similarity-vector

Documentation / similarity/similarity-vector

Other

AutoTokenizer

type AutoTokenizer<> = AutoTokenizer;

Defined in: similarity/similarity-vector.js:2

Type Parameters

Type Parameter

---

calculateCosineSimilarity()

function calculateCosineSimilarity(vectorA: number[], vectorB: number[]): number;

Defined in: similarity/similarity-vector.js:168

Cosine similarity gets similarity of two vectors by whether they have the same direction (similar) or are poles apart. Cosine similarity is often used with text representations to compare how similar two documents or sentences are to each other. The output of cosine similarity ranges from -1 to 1, where -1 means the two vectors are completely dissimilar, and 1 indicates maximum similarity.

Parameters

Parameter	Type	Description
vectorA	number[]
vectorB	number[]

Returns

number

-1 to 1 similarity score

Similarity

addEmbeddingVectorsToIndex()

function addEmbeddingVectorsToIndex(documentVectors: string[], options?: object): Promise<HierarchicalNSW>;

Defined in: similarity/similarity-vector.js:97

VSEARCH: Vector Similarity Embedding Approximation in RAM-Limited Cluster Hierarchy

1. Compile hnswlib-node or NGT algorithm C++ to WASM JS for efficient similarity search.
2. Vector index is split by K-means into regional clusters, each being a specific size to fit in RAM. This is better than popular vector engines that require costly 100gb-RAM servers because they load all the vectors at once.
3. Vectors for centroids of each cluster are stored in a list in SQL, each cluster's binary quantized data is exported as base64 string to SQL, S3, etc.
4. Search: Embed Query, Compare to each cluster centroid to pick top clusters, download base64 strings for those clusters, load each into WASM, find top neighbors per cluster, merge results sorted by distance.

NGT Algorithm NGT Cluster https://qdrant.tech/articles/memory-consumption/ Lancedb Usearch

• ANN Benchmarks

Parameters

Parameter	Type	Description
documentVectors	string[]	An array of document texts to be vectorized.
options?	{ maxElements: number; numDimensions: number; }	Optional parameters for vector generation and indexing.
options.maxElements?	number	The maximum number of data points.
options.numDimensions?	number	The length of data point vector that will be indexed.

Returns

Promise<HierarchicalNSW>

The created HNSW index.

Author

Malkov et al. (2016),

---

convertTextToEmbedding()

function convertTextToEmbedding(text: string, options?: object): Promise<{
  embedding: number[];
  embeddingsDict: {
  };
}>;

Defined in: similarity/similarity-vector.js:26

Text embeddings convert words or phrases into numerical vectors in a high-dimensional space, where each dimension represents a semantic feature extracted by a model like MiniLM-L6-v2. In this concept space, words with similar meanings have vectors that are close together, allowing for quantitative comparisons of semantic similarity. These vector representations enable powerful applications in natural language processing, including semantic search, text classification, and clustering, by leveraging the geometric properties of the embedding space to capture and analyze the relationships between words and concepts. Text Embeddings, Classification, and Semantic Search (Youtube)

Parameters

Parameter	Type	Description
text	string	The text to embed.
options?	{ pipeline: AutoTokenizer; precision: number; }
options.pipeline?	AutoTokenizer	The pipeline to use for embedding.
options.precision?	number	default=4 - The number of decimal places to round to.

Returns

Promise<{ embedding: number[]; embeddingsDict: { }; }>

---

getAllEmbeddings()

function getAllEmbeddings(index: HierarchicalNSW, precision: number): number[][];

Defined in: similarity/similarity-vector.js:148

Retrieves all embeddings from the HNSW index.

Parameters

Parameter	Type	Default value	Description
index	HierarchicalNSW	undefined	The HNSW index containing the embeddings.
precision	number	8	default=8 - The number of decimal places to round to.

Returns

number[][]

An array of embedding vectors. *

---

getEmbeddingModel()

function getEmbeddingModel(options?: object): Promise<AutoTokenizer>;

Defined in: similarity/similarity-vector.js:50

Initialize HuggingFace Transformers pipeline for embedding text.

Parameters

Parameter	Type	Description
options?	{ modelName: string; pipelineName: string; }
options.modelName?	string	default="Xenova/all-MiniLM-L6-v2" - The name of the model to use
options.pipelineName?	string	default "feature-extraction",

Returns

Promise<AutoTokenizer>

The pipeline. *

---

searchVectorIndex()

function searchVectorIndex(
   index: HierarchicalNSW, 
   query: string, 
   options?: object): Promise<object[]>;

Defined in: similarity/similarity-vector.js:132

Searches the vector index for the nearest neighbors of a given query.

Parameters

Parameter	Type	Description
index	HierarchicalNSW	The HNSW index to search.
query	string	The query string to search for.
options?	{ numNeighbors: number; }	Optional parameters for the search.
options.numNeighbors?	number	The number of nearest neighbors to return.

Returns

Promise<object[]>

A promise that resolves to an array of nearest neighbors, each with an id and distance.

Throws

If there's an error during the search process.

Example

const index = await addEmbeddingVectorsToIndex(documentVectors);
const results = await searchVectorIndex(index, 'example query');
console.log(results); // [{id: 3, distance: 0.1}, {id: 7, distance: 0.2}, ...]

---

weighRelevanceConceptVector()

function weighRelevanceConceptVector(
   documents: string[], 
   query: string, 
   options?: any): Promise<object[]>;

Defined in: similarity/similarity-vector.js:189

Rerank documents's chunks based on relevance to query, based on cosine similarity of their concept vectors generated by a 20MB MiniLM transformer model downloaded locally.

A Complete Overview of Word Embeddings

Parameters

Parameter	Type	Description
documents	string[]
query	string
options?	any

Returns

Promise<object[]>