API glossary

A

Agentic search

Agentic search is a retrieval approach in which an AI agent (typically driven by a large language model) actively plans and carries out a multi-step search process rather than performing a single, static lookup. The agent decomposes a query, decides which sources or tools to consult, issues multiple searches, evaluates what it finds, and refines its approach in a loop until it has gathered enough information to answer.

AI web crawler

An AI web crawler is an automated bot that fetches webpages so the retrieved content can support AI systems. Common downstream uses include training large language models (LLMs), powering retrieval indexes for AI answer engines, and refreshing content used to answer user queries. Unlike the narrowest definition of a traditional search crawler (focused on ranking pages and driving click traffic), AI-oriented crawling is usually discussed in terms of how fetched content is reused across workflows for both models and answers.

C

Chunking

Chunking is the process of splitting a large document or body of text into smaller, self-contained segments (called “chunks”) so they can be embedded, indexed, and retrieved individually. In retrieval-augmented generation (RAG) and semantic search, chunking determines the unit of text a system stores and later pulls back as context. This, in turn, directly shapes how relevant and complete the retrieved results are.

Context window

A context window is the maximum amount of text (measured in tokens) that a language model can consider at one time. It spans everything the model is working with in a single conversation: the system instructions, the user’s prompt, any conversation history, any retrieved documents, and the response the model is generating. Once the combined text exceeds the window, the model can no longer “see” the overflow, so the oldest or least relevant material has to be dropped, summarized, or retrieved on demand.

Conversational search

Conversational search is an approach to information retrieval in which a person finds information through a back-and-forth, natural-language dialogue rather than a single keyword query. The system interprets each request in the context of the conversation so far, remembering previous turns, resolving follow-up questions, and refining results as the exchange continues. This allows people to search as though they were talking with a knowledgeable human.

Crawl budget

Crawl budget is the amount of crawling a search engine or bot will do on a given site within a period of time (in effect, the set of URLs it both can and wants to fetch). Often, this entails the combination of two factors: crawl capacity (how much a site’s server can handle without slowing down) and crawl demand (how much the crawler actually wants to fetch, based on a site’s popularity and how often its content changes). For large or frequently updated sites, crawl budget determines how quickly new and changed pages get discovered.

E

Entity extraction

Entity extraction is the natural-language-processing (NLP) task of automatically identifying meaningful things mentioned in text and labeling each one by type. These things can include people, organizations, places, dates, products, and other domain-relevant entities. In many systems, named entity recognition (NER) is a core part of entity extraction, but entity extraction can also be broader depending on the schema and use case.

F

Function calling

Function calling is a capability that lets a large language model (LLM) request the use of an external tool, API, or piece of code by producing a structured, machine-readable call instead of a plain-text reply. The model does not run the code itself; it decides which function to invoke and with what arguments, while the application then executes that function, and the result is handed back to the model to incorporate into its final answer. Function calling is often used interchangeably with “tool use” or tool calling, though strictly speaking it refers to developer-defined functions, a subset of the broader tool-calling pattern.

H

Hallucination

A hallucination is output from an AI model (especially a large language model) that is false, fabricated, or ungrounded, yet presented fluently and confidently as if it were true. (“Ungrounded” in this case meaning the output is unsupported by the model’s training data or the sources it was given.) Hallucinations range from invented citations and fake statistics to plausible-sounding details that no source actually supports, and can also arise when a model misreads, misattributes, or overgeneralizes from stale or partial sources.

Hybrid search

Hybrid search is a retrieval approach that combines keyword (lexical) search and semantic (vector) search as two complementary methods, with results then merged into a single ranked list. By running both at once, hybrid search combines exact-term matches (which keyword search is good at) with meaning-based matches (which semantic search is good at). In mixed-query workloads, this often improves retrieval robustness compared with using either method alone.

I

Intent classification

Intent classification is the natural-language-processing task of determining what a user is trying to accomplish from their input, by assigning a query or message to one of a set of predefined intent categories. It answers the question “what does this person want?” (for example, whether a search query is informational, navigational, or transactional, or whether a chatbot message is a greeting, a complaint, or a request to make a reservation).

K

Knowledge cutoff

A knowledge cutoff is the point in time after which a language model has no training data, and therefore no built-in awareness of events, facts, or changes that occurred later. Because a model’s knowledge is effectively frozen at training time, anything that happened after its cutoff is not part of its built-in knowledge unless newer information is supplied at runtime through external context, retrieval, or tools.

Knowledge graph

A knowledge graph is a structured representation of knowledge as a network of entities (e.g. people, places, products, or concepts) and the relationships that connect them. Instead of storing information as isolated text, a knowledge graph records facts as nodes (the entities) linked by edges (the relationships), often with types and attributes attached. This allows both machines and people to more easily understand how things relate to one another.

L

Large language model (LLM)

A large language model (LLM) is an artificial-intelligence model trained on vast amounts of text to understand and generate human language. It learns the statistical patterns of language so it can predict and produce the next piece of text in a sequence, which (at sufficient scale) lets it answer questions, summarize, translate, write code, and hold a conversation.

Latency

Latency is the delay between a request and the response to it—basically how long a user or system waits for results. In AI search and language-model applications, it usually means the time from sending a query to receiving an answer, and it is a core measure of how responsive and usable a system feels.

