicit relational triples (Subject-Predicate-Object), knowledge graphs enable deterministic multi-hop graph traversal and eliminate large language model hallucinations in enterprise RAG systems.

A LangGraph State Machine is an open-source enterprise orchestration framework designed for building stateful, multi-actor AI applications using directed cyclical graphs. Unlike linear DAG (Directed Acyclic Graph) pipelines, LangGraph supports cyclical loops, fine-grained state persistence across user sessions, conditional branching edges, and deterministic human-in-the-loop state suspension.

A large language model (LLM) is a deep neural network architecture containing billions of parameters trained on vast multi-terabyte text corpuses using self-supervised autoregressive objective functions. Built on Transformer self-attention mechanisms, large language models predict probability distributions over sequential tokens to perform zero-shot reasoning, translation, and code generation.

A Reflection Loop is an architectural design pattern in autonomous AI systems where an agent evaluates, critiques, and refines its own generated outputs prior to final execution or user delivery. Operating in a two-stage loop—Generation (Actor) followed by Evaluation (Critic)—the agent detects syntax bugs, logical contradictions, or policy violations and iteratively revises its work.

A Semantic Router is an enterprise API gateway component that uses lightweight vector embeddings and cosine similarity to classify user query intent in sub-millisecond time. By routing incoming requests to specialized model tiers (e.g., fast 8B models for simple queries, 70B models for complex math), semantic routers reduce latency and token costs.

A vector database is a specialized data storage engine engineered to store, index, and query high-dimensional mathematical vector embeddings using Approximate Nearest Neighbor (ANN) search algorithms. By organizing unstructured data as dense vector points, vector databases execute semantic similarity searches across millions of embeddings in sub-15 millisecond latencies.

Agentic AI refers to autonomous artificial intelligence systems engineered with self-directed reasoning loops, environment state perception, planning capabilities, and external tool execution functions. Unlike static text generation models, agentic AI iteratively formulates multi-step action plans, invokes external APIs, and self-corrects errors to accomplish complex goal-driven workflows.

Agentic RAG (Retrieval-Augmented Generation) is an enterprise AI pattern where autonomous reasoning agents control document retrieval pipelines. Unlike static single-turn RAG that performs one naive vector search, Agentic RAG iteratively formulates search queries, evaluates chunk relevance, reformulates sub-queries, and navigates heterogeneous knowledge bases to resolve multi-hop enterprise inquiries.

Artificial Intelligence (AI) is the computational engineering discipline focused on designing algorithmic software systems capable of executing complex cognitive functions—including multi-step logical reasoning, automated decision-making, natural language comprehension, pattern recognition, and autonomous tool invocation—that historically required human perceptual and analytical intelligence across enterprise operational domains.

Autonomous Tool Use is the capability of an artificial intelligence agent to dynamically discover, select, parameterize, and execute external software functions, APIs, databases, or web services without human intervention. By mapping user objectives to structured function schemas (such as OpenAI Function Calling or Model Context Protocol tools), agents extend their capabilities far beyond text generation.

Activation-aware Weight Quantization (AWQ) is a post-training 4-bit quantization technique for large language models that compresses 16-bit model weights down to 4-bit integers without degrading reasoning accuracy. Unlike uniform quantization, AWQ protects the top 1% salient weights—identified by observing activation channels—reducing VRAM footprint by 70% while accelerating matrix multiplication on modern GPUs.

Cohere Rerank is an enterprise-grade cross-encoder reranking model designed to re-order candidate search results retrieved from vector databases or keyword search engines. Unlike bi-encoder embeddings that process queries and documents independently, Cohere Rerank performs full cross-attention over query-document pairs simultaneously, boosting RAG retrieval precision by 20% to 35%.

Context Window Expansion refers to algorithmic and architectural techniques that extend the maximum sequence token length an LLM can process without retraining from scratch. By modifying positional encoding algorithms—such as Rotary Position Embedding (RoPE) scaling, Linear Interpolation, and YaRN—models trained on 4k token contexts can process up to 128k or 1M tokens with high retrieval accuracy.

Contextual Compression is a RAG optimization technique that dynamically strips out irrelevant text from retrieved document passages before injecting context into an LLM prompt. By analyzing user query intent against candidate passages using small, fast extraction models or LLM sentence compressors, Contextual Compression reduces prompt token length by 60% to 85% while eliminating context noise.

Deep learning is a specialized subfield of machine learning that utilizes multi-layered artificial neural networks to automatically discover hierarchical feature representations from high-dimensional unstructured datasets like text, images, and raw audio. This architectural approach ensures predictable system behavior, verifiable computational outcomes, and continuous operational performance alignment across mission-critical enterprise AI deployments.

