What is Semantic Search?

Keyword search breaks down in predictable ways. People don’t know the exact term a document uses, one word means different things in different departments, and some answers are spread across several sources. In a large organisation with decades of systems and documents behind it, that is the normal case.

Semantic search maps both the query and the content onto a shared model of the domain. A search for “heart attack treatment” finds documents about “myocardial infarction therapy”, because the model records that the two mean the same thing. A search for “Java” from an engineering team returns programming results rather than coffee or Indonesia.

The difference shows up in everyday work. Analysts spend less time hunting for the right source, support teams answer more questions from the knowledge base without escalating, and compliance teams can show that a search for a regulated term also found the documents that use another name for it.

Most semantic search systems combine a few techniques:

Entity recognition. The query is parsed to pick out the things it mentions, such as people, products or concepts, and what type each one is.

Ontology-based expansion. An ontology adds synonyms, broader terms and related concepts to the query. A search for “SUV” also matches “sport utility vehicle”, and can be set to include “crossover” as well.

Vector embeddings. A language model turns queries and documents into dense vectors, and results are ranked by how close they sit to the query. This catches paraphrases that no synonym list would contain.

Knowledge graph lookups. Some questions need facts from several connected records, for example which suppliers of a part also supply a competitor. A knowledge graph can answer that by following the relationships, where keyword or vector search would only find documents that happen to mention both.

Legal research. A law firm can index its precedents and opinions against a legal ontology. A search then finds relevant cases from other jurisdictions, even when they use different terms for the same doctrine.

Retail. On a shop’s website, “lightweight laptop for travel” can be read as a request about weight and battery life, and ranked on those attributes instead of on which product pages repeat the words.

Healthcare. Clinical search can map a free-text query to SNOMED CT concepts and ICD-10 codes, so a clinician finds the relevant records and papers whichever terminology they were written in.