AI Glossary / Graphs and symbolic AI / Intermediate
Similarity Search
Definition
Similarity search retrieves items that are close to a query under a specified comparison measure. In vector search, items and queries are represented as numeric vectors, and the system returns high-similarity or low-distance matches. The result may be the best k matches or all items satisfying a threshold, depending on the task.
What it means
An ordinary exact lookup asks whether a key matches precisely. Similarity search asks which stored items resemble a query according to the selected representation. This supports situations where wording, appearance, or other features vary even when items are relevant to the same need.
Similarity is defined by the system rather than discovered independently of its design. An embedding model may emphasize topic while overlooking a critical difference in date or product version. A mathematically close item can therefore be irrelevant to the user's actual question. Retrieval quality needs evaluation against the task, not just inspection of vector scores.
How it works
Represent stored items and queries in compatible vector spaces. Select a measure, such as Euclidean distance, cosine similarity, or inner product, that fits the representation. Exact search identifies the best matches under that measure. Approximate search uses an index or other shortcuts to reduce search cost while accepting that some exact neighbors may be missed.
Apply necessary metadata restrictions, retrieve candidates, and optionally rerank them using additional information. Track the embedding version and preprocessing rules. A vector database or search library can organize these operations, but its index does not make a poor representation semantically reliable.
Three practical examples
- A help center can retrieve passages whose meaning resembles a user's question even when they use different words.
- An image catalog can find candidate pictures with similar visual features for a designer to inspect.
- A record-cleaning workflow can retrieve near-matching names or descriptions for duplicate review without automatically merging them.
Think of it this way
Think of asking a librarian for books resembling an example you enjoyed. The answer depends on whether “resembling” means topic, writing style, or intended audience. The analogy is limited because vector search follows a numerical representation and ranking rule; it does not independently understand which resemblance matters to you.
Worked example
Consider an original two-dimensional example with query q = (1, 0) and three stored vectors: A = (2, 0), B = (0.8, 0.6), and C = (0, 1). All are nonzero, so cosine similarity is defined.
Cosine similarity divides the dot product by the product of vector lengths. A has cosine 1 with q. B has length 1 and cosine 0.8. C has cosine 0. Thus cosine ranking is A, B, C.
Raw Euclidean distance produces a different order. The distance to A is 1. The distance to B is sqrt((1 − 0.8)^2 + 0.6^2) = sqrt(0.4), about 0.6325. The distance to C is sqrt(2), about 1.4142. Euclidean ranking is B, A, C.
Normalize A to unit length, giving (1, 0). The other vectors already have unit length. Euclidean ranking now agrees with cosine ranking. For unit vectors, squared distance equals 2 − 2 × cosine similarity.
Suppose an approximate top-two cosine search returns {A, C}. The exact top-two set is {A, B}, so neighbor recall for this query is 1/2, or 50%. This checks agreement with exact vector ranking, not whether either item answers a real user's question. The coordinates are invented to illustrate the calculations.
Illustrative case study
In a fictional case, Iris Support adds vector retrieval to its product documentation. A query about exporting a file returns a highly similar article for an older product version. The retrieval calculation works as configured, but the result is unsuitable for the user's context.
The team adds version filtering and tests queries with explicit and missing version information. It measures both retrieval relevance and whether the final answer uses the retrieved passage correctly. This separates candidate retrieval from answer generation. No claim of real operational improvement is made.
Common misconceptions and mistakes
A similarity score is not automatically a calibrated probability. The highest-ranked result can still be a poor match when the collection contains nothing suitable. Returning k items does not guarantee that k relevant items exist.
Exact search is not always too expensive; feasibility depends on collection size, dimension, hardware, and latency requirements. Approximate search is a tradeoff rather than a universal necessity. Faiss is a library with several index families, including HNSW, rather than one indexing algorithm that excludes graph indexes. Retrieval also does not guarantee that a language model's subsequent answer will be correct.
Business use
For a business search system, evaluate representative queries and difficult near misses. Measure relevance, latency, memory, update cost, and approximate-neighbor recall separately. Include tests where similar wording describes the wrong product, language, date, or eligibility condition.
Keep access restrictions and necessary metadata filters within the retrieval design. When the embedding model changes, assess whether stored vectors must be regenerated and whether thresholds still behave sensibly. A score threshold chosen for one model and collection should not be treated as universally meaningful.
People also ask
Is cosine similarity the same as inner product?
They agree for unit-normalized vectors; without normalization, vector lengths can affect inner-product ranking.
Does approximate mean random or unreliable in every case?
No. It means exact-neighbor recovery is not guaranteed by that search configuration.
Can similarity search replace all keyword search?
Not automatically. Exact terms, identifiers, and metadata may remain important, and combined approaches can be useful.
Check your understanding
Why does raw Euclidean distance prefer B while cosine prefers A?
Answer: A points in exactly the query's direction but has a different magnitude. B is nearer as a point in the unnormalized coordinate space.
Related terms
- Nearest neighbor search
- Cosine similarity
- Euclidean distance
- Retrieval-augmented generation
- vector embedding
- approximate nearest neighbour
- HNSW
- reranking
Learn next
Add D = (10, 0). Its cosine similarity is 1, tying A, but its Euclidean distance from q is 9. Use this to explain why the choice of measure and normalization must be specified together.
More in Graphs and symbolic AI
- Significance test - A significance test evaluates how compatible observed data are with a specified null hypothesis and statistical model.
- Similarity graph - A similarity graph represents data items as nodes and selected similarity relationships as edges.
- Similarity networks - In probabilistic AI, similarity networks are collections of local Bayesian networks constructed to distinguish among overlapping subsets of hypotheses.
- Simple Paths - A simple path is a sequence of connected graph nodes in which no node appears more than once. In a directed graph, each step must follow the edge direction.
- Simulated annealing - Simulated annealing is a randomized optimization method that sometimes accepts a worse candidate solution to escape local optima.
- Singular extensions - Singular extensions are selective game-tree search extensions that allocate extra depth to a move appearing substantially better than its alternatives.
Sources
Sources: Hopsworks, Similarity Search, task context: https://www.hopsworks.ai/dictionary/similarity-search . Faiss official documentation, metric definitions and normalization: https://github.com/facebookresearch/faiss/wiki/MetricType-and-distances . Faiss official index catalogue: https://github.com/facebookresearch/faiss/wiki/Faiss-indexes .
Previous: Similarity networks Next: Simple effect