Glossary
AI builder glossary
98 essential terms for building AI-powered apps, explained in plain English and linked to the lessons that go deeper.
A
A/B testing (for AI features)
Comparing two versions of an AI feature on randomly split live traffic to see which drives better outcomes, such as acceptance, resolution, or retention.
AI agent
A system where a model pursues a goal in a loop, choosing its own next steps and tools, checking results, and adjusting until it's done.
AI SDK
Vercel's open-source TypeScript toolkit that gives many model providers one API for text, streaming, tools, and structured output, plus React UI hooks.
AI UX
The design practice for interfaces built on probabilistic model output: setting expectations, showing sources, handling errors, keeping users in control.
AI workflow
A multi-step process where your code fixes the order of steps and calls a model for specific ones, making it more predictable than an agent.
B
C
Chunking
Splitting documents into smaller passages before embedding and indexing them, so retrieval can return the specific part that answers a question.
Citations (grounded answers)
References in an AI answer that point each claim to the specific source passage supporting it, so users and your code can verify it.
Constrained decoding
A generation technique that restricts which tokens a model may produce at each step so its output is guaranteed to match a schema or grammar.
Content credentials (C2PA)
Signed provenance metadata, defined by the C2PA standard, that records how a piece of media was created or edited, including whether AI generated it.
Content moderation
Checking user inputs and model outputs against a content policy, with classifiers or an LLM, to block, flag, or route harmful or off-policy material.
Context engineering
Deciding and assembling everything a model sees on each call, from instructions and history to retrieved data and tool results, within a token budget.
Context window
The maximum number of tokens a model can take in a single request, covering instructions, history, documents, tool definitions, and the reply it writes.
Contextual retrieval
A retrieval technique where a model writes a short situating note for each chunk, prepended before embedding and keyword indexing, so chunks stand alone.
Cosine similarity
A measure of how closely two vectors point in the same direction, the standard way to score how similar two embeddings are in meaning.
D
Denial-of-wallet attack
Abuse that drives up your AI spend, through request floods, huge inputs, or free-tier farming, rather than taking your service offline.
Distillation
Training a small model to imitate a larger one on a specific task, using the large model's outputs as labels, to get similar quality at lower cost.
Drift
A change over time in the inputs an AI feature receives or the outputs it produces, which quietly makes past evals less representative.
E
Effort (reasoning effort)
A request setting that tells a model how much reasoning and work to put into a response, trading quality against latency and output-token cost.
Embedding
A list of numbers that represents the meaning of text or images, so that similar content lands close together and can be compared mathematically.
Error analysis
Systematically reading a sample of real AI outputs, noting what went wrong, and grouping failures into categories to decide what to fix and test next.
EU AI Act
The EU's risk-based law for AI systems, which bans some uses, regulates high-risk ones, and requires disclosure for chatbots and AI-generated content.
Eval-driven development
Building AI features by defining evals for the desired behavior first, then changing prompts, models, or code until they pass, and gating releases on them.
Evals
Repeatable tests that score an AI feature's outputs against defined criteria, so you can measure quality, catch regressions, and compare prompts or models.
Extended thinking
Anthropic's name for Claude's thinking phase, where the model works through a problem in hidden reasoning tokens before writing its final answer.
F
Faithfulness
A quality measure of whether every claim in a generated answer is supported by the context the model was given, rather than added or distorted.
Few-shot prompting
Including a handful of worked input-and-output examples in a prompt so the model copies the format, tone, and judgment you want instead of guessing.
Fine-tuning
Further training of an already-trained model on examples you supply, changing its weights so it handles a specific task, format, or style more reliably.
Frontier model
One of the most capable models available at a given time, typically from a major lab, offering top quality at higher prices and often more latency.
Full-text search (BM25)
Keyword search over word indexes that ranks documents by query-term matches, with BM25 as the classic relevance formula most search engines use.
G
Generative UI
An interface pattern where the model chooses which of your components to render, and with what data, instead of replying only with text.