Dense Retrieval is a neural search methodology that projects queries and document passages into continuous, dense vector spaces using bi-encoder neural network embedding models. Unlike sparse keyword matching, dense retrieval measures semantic similarity by computing vector distance metrics (such as Cosine Similarity or Inner Product), capturing high-level conceptual intent even when query terms do not match document vocabulary.

Differential Privacy (DP) in machine learning is a mathematical framework that guarantees model outputs do not reveal whether any single individual's record was included in the training dataset. By injecting calibrated noise (DP-SGD) and enforcing strict privacy budget (epsilon ε) bounds during gradient updates, DP prevents membership inference attacks.

Direct Preference Optimization (DPO) is a stable, parameter-efficient algorithm for aligning large language models with human preference data. Unlike traditional RLHF (Reinforcement Learning from Human Feedback), which requires training a separate reward model and running complex PPO (Proximal Policy Optimization) reinforcement learning loops, DPO mathematically reparameterizes the reward function to optimize model policy directly using a binary cross-entropy loss.

EU AI Act Compliance represents the regulatory framework and technical enforcement standards mandated by the European Union AI Act. It categorizes AI applications into risk tiers (Unacceptable, High Risk, Specific Transparency, Minimal Risk), imposing strict governance, data lineage, risk management, and technical documentation requirements.

Fine-tuning is a machine learning technique in which a pre-trained foundation model undergoes additional gradient training on a specialized dataset to adjust its internal parameters. By modifying weights through Low-Rank Adaptation (LoRA) or full parameter updates, fine-tuning aligns model response style, domain vocabulary, and task compliance with specialized enterprise requirements.

GGUF (GPT-Generated Unified Format) is a binary file format specification created by the llama.cpp project for storing quantized large language models. Designed for single-file deployment across CPUs, Apple Silicon, and GPUs, GGUF encapsulates all model tensor weights, hyperparameter metadata, and tokenizer vocabularies into a unified, backward-compatible binary file.

GraphRAG is an advanced retrieval-augmented generation framework that combines vector embeddings with structured Knowledge Graphs. Originally introduced by Microsoft Research, GraphRAG extracts entities, relationships, and semantic claim triples from raw text, constructing a hierarchical knowledge graph. This enables global multi-hop reasoning and holistic summaries over entire document collections.

Hallucination Mitigation refers to a suite of algorithmic, architectural, and decoding techniques designed to prevent LLMs from generating false, ungrounded, or factually incorrect claims. Key mitigation strategies include Retrieval-Augmented Generation (RAG) context grounding, logit decoding constraints, self-consistency voting, and automated post-generation fact verifiers.

Hierarchical Navigable Small World (HNSW) is a graph-based Approximate Nearest Neighbor (ANN) search index structure designed for high-dimensional vector databases. By organizing vector embeddings into a multi-layer graph hierarchy—where top layers act as sparse express paths and bottom layers form dense local neighborhood connections—HNSW achieves logarithmic O(log N) search complexity with high recall (>98%).

Human-in-the-Loop (HITL) is an enterprise AI governance design pattern that inserts explicit human verification, authorization, or editing checkpoints into autonomous agent state execution workflows. By suspending execution state prior to executing high-risk operations—such as financial wire transfers, database writes, or patient notifications—HITL guarantees enterprise policy compliance and risk mitigation.

Hybrid Vector Search is an advanced retrieval paradigm that combines semantic dense vector search (using neural embeddings) with lexical sparse keyword search (using BM25 or TF-IDF). By merging candidate results using rank fusion algorithms such as Reciprocal Rank Fusion (RRF), hybrid search captures both high-level semantic intent and exact keyword matches (e.g., product SKUs, acronyms, and proper nouns).

Inter-Token Latency (ITL), also known as Time-Per-Output-Token (TPOT), is a performance metric measuring the average time elapsed between generating consecutive output tokens during an LLM's auto-regressive decoding phase. ITL determines the visual streaming reading speed experienced by users and is bounded primarily by GPU memory bandwidth.

ISO/IEC 42001 is the official international standard specifying requirements for establishing, implementing, maintaining, and continually improving an Artificial Intelligence Management System (AIMS). It provides a comprehensive structured risk management framework for enterprise organizations to ensure responsible, ethical, transparent, and compliant AI deployment.

KV Cache Optimization refers to memory compression and management techniques designed to reduce the GPU VRAM footprint of Key-Value (KV) tensors generated during auto-regressive LLM decoding. By applying INT8/FP8 cache quantization, PagedAttention memory virtualizing, and prompt caching, inference servers increase batch capacity and process longer context lengths.

LLM Observability is the practice of monitoring, tracing, and evaluating the runtime behavior, latency, token costs, and output quality of large language model applications in production. By capturing fine-grained OpenTelemetry trace spans across RAG steps, tool calls, and model invocations, observability platforms provide full operational visibility.

