Elasticsearch supports both vector and full-text search through distinct but complementary mechanisms. For full-text search, it relies on inverted indexes and scoring algorithms, while vector search uses dense vector representations and similarity metrics. These approaches address different use cases, such as keyword-based queries and semantic similarity matching, and can be combined for hybrid search scenarios.
For full-text search, Elasticsearch builds inverted indexes to enable fast text-based queries. When a document is indexed, its text fields are analyzed—split into tokens, normalized, and stored in a structure that maps terms to the documents containing them. For example, a field containing “quick brown fox” might be tokenized into ["quick", "brown", “fox”] with positional data. Queries like match or term leverage these indexes to find documents containing specific words. Elasticsearch uses the BM25 algorithm to rank results based on term frequency and document length, prioritizing documents where search terms appear prominently. Filters and aggregations can further refine results. This approach works well for exact matches, phrase searches, or fuzzy queries with typo tolerance.
Vector search is supported through the dense_vector field type, which stores arrays of floating-point numbers representing embeddings (e.g., from machine learning models). To perform similarity searches, Elasticsearch offers the knn (k-nearest neighbors) query, which compares a query vector against stored vectors using metrics like cosine similarity. For example, a vector representing “canine companions” could be compared to product descriptions encoded as vectors to find semantically related items. Under the hood, Elasticsearch uses the HNSW algorithm for approximate nearest neighbor search, balancing speed and accuracy. Vectors can be indexed with optional parameters like dimensions (e.g., 768 for BERT embeddings) and similarity (e.g., cosine). This enables use cases like image similarity search or natural language queries where keyword matching falls short.
Elasticsearch allows combining vector and full-text search in a single query. For instance, you might filter documents using a match query for specific keywords, then apply a knn query to re-rank results based on semantic similarity. The hybrid scoring method or custom script_score functions can blend relevance scores from both approaches. This is useful for applications like e-commerce, where a search for “wireless headphones” could prioritize products with exact keyword matches while also surfacing items semantically related to “Bluetooth earbuds.” Developers can tune weights between vector and text-based scores to optimize results. Both search types benefit from Elasticsearch’s distributed architecture, enabling scalability across large datasets.
