Agent Memory
- Definition
- Agent memory is the stored information an AI agent carries beyond a single prompt, from the current conversation to facts, preferences, and past outcomes it retrieves in later sessions.
Showing 20 of 77 terms
AI Alignment
- Definition
- AI alignment is the effort to make an AI system's goals and behavior match what its operators and users actually intend, including their values and constraints, rather than a literal or convenient…
AI Gateway
- Definition
- An AI gateway is a proxy that sits between applications and model endpoints to enforce authentication, routing, rate limits, logging, and data policies for every request to an AI model.
Attention Mechanism
- Definition
- An attention mechanism lets a neural network decide how much weight to give each part of its input when producing each part of its output, and it is the central operation inside…
Benchmark
- Definition
- A benchmark is a standardized set of tasks and a scoring method used to compare AI models against each other, giving a repeatable measure of capability on a defined problem.
BERT
- Definition
- BERT (Bidirectional Encoder Representations from Transformers) is a language model from Google that reads text in both directions at once and became the standard base for text classification, entity extraction, and search…
Chain-of-Thought Prompting
- Definition
- Chain-of-thought prompting asks a language model to write out its intermediate reasoning steps before giving a final answer, which improves accuracy on arithmetic, logic, and multi-step business rules.
ChatGPT
- Definition
- ChatGPT is a conversational AI assistant from OpenAI, released in November 2022, that lets people interact with GPT language models through a chat interface on the web, desktop, and mobile.
Chunking
- Definition
- Chunking is the step in a retrieval-augmented generation pipeline that splits documents into smaller passages so they can be embedded, searched, and supplied to a language model within its context window.
Claude
- Definition
- Claude is a family of large language models and a conversational AI assistant developed by Anthropic, first released in March 2023 and offered through a chat product and an API.
Closed-Source Model
- Definition
- A closed-source model is an AI model whose weights and training details stay private, so customers can use it only through the provider's hosted API or product under the provider's terms.
Context Window
- Definition
- The context window is the maximum number of tokens a language model can take in at once, covering the system prompt, the conversation so far, any retrieved documents, and the answer it…
Continuous Batching
- Definition
- Continuous batching is an inference server technique that adds new requests to the running batch and removes finished ones at every generation step, keeping the GPU busy and raising throughput.
Direct Preference Optimization (DPO)
- Definition
- Direct preference optimization (DPO) is a method for aligning a language model to human preferences by training directly on pairs of preferred and rejected responses, without a separate reward model.
Embeddings
- Definition
- Embeddings are numeric vectors that represent the meaning of text, images, or other data so that items with similar meaning sit close together and can be compared, searched, and clustered by a…
Few-Shot Learning
- Definition
- Few-shot learning is a model's ability to learn a new task or category from only a handful of examples, with one-shot learning as the special case of a single example.
Fine-Tuning
- Definition
- Fine-tuning is the process of continuing to train a pre-trained model on a smaller, task-specific dataset so that it adapts to a company's language, formats, and decisions.
Foundation Model
- Definition
- A foundation model is a large AI model trained on broad data that can be adapted to many different tasks through prompting, retrieval, or fine-tuning instead of being built for a single…
Generative AI
- Definition
- Generative AI is a class of AI models that produce new content, such as text, images, code, or audio, by learning the patterns in large collections of existing examples.
GPT (Generative Pre-Trained Transformer)
- Definition
- GPT, short for Generative Pre-Trained Transformer, is a family of large language models built on the transformer architecture and trained to predict the next token in a sequence of text.
GPU Memory (VRAM)
- Definition
- GPU memory (VRAM) is the fast memory on a graphics card that holds a model's weights and working data during inference, and its size decides which models can run and how many…
Grounding
- Definition
- Grounding is the practice of tying a generative AI model's answers to verified sources, such as retrieved documents or database records, so the output reflects real facts instead of the model's memory.
Hallucination
- Definition
- A hallucination is output from a generative AI model that is fluent and confident but false, such as an invented fact, citation, number, or claim that an action was completed.
Inference Endpoint
- Definition
- An inference endpoint is the network address where a deployed model accepts requests and returns predictions or generated text, whether the model runs in a vendor's cloud or on the company's own…
Inference Server
- Definition
- An inference server is the software that loads a trained model onto GPUs, accepts requests over the network, and schedules them for efficient generation, turning model weights into a usable service.
Instruction Tuning
- Definition
- Instruction tuning is a training stage that teaches a pre-trained language model to follow directions by showing it many examples of instructions paired with good responses.