Golden dataset
A curated set of real inputs paired with known-good outputs or criteria, used as the fixed benchmark your AI feature is evaluated against.
Grounding
Anchoring a model's answer in specific source material you supply, such as documents, records, or tool results, rather than its general training data.
H
Hallucination
Output that sounds confident and plausible but isn't true or isn't supported by the input, like an invented policy, number, quote, or record.
HNSW index
Hierarchical Navigable Small World: a graph-based vector index that finds approximate nearest neighbors fast by walking layered links.
Human-in-the-loop
A design where a person reviews, edits, or approves AI output before it takes effect, used where errors are costly or accuracy isn't proven yet.
Hybrid search
Running keyword search and vector search together and merging their results, so queries match on both exact terms and overall meaning.
I
J
Jailbreak
A prompt crafted to make a model ignore its safety training or your product's rules, often through role-play or slow multi-turn escalation.
JSON Schema
A standard vocabulary for describing the shape of JSON data, used by model APIs to define structured outputs and the arguments that tools accept.
K
L
Large language model (LLM)
A neural network trained on vast amounts of text that generates output token by token, powering chat, summaries, extraction, and search features.
LLM gateway
A proxy layer between your app and model providers that offers one API for many models, plus fallbacks, key management, budgets, and usage logging.
LLM observability (tracing)
Recording each model call with its prompt, output, tokens, cost, latency, and triggering request, so you can debug, evaluate, and audit AI features.
LLM-as-judge
Using a language model to grade another model's output against written criteria, for qualities like tone or faithfulness that code can't check.
Local LLM
A language model running on hardware you control, such as a laptop or your own server, with no third-party API involved in generating the output.
Long context
Using a model's very large context window to send whole documents or datasets in one request instead of retrieving a few passages.
LoRA
Low-rank adaptation: a fine-tuning method that freezes a model's weights and trains a small add-on adapter, cutting memory needs and output size.
M
Max tokens
The required cap on how many tokens a model may generate in one response, which limits output length, cost, and runtime but can truncate answers.
Mean reciprocal rank (MRR)
A retrieval metric averaging one over the rank of the first relevant result across queries, rewarding search that puts the right answer first.
Model cascade
A routing pattern that sends each request to a cheap model first and escalates to a stronger model only when the cheap answer fails a quality check.
Model Context Protocol (MCP)
An open standard for connecting AI applications to external tools and data through servers, so one integration works across many AI clients.
Model routing
Choosing which model handles each request at run time, using rules or a classifier, so easy requests go to cheap models and hard ones to strong models.
Multimodal model
A model that can take in, and sometimes produce, more than one kind of data, such as text, images, PDFs, audio, or video, within a single request.
N
O
OCR (optical character recognition)
Converting images of text, such as scans or phone photos, into machine-readable characters that your app can search, store, and process.
On-device AI
Running a model directly on the user's phone, laptop, or browser instead of a server, trading model size for privacy, speed, and zero cost per call.
Open-weight model
A model you can download because its trained weights are public, so you can run, fine-tune, and host it yourself instead of calling a vendor's API.
P
Pairwise evaluation
Judging which of two outputs for the same input is better, rather than scoring each alone, to decide whether a new version beats the current one.
pgvector
An open-source Postgres extension that adds a vector column type, distance operators, and approximate nearest neighbor indexes for embeddings.
PII redaction
Removing or masking personal data, such as names, emails, phone numbers, or account numbers, from text before it's sent to a model, logged, or stored.
Precision and recall
Two complementary metrics for classifiers and search: precision is the share of flagged items that were right; recall is the share of right items flagged.
Prompt caching
A provider feature that reuses already-processed prompt prefixes across requests, cutting the cost and latency of long, repeated instructions and context.
Prompt chaining
Splitting a task into a sequence of focused model calls, where your code passes each step's output to the next, instead of one prompt that does everything.
Prompt engineering
The practice of writing, testing, and refining the instructions and examples a model receives so a feature produces reliable, useful output at scale.
Prompt injection
An attack where text the model reads, from a user or from documents, emails, or web pages, carries instructions that override what the developer intended.