P

Programmable search engine

A programmable search engine is a search engine that developers configure and query through code rather than only through a consumer website. This allows developers to build search directly into their own applications, control how it behaves, and receive results in a structured, machine-readable form. The term “programmable search engine” more broadly describes any search engine exposed for programmatic use, typically through a search API.

R

Rate limiting

Rate limiting is the practice of capping how many requests a client (often an app on a computer or mobile device) can make to a service within a set period of time. When a client exceeds the limit, the service rejects or delays further requests (usually returning an HTTP 429 Too Many Requests response) until the window resets. Rate limiting protects a service from overload, ensures fair access among users, and enforces the usage tiers that many APIs sell.

Reranking

Reranking is a second-stage process that reorders an initial list of search results to put the most relevant ones first. A fast first-stage retriever pulls a broad set of candidate results, and the reranker then scores each candidate more carefully against the query (using a slower but more accurate model) and sorts them by that score. Reranking is the precision step that comes after the recall step.

Retrieval score

A retrieval score is a number that a search or retrieval system assigns to each candidate result, expressing how relevant or similar the result is to the original user query. The system ranks results by this score, putting the highest-scoring items first, and can use it to decide which results are strong enough to keep. Depending on the method, the score may measure keyword overlap, semantic similarity, or a blend of both.

Retrieval-augmented generation (RAG)

Retrieval-augmented generation (RAG) is a technique that improves a language model’s answers by retrieving relevant information from an external source, and supplying this information to the model as context before it generates a response. Rather than relying exclusively on what it learned in training, the model answers using documents fetched at query time, so its output can be current, specific, and grounded in real sources.

S

Scraper

A scraper is a program that automatically extracts data from websites (including search engine results pages), usually by downloading their pages and parsing the HTML written for human browsers; scrapers can turn content meant for human readability into structured data a computer program can store and reuse. Scrapers have many legitimate uses (including price comparisons, research, archiving, and system monitoring), but depending on what is scraped and how, they can raise legal, technical, and policy issues. In the search context, a scraper pulls results by loading and parsing a search engine’s results pages—an approach that many major providers restrict or prohibit through terms of service and anti-automation policies.

Search engine results page (SERP)

Semantic search is an approach to search that matches results by meaning and intent, not just exact keyword overlap. In modern systems, this is commonly implemented by representing queries and content as embeddings (numeric vectors that capture semantic relationships), then retrieving items with the most similar vectors. That lets a query like “best wine for seafood” match content such as “good with fish,” even when exact terms differ.

Semantic search

A search engine results page (SERP) is the page a search engine returns in response to a query. On modern search engines, it is usually a blend of elements rather than a simple list: ranked organic (unpaid) links, paid ads, and SERP features such as featured snippets or knowledge panels that answer parts of the query directly. On some queries and engines, AI-generated overviews may also appear, often near the top of the page.

SERP features

SERP features are enriched modules on a search engine results page that go beyond a standard list of organic links. Common examples include featured snippets, knowledge panels, “People Also Ask” boxes, image and video carousels, local map packs, top-stories blocks, shopping modules, and AI-generated overviews.

Source attribution

Source attribution is the practice of identifying and crediting the original sources behind a piece of information, so a reader can see where a claim originated. In AI search and answer engines, it means pairing generated answers with citations (links or references back to the specific pages and passages the answer draws on) rather than presenting conclusions with no traceable origin.

Structured output

Structured output is output from a language model that follows a specific, predefined format (most often JSON matching a defined schema) instead of free-form prose. By constraining response shape, structured output makes model output easier for software to parse and act on without brittle cleanup logic.

T

Tokenization

Tokenization is the process of breaking text into smaller units called tokens, the pieces a system actually processes. For a language model, tokens are usually sub-word fragments that the model reads as numbers rather than whole words. In classic search, tokenization instead splits text into the terms used to build and query an index. Either way, it is the step that turns raw text into units a system can work with.

Tool calling

Tool calling is the capability that lets a large language model (LLM) use external tools like APIs, functions, databases, or services. By producing a structured request to invoke a tool, an application or the model provider can then execute a task and return a result to the model. This is how a model moves beyond generating text to fetching live data and taking actions.

V

Vector database

A vector database is a database built to store and search embeddings—the high-dimensional numeric vectors that represent the meaning of text, images, or other data. Instead of matching records by exact values like a traditional database, a vector database finds the records whose vectors are most similar to a query, which is what makes meaning-based (semantic) search possible at scale.

Vector embeddings

A vector embedding (or just an embedding) is a list of numbers that represents the meaning of a piece of data (such as text, image, or audio) as a point in a high-dimensional space. An embedding model produces these vectors so that items with similar meaning land close together, which lets software compare meaning by measuring distance between vectors rather than matching words.

W

Web grounding

Web grounding is the practice of basing a language model’s answers on live web content retrieved at query time, so its responses reflect current, verifiable sources rather than only what it learned in training. A grounded model searches the Web for a query, pulls in the relevant pages, and generates its answer from that retrieved content—often citing the sources it used.

Web search API

A web search API is a service that lets software query a search engine’s index of the Web and receive the results as structured data, rather than through a browser or by scraping a search results page. A program issues a request containing a query and parameters and gets back machine-readable results (typically URLs, titles, and text snippets, alongside news, images, videos, and richer content) ready to use in an AI application.

Z

Zero-click search

Zero-click search is a search session that ends without the user clicking an outbound result. This can happen because the results page provides enough information directly, because the user abandons the query, or because they reformulate and search again.