Home / Semantic search glossary

Semantic search glossary

This glossary defines 81 terms used across the Semantic Search Guide, from embeddings and vector indexes to ranking metrics and reranking. Each definition is short and links to the guide that covers the term in full. Terms are listed alphabetically, and every entry has its own anchor, so a definition can be linked directly, for example /glossary/#embedding.

A B C D E F H I K L M N P Q R S T V W

A

Agentic retrieval: A retrieval approach in which a language model agent decides which searches to run, which sources to query and when it has enough evidence, instead of following one fixed query. Full guide: Agentic retrieval and agentic search.

ANN recall: The share of the true nearest neighbors, found by exact search, that an approximate index actually returns. It measures what the index misses, not whether results are relevant. Full guide: ANN recall.

Approximate nearest neighbor (ANN) search: Vector search that uses an index to find vectors that are very likely the closest to a query, trading a small loss of recall for a large gain in speed over checking every vector. Full guide: Approximate nearest neighbor search.

Asymmetric search: Semantic search where queries and documents differ in form, such as short questions against long passages. Many embedding models handle it with separate query and document prefixes or encoders. Full guide: Asymmetric vs symmetric semantic search.

B

BERT: A transformer encoder model, released by Google in 2018, that reads text in both directions. It is the basis of many rerankers and sentence embedding models used in search. Full guide: Transformers and BERT in search.

Bi-encoder: A model that encodes the query and each document separately into vectors, so document vectors can be computed in advance and compared quickly at query time. Full guide: Bi-encoders.

Bitmap filter: A filter stored as a bit array with one bit per document, set when the document matches. Combining filters becomes fast bitwise AND and OR operations. Full guide: Bitmap filters.

BM25: A keyword ranking function that scores documents by how often query terms appear, with diminishing returns for repeats, weighted by term rarity and adjusted for document length. Full guide: The BM25 algorithm.

C

Chunking: Splitting long documents into smaller passages before embedding, so each vector represents one focused piece of text. Chunk size and overlap affect both recall and cost. Full guide: Chunking for semantic search.

ColBERT and late interaction: A retrieval model that keeps one vector per token and scores a document by matching each query token to its most similar document token, a pattern called late interaction. Full guide: ColBERT and late interaction.

Cosine similarity: A similarity measure equal to the cosine of the angle between two vectors. It ignores vector length, and for normalised vectors it gives the same ranking as dot product. Full guide: Cosine vs dot product vs Euclidean distance.

Cross-encoder: A model that reads the query and a document together and outputs one relevance score. It is more accurate than a bi-encoder but too slow to score a whole collection, so it is used for reranking. Full guide: Cross-encoders.

Curse of dimensionality: The set of effects that make high-dimensional spaces behave unlike everyday space, such as distances between random points becoming similar, which complicates nearest neighbor search. Full guide: Curse of dimensionality.

D

Dense retrieval: Retrieval that represents queries and documents as dense embedding vectors and ranks documents by vector similarity. Full guide: Dense retrieval.

Dense vector: A vector in which almost every dimension holds a non-zero value, as produced by embedding models. Individual dimensions have no human-readable meaning. Full guide: Dense vectors.

Dot product: The sum of the products of two vectors' matching dimensions. It rewards both direction and length, and equals cosine similarity when vectors are normalised. Full guide: Cosine vs dot product vs Euclidean distance.

E

Embedding: A list of numbers produced by a model to represent text, images or other data, placed so that items with similar meaning end up close together. Full guide: What are embeddings.

Embedding dimensions: The number of values in an embedding vector. More dimensions can hold more detail but cost more storage, memory and compute. Full guide: Embedding dimensions.

Embedding model: The model that turns data into embeddings. Its training data and objective decide what counts as similar, so choosing it is one of the largest decisions in semantic search. Full guide: Choosing an embedding model.

Euclidean distance: The straight-line distance between two vectors, also called L2 distance. Smaller values mean more similar items. Full guide: Cosine vs dot product vs Euclidean distance.

Exact search: Nearest neighbor search that compares the query with every vector, also called brute-force or flat search. It always finds the true nearest neighbors but slows as the collection grows. Full guide: Exact vs approximate search.

F

Faceted search: Search that shows counts for attribute values, such as brand or price band, and lets users narrow results by selecting them. Full guide: Faceted semantic search.

Filter selectivity: The share of a collection that passes a filter. Very restrictive filters can hurt approximate vector search, so engines often change strategy based on selectivity. Full guide: Filter selectivity.