Prompt template
A reusable prompt with placeholders that your code fills with per-request data, such as a ticket, user details, or retrieved documents, at runtime.
Q
Quantization
Storing a model's weights at lower precision, such as 4 or 8 bits instead of 16, so it uses less memory and runs faster at a small cost in quality.
Query rewriting
Using a model to turn a user's raw question into one or more better search queries before retrieval, such as standalone or expanded versions.
R
Rate limiting
Capping how many requests or tokens can be used in a time window, covering both limits providers set on you and limits you set on your own users.
Reasoning model
A model that works through a problem privately before answering, improving hard multi-step tasks at the cost of extra latency and output tokens.
Recall@k
A retrieval metric that measures the share of queries, or relevant items, found within the top k results, showing whether search surfaces what's needed.
Reciprocal rank fusion (RRF)
A method for merging ranked result lists by summing one divided by a constant plus each item's rank, so no score normalization is needed.
Red-teaming
Deliberately attacking your own AI feature with injections, jailbreaks, data-leak attempts, and cost abuse to find failures while they're cheap to fix.
Reranking
A second retrieval pass where a slower, more precise model rescores a shortlist of candidates against the query and reorders them before generation.
Retrieval-augmented generation (RAG)
A pattern that retrieves relevant passages from your own data and includes them in the prompt, so the model answers from current, specific sources.
S
Semantic caching
Reusing a stored answer when a new request means the same thing as an earlier one, matched by embedding similarity rather than exact text.
Semantic search
Search that matches on meaning rather than exact words, typically by embedding the query and finding the stored content with the nearest vectors.
Server-sent events (SSE)
A simple HTTP protocol for pushing a one-way stream of text events from server to client, widely used to deliver streamed model output.
Shadow mode
Running a candidate prompt or model on real production inputs in the background, storing its output unseen, to compare it with the live version.
Speaker diarization
Splitting a transcript by who is speaking, labeling segments as Speaker A, Speaker B, and so on, so multi-person recordings become usable.
Speech-to-speech model
A model that takes spoken audio in and produces spoken audio out directly, without converting to text in between, for low-latency voice conversations.
Speech-to-text (STT)
Converting spoken audio into written text, the building block behind voice notes, call transcripts, dictation, captions, and voice interfaces.
Stop reason
The field in a model response that says why generation ended, such as finishing normally, hitting the token cap, calling a tool, or refusing.
Streaming
Delivering a model's output incrementally as it's generated, so users see text appear within moments instead of waiting for the full response.
Structured outputs
An API feature that makes a model return data matching a schema you define, so your code gets typed fields instead of free text to parse.
System prompt
The standing instructions sent with every request to define a feature's role, rules, output format, and boundaries, kept separate from user input.
T
Temperature
A legacy sampling setting that made output more predictable or more varied, now rejected or discouraged on many current reasoning models.
Text-to-speech (TTS)
Generating spoken audio from text, used for read-aloud features, spoken replies, accessibility, and the speaking half of a voice interface.
Time to first token (TTFT)
The delay between sending a request and receiving the first token of output, the latency number that most shapes how fast an AI feature feels.
Token
The chunk of text a model processes as a single unit, often a word fragment, which sets how much fits in a request and how much each call costs.
Tool use (function calling)
Letting a model request calls to functions you define, such as a database lookup or API action, with your code running them and returning the results.
U
Unit economics (of AI features)
The cost and revenue of an AI feature per unit, such as per request, per user, or per outcome, showing whether each customer and feature makes money.
Usage-based pricing
Charging customers according to how much they use, such as actions, documents, minutes, or credits, so revenue moves with the variable cost of AI features.
V
Vector database
A database or index built to store embeddings and quickly find the vectors nearest a query, usually alongside metadata used for filtering.
Vision-language model
A model that takes images and text together as input and answers in text, so it can read screenshots, photos, charts, and scanned documents.
Voice activity detection (VAD)
Detecting when someone is speaking in an audio stream, used to decide when a user's turn has ended and to skip silence before transcription.