LLM & Generative AI

Agent evaluation assesses LLM-based agents that plan, remember, and call tools in external environments by scoring full …

An agent loop is the control cycle that assembles input and accumulated context, invokes an LLM to plan and choose the n…

Agentic RAG is a retrieval-augmented generation architecture with an explicit planner (policy) that interleaves multi-st…

Agentic workflows are dynamic workflows in which multiple specialized AI agents collaborate to plan, reason, use tools, …

An AI agent is a system that pursues a stated goal by perceiving its environment, reasoning and planning, and repeatedly…

AI inference is the runtime phase in which a trained model with fixed weights processes new inputs to produce prediction…

Anthropic is an AI company that provides the Claude family of large language models and a developer platform, distributi…

Attention is a neural mechanism that computes a weighted aggregation of information by scoring the similarity between a …

AUC represents the area under the ROC curve and is used as a metric to evaluate the performance of a classification mode…

Batch inference is an offline prediction method that aggregates a large, fixed set of inputs and generates outputs in bu…

Amazon Bedrock is a fully managed AWS service that provides secure, enterprise-grade access to multiple foundation model…

BM25 is a probabilistic information-retrieval scoring function that ranks documents by summing per-query-term contributi…

A browser agent is an autonomous system that controls a real web browser to complete tasks by running a closed-loop of o…

ChatGPT is OpenAI's conversational AI application that turns natural-language user requests into answers or task outputs…

Claude is Anthropic’s family of large language models delivered through a developer platform — exposed via the Messages …

Computer Use is a tool-and-harness integration pattern where a model perceives the current UI via screenshots, proposes …

Context engineering is the disciplined selection, organization, and formatting of everything a language model reads on a…

A context window is the finite working memory of a language model that encompasses all tokens it can reference during ge…

CoT, or Chain-of-Thought, is a reasoning technique that prompts or trains large language models to produce or emulate in…

A Cross-Encoder is an interaction-based neural ranker that concatenates the query and document into a single Transformer…

Edge deployment means running AI models or apps close to where data is created — for example on factory lines, inside re…

An embedding is a learned vector representation that maps discrete objects or high-dimensional inputs into a continuous …

Evals are the practice of turning benchmark or user‑study measurements into decision‑ready evidence about a model’s capa…

An evaluation harness is a testing framework that runs language models or agents against standardized datasets, prompts,…

Fine-tuning is the process of continuing training from a pretrained model to adapt it to a specific task, domain, style,…

Gemini is Google’s family of multimodal generative models delivered through the Gemini API and Vertex AI, handling text,…

GPT-4o is OpenAI’s latest large language model that can handle text, speech, and images all at once. It’s designed to be…

Grouped-query attention is a method used in large language models (LLMs) and transformer-based AI systems to process sev…

Hallucination is a failure mode where an LLM produces fluent content that conflicts with source evidence, real-world fac…

Hybrid search is an information-retrieval design that runs lexical (keyword/BM25 over an inverted index) and vector (emb…

In-context learning is a pre-trained language model?s ability to adapt to a task at inference by conditioning on instruc…

Inference is the execution phase where a trained model receives new inputs and computes predictions, classifications, or…

Inference cost is the operational compute-and-infrastructure expense incurred each time a deployed LLM tokenizes a promp…

Inference latency is the actual time it takes for an AI model to process an input and return an output. It typically ref…

Inference-Time Scaling is a technique that improves a trained model’s outputs without retraining by allocating more comp…

A KV cache is the inference-time memory structure that stores previously computed attention key/value tensors in an auto…

KV offloading is an inference technique that tiers the self-attention Key/Value cache from GPU memory to CPU RAM or stor…

A large language model is a deep learning system trained on vast text corpora to understand and generate natural languag…

LoRA is a parameter-efficient fine-tuning method that freezes the base model and trains small low-rank adapters instead.

Model Context Protocol (MCP) is a stateful JSON-RPC 2.0 client-server protocol that lets AI hosts discover and invoke se…

An MCP server is the service-side component of Model Context Protocol that exposes server capabilities such as tools, re…

Mistral AI is a platform company offering a family of large language models via a first-party API and enterprise product…

Model cascading is a dynamic routing technique that speculatively runs small, low-cost models first, validates draft res…

Model distillation is a training method that teaches a smaller student model to imitate a larger teacher model's output …

Model parallelism is a distributed technique that enables training or serving neural networks too large for a single GPU…

A model router is an orchestration layer that selects which model should handle a request based on difficulty, modality,…

