Generative AI Engineering
Understand the model, input, context, generation, and operation terms used when working with generative AI and large language models.
12 matching terms
Models
Generative AI
AI that creates new text, images, audio, code, or other content from learned patterns.
Use the term when distinguishing content-producing systems from classification or prediction systems.
Use generative AI to draft three product descriptions, then review them for factual accuracy.Models
Foundation model
A broadly trained model that can be adapted to many downstream tasks.
Use it when discussing the base model behind a chatbot, agent, or specialized application.
Select a foundation model, then add retrieval and task-specific instructions for customer support.Models
Large language model (LLM)
A model trained on large text collections to understand and generate language sequences.
Use it for text generation, summarization, extraction, translation, and conversational tasks.
Ask the LLM to summarize the meeting transcript into decisions, owners, and deadlines.Models
Multimodal model
A model that can process or produce more than one data type, such as text and images.
Use it when a task combines screenshots, documents, audio, video, or text.
Provide a chart image and ask the multimodal model to explain the visible trend.Input and context
Token
A unit of text processed by a language model; it may be a word, part of a word, or punctuation.
Use token counts to estimate context capacity, latency, and usage cost.
Trim repeated instructions when the input approaches the model's token limit.Input and context
Context window
The maximum amount of input and generated output a model can consider in one request.
Use it to decide whether documents should be summarized, chunked, or retrieved selectively.
A larger context window does not guarantee that every included detail will be used correctly.
Retrieve the five most relevant passages instead of placing the entire archive in the context window.Input and context
System instruction
A high-priority instruction that defines the model's role, rules, and response behavior.
Use it to set durable behavior shared by many user requests.
You are a support assistant. Cite the supplied policy and say when evidence is missing.Operation
Model weights
Learned numerical values that encode patterns acquired during training.
Use the term when comparing base, fine-tuned, quantized, or open-weight models.
Load the approved model weights in the isolated inference environment.Operation
Inference
The process of running a trained model on new input to produce an output.
Use it when discussing serving, latency, hardware, or per-request cost.
Batch non-urgent inference requests to improve throughput.Operation
Knowledge cutoff
The latest point in time substantially represented in a model's training knowledge.
Check it before asking about recent events, prices, policies, or software versions.
Use live search for events that occurred after the model's knowledge cutoff.Generation
Sampling
The method used to choose each next token from the model's probability distribution.
Adjust sampling when you need more deterministic or more varied outputs.
Use conservative sampling for policy extraction and broader sampling for ideation.Generation
Temperature
A generation setting that generally controls how concentrated or varied token choices are.
Lower it for repeatable structured tasks and raise it cautiously for exploration.
The exact effect and supported range vary by model and provider.
Set a low temperature for converting invoices into a fixed JSON schema.