Vector Search

Vector search finds items whose numerical representations are close to a query vector under a chosen similarity measure. An embedding is a vector produced by a model to represent characteristics of an input, such as the meaning of a paragraph or the visual features of an image. For example, a search for “recover account access” might retrieve a document titled “Reset your password” even without matching the query's exact wording. Retrieval quality depends on what the embedding model captures; numerical proximity does not prove relevance or factual correctness.

A typical cloud implementation starts with an ingestion pipeline. The pipeline extracts content, divides long documents into useful chunks, generates an embedding for each chunk, and stores each vector with an identifier and metadata. Source text can reside alongside the vector or in another system accessible through the identifier. At query time, the application embeds the query into a compatible vector space and retrieves the closest matching records. Query and document embeddings must use compatible models and configurations; matching vector dimensions alone does not make unrelated embedding spaces comparable.

In retrieval-augmented generation, or RAG, the application supplies retrieved passages to a foundation model as context. For an internal support assistant, vector search can retrieve troubleshooting instructions that the model uses to draft an answer with source references. The vector database stores and searches embeddings; it does not inherently generate embeddings or write answers. Retrieval does not retrain the foundation model, and providing relevant context does not guarantee that the generated answer will be correct.

Several cloud storage patterns support this design. Amazon OpenSearch Service provides vector indexing and search alongside lexical search and metadata filtering. Amazon Aurora PostgreSQL-Compatible Edition and Amazon RDS for PostgreSQL support the pgvector extension, allowing embeddings to reside near relational records and participate in SQL queries. Within the Amazon Neptune family, Neptune Analytics supports vector similarity search over embeddings associated with graph vertices. This enables similarity retrieval to be combined with relationship analysis. Capabilities differ by engine, deployment option, and extension version, so a service-family name alone is not sufficient to establish support.

Search designs balance accuracy, latency, memory, storage, and update cost. Exact nearest-neighbor search returns the true nearest neighbors under the selected metric. Brute-force exact search compares the query against every eligible vector, while some exact indexing methods can prune candidates without sacrificing exactness. Approximate nearest-neighbor indexes reduce search work at scale but may miss some true nearest neighbors. Their usefulness depends on measured recall and latency for the workload. Cosine similarity, dot product, and Euclidean distance are common comparison methods; select the method and normalization that match the embedding model's guidance.

Vector search does not automatically replace keyword search. Product codes, error identifiers, and precise names often benefit from lexical matching. Hybrid retrieval combines lexical and vector results, and a reranking model can reorder retrieved candidates. Evaluate retrieval with representative questions and relevance judgments rather than interpreting similarity scores as probabilities of correctness. Chunk size, metadata filters, candidate counts, and embedding quality all affect results.

Operational controls are essential. Enforce tenant and document permissions before retrieved content reaches the model; semantic similarity is not authorization. Propagate source updates and deletions to the vector store, and track embedding model versions. Changing models generally requires re-embedding the corpus and rebuilding or migrating indexes. Budget for embedding generation, ingestion, index capacity, replication, and queries. Protect embeddings as potentially sensitive derived data rather than assuming they anonymize their source content.