Model serving is the operational system that deploys a trained model behind APIs, batch jobs, or streaming endpoints and…

Mixture of Experts (MoE) is a sparse conditional-computation architecture in which a gating/router network selects a sma…

A multi-agent system is a network of multiple artificial intelligence agents that interact within a shared environment, …

Multi-hop retrieval is a technique where an AI system answers complex queries by sequentially retrieving and connecting …

Multi-stage training is a method for developing AI models—especially large language models (LLMs)—by progressively impro…

A multimodal model is an AI model designed to process or generate across two or more data modalities, such as text, imag…

Multimodal RAG is an extension of retrieval-augmented generation that embeds and indexes heterogeneous data such as text…

Natural Language Processing (NLP) is an AI discipline that enables computers to interpret and produce human language by …

NVIDIA provides an end-to-end AI software stack—NVIDIA AI Enterprise—spanning deployment microservices (NIM) and develop…

On-device AI means running artificial intelligence directly on your own device—like a phone, laptop, or tablet—instead o…

An open-source large language model (open-source LLM) is a type of AI language model whose underlying code and trained d…

OpenAI is an AI platform and API provider that offers models such as GPT‑5.5 and hosted tools to developers, exposing a …

OpenAI Codex is a cloud-based coding agent optimized for software engineering that can implement features, fix bugs, exp…

Output tokens are pieces of text generated by an AI model in response to input, where the model predicts the next most l…

PagedAttention is an LLM-serving memory technique that partitions the attention key–value (KV) cache into fixed-size pag…

Post-training is the stage that adapts a pretrained model to instructions, safety requirements, domain behavior, and hum…

Pre-training is the upstream phase that optimizes a model on large, broad data with self-supervised objectives such as n…

Prompt caching is an inference optimization where the provider reuses the model’s prefill state for an exact, sufficient…

Prompt injection is an attack that causes an LLM application to treat untrusted user input or external content as higher…

PyTorch is an open-source deep learning framework used to build and train neural networks. With its Python-based intuiti…

Retrieval-Augmented Generation (RAG) couples a retriever with a generator so an LLM conditions on top‑K query‑relevant c…

Re-ranking is a second-stage ranking step in RAG and search pipelines that jointly processes the user query with each in…

Real-time inference is a serving paradigm that exposes a trained model as an API to execute and respond immediately upon…

A reasoning model is a specialization of large language models that augments next‑token generation with intermediate rea…

A recurrent mechanism refers to an architectural design in AI models where the output from a previous step is fed back a…

Reinforcement Learning from Human Feedback (RLHF) is a post-training alignment method that treats a language model as a …

RoPE, or Rotary Position Embedding, is a Transformer positional encoding method that rotates query and key vectors by po…

Reciprocal Rank Fusion (RRF) is a rank-aggregation algorithm that merges result lists from different retrievers by ignor…

Self-attention is a mechanism where each element in an input sequence compares itself with all other elements to compute…

Self-supervised pretext tasks are label-free training objectives that exploit intrinsic structure in unlabeled data to l…

A Small Language Model (SLM) is a language model that performs natural-language understanding or generation with a small…

Speculative decoding is an inference acceleration method where a smaller drafter proposes multiple candidate tokens and …

Structured outputs let an LLM generate only data that matches a user‑provided schema (typically JSON Schema), so require…

Supervised fine-tuning is the process of further training a pre-trained AI model using additional labeled data, where hu…

SWE-bench is a software engineering benchmark that evaluates language models and agents on real GitHub issues by providi…

Synthetic data is artificially generated data created from rules, simulations, statistical models, or generative models …

TensorFlow is an open-source machine learning and deep learning framework developed by the Google Brain team, designed f…

Test-time compute is the inference-time budget of model evaluations, generated tokens, and wall-clock latency an LLM spe…

A token is the basic unit an LLM uses to represent and process text instead of reading raw characters or full words dire…

Tool calling is an interaction mechanism where an LLM, given definitions and input schemas of external tools, emits a st…

Tool use is an interaction pattern where an LLM emits structured calls against declared tool interfaces while the host a…

A Transformer is a neural network architecture that stacks self-attention and feed-forward blocks to learn relationships…

A vector database is a specialized storage and retrieval system for embeddings that supports low-latency nearest-neighbo…

A vision-language model is an artificial intelligence model designed to simultaneously understand and process both visua…

Visual instruction tuning is a supervised fine-tuning approach that aligns a vision encoder with a large language model …

vLLM is an open-source LLM serving engine that boosts throughput by managing the attention KV cache with PagedAttention—…