LLM-as-a-Judge is an automated evaluation methodology where a high-capability frontier model (such as GPT-4o or Claude 3.5 Sonnet) is used to score, rank, and evaluate the quality, correctness, and safety of candidate model outputs based on structured evaluation rubrics and pairwise comparison prompts.

LoRA (Low-Rank Adaptation) fine-tuning is a parameter-efficient fine-tuning (PEFT) technique that adapts pre-trained large language models by freezing base model weights and inserting trainable low-rank decomposition matrices into attention layers. LoRA reduces trainable parameter count by 99% and GPU memory demands by 3x while matching full fine-tuning accuracy.

Machine learning (ML) is a branch of artificial intelligence focused on building mathematical algorithms that learn statistical patterns from historical data to make automated predictions, classifications, or decisions without explicit rule-based programming. This architectural approach ensures predictable system behavior, verifiable computational outcomes, and continuous operational performance alignment across mission-critical enterprise AI deployments.

Mixture of Experts (MoE) is a sparse neural network architecture that replaces dense Feed-Forward Network (FFN) layers in transformers with multiple sub-networks called 'experts'. A top-k gating router evaluates incoming input tokens dynamically, routing each token to only 1 or 2 specialized experts per layer, achieving high total model capacity while keeping active compute overhead low.

The Model Context Protocol (MCP) is an open JSON-RPC 2.0 communication standard created to securely connect artificial intelligence models and agents to external data sources, enterprise tools, and API environments. Functioning as a universal USB-C cable for AI software, MCP standardizes how tools, vector resources, and context prompts are exposed, executed, and authenticated across heterogeneous cloud infrastructure.

Model Fallback Routing is an infrastructure resilience pattern that automatically redirects enterprise LLM requests to alternative model providers or secondary cluster endpoints when the primary model experiences rate limits, API timeouts, or service outages. By implementing circuit breakers and retry policies, fallback routers maintain high availability SLAs.

PagedAttention is a memory management algorithm inspired by virtual memory paging in operating systems, designed by the vLLM team to solve GPU memory waste in LLM inference. By storing Key-Value (KV) cache tensors in non-contiguous physical memory blocks rather than continuous VRAM chunks, PagedAttention eliminates internal and external memory fragmentation, reducing VRAM waste to under 4%.

pgvector Indexing is an open-source PostgreSQL extension that enables native vector similarity search directly within relational database tables. By adding high-performance vector indexing methods—specifically HNSW (Hierarchical Navigable Small World) and IVFFlat (Inverted File Flat)—pgvector allows enterprises to perform vector similarity queries alongside standard relational SQL joins, ACID transactions, and row-level security policies.

Plan-and-Solve Prompting is an advanced AI agent design pattern that explicitly separates task planning from task execution. The framework first formulates an overall step-by-step plan (decomposing complex enterprise goals into sub-tasks) and then systematically executes each sub-task sequentially, preventing LLMs from prematurely committing to flawed execution paths.

Prompt Caching is an optimization technique that stores pre-computed Key-Value (KV) attention tensors of static prompt prefixes—such as system instructions, broad context documents, or API schemas—in GPU VRAM or host RAM. Subsequent user requests sharing the exact prefix reuse cached tensors, bypassing prefill matrix computations and reducing cost and latency.

Prompt injection is a security vulnerability in large language model applications where untrusted user input or external document payload manipulates the model's system instructions, causing it to bypass authorization guardrails or leak sensitive data, requiring robust input sanitization and architectural guardrail boundaries.

Prompt Injection Defense encompasses architectural and algorithmic security techniques designed to prevent malicious users or untrusted data inputs from overriding an LLM's system instructions. By implementing input sanitization boundaries, dual-LLM evaluation nodes, structural delimiter tagging, and output guardrails, organizations protect downstream AI workflows.

QLoRA (Quantized Low-Rank Adaptation) is an advanced fine-tuning technique that compresses a frozen base LLM to 4-bit precision using NormalFloat4 (NF4) quantization while training 16-bit LoRA adapter parameters. QLoRA enables fine-tuning 70B parameter models on a single 48GB GPU without sacrificing accuracy compared to 16-bit fine-tuning.

Retrieval-Augmented Generation (RAG) is an enterprise AI architectural pattern that enhances large language model responses by fetching relevant, verified document chunks from a vector database or search index before generating an answer, neutralizing hallucination risks and expanding the LLM context window with real-time enterprise facts.

Self-RAG (Self-Reflective Retrieval-Augmented Generation) is an autonomous RAG framework that trains language models to dynamically decide WHEN to retrieve documents, evaluate whether retrieved passages are relevant, and critique their own generated responses. By emitting special Reflection Tokens (such as Retrieve, ISREL, ISSUP, and ISUSE), Self-RAG eliminates unnecessary database retrievals while boosting generation accuracy.