Filtered ANN search: Approximate nearest neighbor search restricted to objects that match a filter, which needs special handling to keep recall high when few objects match. Full guide: Filtered ANN search.

Fine-tuning an embedding model: Further training an existing embedding model on pairs from your own domain so that it places your relevant query and document pairs closer together. Full guide: Fine-tuning embedding models.

Full-text search: Search over the words inside documents using an inverted index and text analysis such as tokenization, lowercasing and stopword handling. Full guide: Full-text search.

H

HNSW: Hierarchical Navigable Small World, a graph-based vector index. It links each vector to near neighbors across several layers and finds close vectors by walking the graph. Full guide: Semantic indexing and vector indexes.

Hybrid search: Search that runs keyword and vector retrieval for the same query and fuses the two ranked lists, so exact terms and meaning both count. Full guide: What is hybrid search.

HyDE: Hypothetical Document Embeddings: a language model writes a plausible answer to the query, and that text is embedded and used for search instead of the short query. Full guide: HyDE.

I

Index freshness: How quickly new or changed data becomes searchable after it is written. Full guide: Real-time indexing and index freshness.

Information retrieval: The field that studies finding relevant documents or information in large collections, and the source of most search ranking and evaluation methods. Full guide: What is information retrieval.

Inverted index: A data structure that maps each term to the list of documents containing it. It is the core of keyword search. Full guide: The inverted index.

K

Keyword search: Search that matches the literal terms in a query against terms in documents, also called lexical search. It is precise for names and codes and misses different wording. Full guide: What is keyword search.

Knowledge graph: A network of entities and the typed relationships between them, used in search to answer factual questions and connect related things. Full guide: Knowledge graphs in search.

L

Latent semantic indexing: An early semantic method that applies singular value decomposition to a term-document matrix, so documents can match related terms they do not contain. Full guide: Latent semantic indexing.

LLM reranking: Using a large language model to judge and reorder retrieved results, scoring them one at a time, in pairs or as a whole list. Full guide: LLM reranking.

M

Matryoshka embeddings: Embeddings trained so that the first part of the vector is useful on its own, letting you shorten vectors to save storage with a modest loss of quality. Full guide: Matryoshka embeddings.

Maximal marginal relevance: A reranking method that picks each next result by balancing relevance to the query against similarity to results already chosen, to reduce near-duplicates. Full guide: Maximal marginal relevance.

Mean reciprocal rank: An evaluation metric that averages one divided by the rank of the first relevant result across queries. It rewards putting a right answer near the top. Full guide: Mean reciprocal rank.

Metadata filtering: Restricting search results by structured fields, such as date, category or price, alongside vector or keyword relevance. Full guide: Metadata filtering.

MTEB: The Massive Text Embedding Benchmark, a public leaderboard that scores embedding models across many tasks, including retrieval. Full guide: The MTEB leaderboard.

Multi-query retrieval: Generating several rewordings of one question, searching with each and merging the results to raise recall. Full guide: Multi-query retrieval.

Multi-tenancy: Serving many separate customers or groups from one search system while keeping each tenant's data isolated. Full guide: Multi-tenant semantic search.

Multi-vector search: Representing each object with several vectors, such as one per token, passage or field, and searching across them. Full guide: Multi-vector search.

Multimodal embeddings: Embeddings that place different data types, such as text and images, in one shared space so a text query can find images and the reverse. Full guide: Multimodal embeddings.

N

nDCG: Normalized discounted cumulative gain, an evaluation metric that rewards relevant results more the higher they rank and supports graded relevance labels. Full guide: nDCG.

P

Pre-filtering and post-filtering: Two ways to combine filters with vector search: apply the filter before the vector search, or run the vector search first and drop non-matching results afterwards. Full guide: Pre-filtering vs post-filtering.

Precision@k: The share of the top k results that are relevant. Full guide: Precision@k.

Prompt caching: A language model provider feature that reuses computation for a repeated prompt prefix. Unlike semantic caching, it needs an exact match and still generates a new answer. Full guide: Semantic caching vs prompt caching.

Q

Quantization: Compressing vectors by storing values with fewer bits, such as scalar, binary or product quantization, to cut memory at some cost in accuracy. Full guide: Embedding and vector quantization.

Query decomposition: Splitting a complex question into simpler sub-questions that are searched separately and then combined. Full guide: Query decomposition.