Knowledge Cutoff
- Definition
- The knowledge cutoff is the date after which a language model's training data stops, so the model has no built-in knowledge of events, products, rules, or prices that appeared later.
Knowledge Distillation
- Definition
- Knowledge distillation is a technique that trains a smaller student model to reproduce the behavior of a larger teacher model, so the result runs faster and cheaper while keeping most of the…
KV Cache
- Definition
- The KV cache is the GPU memory an inference server uses to store the attention keys and values of every token already processed, so each new token is generated without recomputing the…
Large Language Model (LLM)
- Definition
- A large language model (LLM) is a neural network trained on very large amounts of text to predict the next token, which lets it draft, summarize, translate, classify, and answer questions in…
Large Models
- Definition
- Large models are AI models with billions or trillions of parameters trained on very large datasets, a category that includes language, vision, and multimodal models. Size brings broader capability and higher running…
LLM Observability
- Definition
- LLM observability is the practice of recording and analyzing what a language model application does in production, including prompts, responses, tool calls, latency, token usage, errors, and quality signals.
LLM-as-a-Judge
- Definition
- LLM-as-a-judge is an evaluation method in which a language model scores or ranks the outputs of another model against a rubric, replacing or supplementing human review at scale.
Long Context
- Definition
- Long context describes language models and workloads that handle prompts of hundreds of thousands of tokens, enough to supply whole documents or long histories instead of retrieving short excerpts.
LoRA (Low-Rank Adaptation)
- Definition
- LoRA (Low-Rank Adaptation) is a fine-tuning method that freezes a model's original weights and trains small added matrices instead, producing a compact adapter that changes the model's behavior.
Mixture of Experts (MoE)
- Definition
- Mixture of experts (MoE) is a model architecture that splits parts of a neural network into several expert sub-networks and activates only a few of them for each token, lowering compute per…
Model Card
- Definition
- A model card is a structured document published with an AI model that describes its intended uses, training data, evaluation results, limitations, and license, so buyers and auditors can judge its fitness.
Model Evaluation
- Definition
- Model evaluation is the measurement of how well a model performs against defined criteria, using held-out data, task-specific tests, and safety checks, before and after it is put into service.
Model Serving
- Definition
- Model serving is the software and infrastructure that keeps a trained model loaded in memory, accepts requests over a network, runs inference, and returns results reliably enough for applications to depend on.
Model Size
- Definition
- Model size is the scale of a model, usually expressed as its parameter count, which determines how much memory it needs, how fast it runs, and what hardware can serve it.
Model Weights
- Definition
- Model weights are the learned numerical parameters of a trained model, saved as files that can be loaded to run the model, shared under a license, or fine-tuned further.
Multimodal AI
- Definition
- Multimodal AI is AI that can take in, reason over, or produce more than one type of data, such as text, images, audio, and video, within a single model or system.
Open-Source Model
- Definition
- An open-source model is an AI model whose weights, training code, and ideally training data are published under a license that permits anyone to use, study, modify, and redistribute it.
Open-Weight Model
- Definition
- An open-weight model is an AI model whose trained weights are published for download and self-hosting, usually without the training data, and often under a license that restricts certain uses.
OpenAI-Compatible API
- Definition
- An OpenAI-compatible API is a model-serving interface that accepts the same request and response format as OpenAI's API, so most client libraries and applications can switch models by changing a URL.
Parameter-Efficient Fine-Tuning (PEFT)
- Definition
- Parameter-efficient fine-tuning (PEFT) is a family of methods that adapt a large pre-trained model by training only a small number of added or selected parameters instead of the whole network.
Parameters (Model)
- Definition
- Model parameters are the numerical values, mainly weights and biases, that a model learns during training, and their count is the standard measure of a model's size.
Pre-Trained Model
- Definition
- A pre-trained model is a model that has already been trained on a large general dataset and is distributed for others to use directly or adapt to their own tasks.
Pre-Training
- Definition
- Pre-training is the first and largest phase of building a model, in which it learns general patterns from a huge unlabeled dataset before any task-specific training.
Prompt
- Definition
- A prompt is the text sent to a generative AI model to get a response, including the instructions, the question, any supplied documents or examples, and the required output format.
Prompt Caching
- Definition
- Prompt caching stores the processed form of a prompt prefix, such as a long system prompt or document, so later requests that begin the same way skip that work and respond faster.