Shadow Evaluation is a deployment testing pattern where live production user traffic is mirrored asynchronously to a new candidate AI model alongside the active production model. The candidate model generates inferences silently without serving responses to end users, allowing engineers to benchmark performance, latency, and accuracy under real-world traffic.

Sparse BM25 Retrieval is a classic lexical search algorithm based on the Okapi BM25 probabilistic ranking function. BM25 evaluates document relevance against a user query by calculating Term Frequency (TF), Inverse Document Frequency (IDF), and document length normalization. Represented as sparse high-dimensional vectors, BM25 excels at exact keyword matching.

Speculative decoding is an inference optimization technique that accelerates large language model generation by employing a smaller draft model to propose candidate token sequences, which are verified in parallel by the target model. This parallel validation preserves the target model's exact output distribution while significantly reducing per-token latency.

Synthetic Data Generation is the algorithmic process of using frontier LLMs or generative models to produce artificial, high-quality training datasets, instruction pairs, or evaluation benchmarks. Combined with automated deduplication and quality validation filters, synthetic data overcomes real-world data scarcity and privacy constraints.

TensorRT-LLM is an open-source, highly optimized C++ library developed by NVIDIA for compiling and executing large language model inference on NVIDIA GPUs. It combines custom CUDA kernels, FP8 precision GEMM matrix math, in-flight (continuous) batching, and multi-GPU tensor parallelism to deliver peak hardware throughput.

The Agent Supervisor Pattern is a hierarchical multi-agent architecture where a centralized management node (the Supervisor Agent) directs, coordinates, and monitors a team of domain-specialized worker agents. The supervisor receives user objectives, delegates sub-tasks to individual worker agents, evaluates returned worker artifacts, and routes state execution until the overall goal is achieved.

The ReAct (Reasoning and Acting) pattern is an AI agent execution framework that combines step-by-step verbal reasoning (thought generation) with environment interaction (tool acting). By alternating between explicit reasoning steps and external API execution, ReAct allows large language models to dynamically adjust execution paths, overcome unexpected tool errors, and solve multi-step engineering tasks.

Time-to-First-Token (TTFT) is a critical performance metric in large language model inference that measures the total elapsed time between a client sending an input query payload and receiving the very first generated token response. TTFT is dominated by the computational prefill phase, where the model processes all input context tokens simultaneously.

Token Budgeting is an enterprise financial and technical control framework designed to monitor, limit, and optimize token consumption across multi-tenant LLM applications. By enforcing per-user token quotas, sliding-window prompt truncation, token-aware rate limiting, and dynamic context prioritization, organizations prevent runaway API billing and GPU resource exhaustion.

vLLM serving is an open-source, high-throughput large language model inference engine that optimizes GPU memory management using PagedAttention. By allocating Key-Value (KV) cache memory in non-contiguous virtual blocks, vLLM eliminates memory fragmentation, enables continuous batching across dynamic requests, and increases model serving throughput by 2x to 4x compared to HuggingFace Transformers.

Chunking Strategies are algorithms and data preprocessing methodologies used to divide large unstructured documents into smaller, semantically coherent text segments (chunks) prior to vector embedding generation. Choosing the correct chunking strategy—such as Fixed-Size with Overlap, Recursive Character Splitting, Semantic Sentence Splitting, or Document Structure Parsing—directly determines RAG retrieval accuracy.

Jailbreak Guardrails are multi-layered safety mechanisms engineered to detect and block adversarial prompt injection attacks aimed at bypassing an LLM's safety alignment. By evaluating input payloads against prefix attack vectors, hypothetical role-play framing, and semantic toxicity classifiers, guardrails preserve model alignment boundaries.

Multi-Agent Swarms are decentralized artificial intelligence architectures composed of multiple specialized AI agents collaborating to execute complex, multi-domain workflows. By dividing high-level goals into sub-tasks assigned to domain-specific agents—such as research agents, code generator agents, and auditor agents—swarms achieve higher accuracy, parallel execution speed, and fault isolation compared to single monolithic LLMs.

Schema Guardrails are token-level decoding constraints that force large language models to generate output structured strictly according to predefined JSON schemas, Pydantic models, or context-free grammars. By masking invalid candidate tokens at each generation step, schema guardrails eliminate JSON syntax errors and ensure 100% API schema compliance.

Zero Data Retention (ZDR) is a cloud security policy and contractual guarantee standard where artificial intelligence API providers process user prompts and outputs strictly in volatile RAM without saving or logging any payload data to persistent disk storage, safeguarding sensitive enterprise credentials and regulatory compliance.

Zero Data Retention (ZDR) VPC is a cloud deployment architecture where large language models execute inside an isolated Virtual Private Cloud (VPC) with zero persistent storage of prompt or response payloads. ZDR guarantees that inferences remain strictly transient, satisfying enterprise data privacy and non-logging SLAs.