Query expansion: Adding related terms or phrasings to a query so it matches more relevant documents. Full guide: Query expansion.

Query intent classification: Detecting what a searcher wants, such as a specific item, a topic or a count, so the query can be routed to the best retrieval method. Full guide: Query intent classification.

Query rewriting: Reformulating a query, often with a language model, into a clearer or self-contained version before searching. Full guide: Query rewriting.

R

Recall@k: The share of all relevant documents for a query that appear in the top k results. Full guide: Recall@k.

Reciprocal rank fusion: A method for merging ranked lists that gives each document a score of one divided by a constant plus its rank in each list, then sums the scores. It uses ranks only, not raw scores. Full guide: Reciprocal rank fusion.

Relative score fusion: A method for merging ranked lists that rescales each list's scores to a common range and adds them, keeping information about how far apart results are. Full guide: Relative score fusion.

Relevance judgments: Labels saying which documents are relevant to which queries, used as the answer key when evaluating search. Full guide: Search test sets and relevance judgments.

Relevance score: The number a search system assigns to each result to rank it. Its scale depends on the method, so scores from different methods are not directly comparable. Full guide: Relevance scores.

Reranking: A second stage that rescores the top results from a first search with a more accurate but slower model. Full guide: What is reranking.

Retrieval-augmented generation (RAG): A pattern in which a system retrieves relevant passages and gives them to a language model as context for writing an answer. Full guide: Semantic search vs RAG.

S

Score normalization: Rescaling scores from different retrieval methods to a common range so they can be combined fairly. Full guide: Score normalization in hybrid search.

Semantic caching: Reusing a stored answer when a new query means the same as an earlier one, judged by embedding similarity rather than an exact text match. Full guide: What is semantic caching.

Semantic search: Search that matches the meaning of a query rather than only its exact words, usually by comparing embeddings. Full guide: What is semantic search.

Sentence embedding: One vector for a whole sentence or passage, as opposed to one vector per word. Full guide: Word vs sentence embeddings.

Sentence-BERT: A method that fine-tunes BERT-style models in a siamese setup to produce sentence embeddings that can be compared with cosine similarity. Full guide: Sentence-BERT.

Similarity search: Finding the items most similar to a query item, measured by a distance or similarity function over their vectors. Full guide: Similarity search.

Similarity threshold: The minimum similarity at which a semantic cache treats a new query as a match for a stored one. Too low returns wrong answers; too high misses reuse. Full guide: Similarity thresholds for caching.

Sparse vector: A vector with one dimension per vocabulary term in which most values are zero, as used in keyword and learned sparse retrieval. Full guide: Sparse vectors.

SPLADE: A learned sparse retrieval model that uses a transformer to weight query and document terms and add related terms, so an inverted index can match by meaning. Full guide: SPLADE and learned sparse retrieval.

Stemming and lemmatization: Reducing words to a shared base form before indexing: stemming strips suffixes by rule, while lemmatization looks up the dictionary form. Full guide: Stemming and lemmatization.

T

TF-IDF: Term frequency times inverse document frequency, a weighting that scores a term higher when it is frequent in a document and rare across the collection. Full guide: TF-IDF.

Tokenization: Splitting text into units, called tokens, that a model or index processes. Embedding models and keyword indexes tokenize differently. Full guide: Tokenization vs embedding.

Top-k retrieval: Returning the k highest-scoring results for a query. Full guide: Top-k retrieval.

Typo tolerance: Handling misspelled queries, for example with fuzzy matching, character n-grams or spelling correction, so they still find the right results. Full guide: Typo tolerance and spelling correction.

V

Vector database: A database built to store vectors with their data and run fast similarity search, usually with filtering and other search features. Full guide: Choosing a vector database for semantic search.

Vector index: A data structure, such as an HNSW graph or a set of clusters, that speeds up nearest neighbor search over vectors. Full guide: Semantic indexing and vector indexes.

Vector search: Search that converts a query into a vector and returns the stored vectors closest to it. Full guide: Vector search.

Vector space model: The classic retrieval model that represents queries and documents as vectors of term weights and ranks documents by their similarity to the query. Full guide: Vector space model.

Vocabulary mismatch: The problem that users and documents often use different words for the same thing, so keyword search misses relevant results. Full guide: The vocabulary mismatch problem.

W

Word2vec and GloVe: Two influential methods for learning word embeddings from large text collections, so that words used in similar contexts get similar vectors. Full guide: Word2vec and GloVe.