Prompt Chaining
- Definition
- Prompt chaining is the technique of breaking a task into a sequence of smaller prompts, where the output of each step becomes the input to the next, with checks between steps.
Prompt Engineering
- Definition
- Prompt engineering is the discipline of designing, testing, and maintaining the instructions given to a generative AI model so that it produces correct, consistent output for a specific business task.
Prompt Injection
- Definition
- Prompt injection is an attack in which instructions hidden in content a language model processes, such as an email, document, or web page, cause it to act against its operator's intent.
Prompt Template
- Definition
- A prompt template is a reusable prompt with placeholders that an application fills with the user's question, retrieved documents, or record data at run time, so every request is built the same…
Quantization
- Definition
- Quantization stores a model's weights in lower-precision numbers such as FP8, INT8, or INT4 so the model needs less memory and runs faster, at the cost of a small loss in accuracy.
Rate Limiting
- Definition
- Rate limiting caps how many requests or tokens a user, application, or team may send to an AI model in a given period, protecting capacity and controlling cost.
Reasoning Model
- Definition
- A reasoning model is a large language model trained to work through a problem in intermediate steps before giving its final answer, trading speed and cost for accuracy on harder tasks.
Reinforcement Learning from Human Feedback (RLHF)
- Definition
- Reinforcement learning from human feedback (RLHF) is a training method that teaches a language model to produce answers people prefer by learning from human rankings of its outputs.
Retrieval-Augmented Generation (RAG)
- Definition
- Retrieval-augmented generation (RAG) is a technique that finds relevant passages from a company's own documents and supplies them to a language model so that it answers from current, verifiable sources.
Self-Supervised Learning
- Definition
- Self-supervised learning trains a model on unlabeled data by hiding part of each example and teaching the model to predict it, which is how large language models and other foundation models are…
Session (AI Conversation)
- Definition
- A session is one continuous conversation between a user or system and an AI model or agent, with its own history, state, and identity, kept separate from every other conversation.
Small Language Model (SLM)
- Definition
- A small language model (SLM) is a language model with a few billion parameters or fewer, sized to run on a single GPU, a CPU, or an edge device while handling narrow,…
Speculative Decoding
- Definition
- Speculative decoding speeds up text generation by letting a small draft model propose several tokens that the large model then verifies in one pass, producing the same output faster.
Streaming (LLM Responses)
- Definition
- Streaming is the delivery of a language model's response token by token as it is generated, so the user sees text appear immediately instead of waiting for the full answer.
System Prompt
- Definition
- A system prompt is the standing set of instructions an application gives a language model before any user message, defining the model's role, rules, tone, tools, and the boundaries of what it…
Temperature and Sampling
- Definition
- Temperature and sampling are the settings that control how a language model picks each next token, from always choosing the most likely option to drawing from a wider spread of possibilities.
Time to First Token (Latency)
- Definition
- Time to first token is the delay between sending a prompt and receiving the first piece of the model's answer, the latency figure users notice most in an interactive AI feature.
Token-Based Pricing
- Definition
- Token-based pricing is the billing model for hosted AI services in which the customer pays for the number of tokens sent to and generated by the model, with input and output often…
Tokenization
- Definition
- Tokenization is the step that splits text into tokens, the word pieces a language model reads and produces, and the unit in which context windows, throughput, and usage-based pricing are measured.
Tokens
- Definition
- Tokens are the units of text a language model reads and writes, usually a word or a fragment of a word. Token counts determine context limits, speed, and the cost of each…
Tokens per Second (Throughput)
- Definition
- Tokens per second measures how fast a language model produces text, either for a single response or as the total output of a server across all users at once.
Tool Use
- Definition
- Tool use is the ability of a language model to call external functions, such as a database query, a search, or an API, and use the returned data to complete a task.
Transfer Learning
- Definition
- Transfer learning reuses a model trained on one large task as the starting point for a related task, so the new task needs far less data and compute.
Transformer
- Definition
- A transformer is a neural network architecture that uses attention to weigh the relationship between every element in a sequence at once, and it is the foundation of modern large language models.
Vision-Language Model (VLM)
- Definition
- A vision-language model (VLM) is a large language model that also accepts images, so it can describe photos, read documents, answer questions about charts, and follow instructions that mix text and pictures.
Zero-Shot Learning
- Definition
- Zero-shot learning is a model's ability to perform a task or recognize a category it was never explicitly trained on, using only a description of what is wanted.
No terms match your search.
Try a broader keyword or another category.