Semantic search focuses on what a query means rather than requiring the same words to appear in a document. For example, an employee searching for “Can I expense a train ticket?” might retrieve a policy titled “Reimbursement for business travel.” Traditional keyword search can also handle synonyms and linguistic variations, but semantic search uses learned representations or relevance models to recognize broader relationships between queries and content.
A common implementation converts documents and queries into embeddings: numerical vectors that represent characteristics of their meaning. Documents are often split into smaller passages, called chunks, before embedding. A vector index retrieves passages whose embeddings are close to the query embedding according to a similarity measure. Query and document embeddings must come from compatible models designed for retrieval; equal vector dimensions alone do not make two models compatible.
In the cloud, an ingestion pipeline reads documents from repositories or object storage, extracts text, creates chunks, generates embeddings, and stores vectors alongside source references and metadata. At query time, the application embeds the request, applies authorization and metadata constraints, retrieves candidates, and optionally reranks them with another relevance model. Managed model APIs, search services, and databases with vector capabilities can implement these stages without requiring engineers to manage every underlying component.
Semantic search is an important retrieval technique for retrieval augmented generation, or RAG. RAG retrieves relevant information from an external knowledge source and supplies it as context to a generative model, which uses that context to produce an answer. A support assistant might retrieve troubleshooting instructions and then explain the appropriate steps, with links to the source passages. Amazon Bedrock Knowledge Bases is one managed implementation of this pattern. RAG can also use keyword or hybrid retrieval; it does not require semantic search exclusively.
Business applications include searching internal procedures, finding relevant product documentation, and helping customer service agents locate answers in approved material. Similar semantic representations can support content recommendations, such as finding articles related to one a user is reading, although a recommendation system may also need behavioral and personalization signals. With suitable multimodal models, retrieval can extend beyond text to images or other media. Semantic search itself does not generate text, images, video, or audio; generation is a separate capability that an application may combine with retrieval.
Hybrid search combines lexical and semantic retrieval because neither approach is best for every query. Semantic retrieval can handle paraphrases well but miss exact distinctions involving product identifiers, error codes, dates, or negation. Lexical search can preserve those exact matches. Reranking may improve relevance but adds latency and inference cost. Approximate nearest-neighbor indexes improve search efficiency at scale but can miss relevant candidates, so engineers must balance recall, response time, memory, and cost.
Retrieval quality depends on document quality, chunk boundaries, embedding-model suitability, and index freshness. Small chunks may lose context, while large chunks may mix unrelated subjects. Changing the embedding model generally requires re-embedding the indexed content. Evaluate retrieval using representative queries and relevance judgments, including recall at a chosen result count, and evaluate generated answers separately for correctness and support from retrieved evidence.
Semantic similarity is not proof of factual accuracy, and RAG does not eliminate hallucinations. Retrieved documents can be stale, incorrect, or contain malicious instructions. Enforce document-level access controls during retrieval, preserve source references, treat retrieved text as untrusted data, and define behavior for insufficient evidence. Retrieving private documents into a prompt does not itself retrain the foundation model, but the application's data handling and model-provider policies still require review.
Amazon OpenSearch Service