When someone searches your enterprise knowledge base for "staff holiday entitlements" and finds a document titled "Annual Leave Policy," that's vector search at work. It didn't match keywords. It matched meaning. This is the technology that makes enterprise AI actually useful.
The Definition
Vector search (also called semantic search or similarity search) is a method for finding information based on meaning rather than exact keyword matches. It works by converting text into numerical representations called embeddings (lists of numbers that capture the semantic meaning of the content). When you query the system, it converts your question into the same format and finds the closest matches.
Think of it this way: keyword search asks "which documents contain these words?" Vector search asks "which documents mean the same thing?"
Why Enterprise Leaders Should Care
Every enterprise AI capability that touches your internal knowledge depends on vector search. RAG systems, intelligent document retrieval, claims processing, compliance checking - they all use vector search to find the right information before the language model generates a response.
70%
of enterprise knowledge queries fail with keyword search alone
Source: Elastic, Enterprise Search Report, 2023
The gap is obvious when you consider how people actually ask questions. An employee asking "how many days off do I get when my partner has a baby?" won't find a document titled "Parental Leave - Secondary Carer Entitlements" with keyword search. Vector search bridges that gap because it understands that both are about the same thing.
How It Works
1. Embedding
An embedding model converts text into a vector, a list of hundreds or thousands of numbers that represent the text's meaning in mathematical space. Texts with similar meanings produce vectors that are close together. "Company leave policy" and "employee holiday entitlements" produce nearby vectors. "Quarterly revenue forecast" does not.
2. Indexing
These vectors are stored in a vector database, a specialised database optimised for similarity search. Options include pgvector (a PostgreSQL extension), Pinecone, Weaviate, and Qdrant. The database builds index structures that make nearest-neighbour lookup fast, even across millions of documents.
3. Querying
When a user asks a question, the question is embedded using the same model. The vector database returns the stored vectors closest to the query vector, the most semantically relevant documents. This is typically measured using cosine similarity.
Vector Search vs Keyword Search
| Aspect | Keyword Search | Vector Search |
|---|---|---|
| Matches on | Exact words and stems | Meaning and concepts |
| Synonyms | Only with manual configuration | Handled automatically |
| Natural language queries | Poor | Excellent |
| Setup effort | Low | Moderate (needs embedding model) |
| Best for | Precise, known-term queries | Conversational questions |
In practice, the best enterprise systems use hybrid search, combining both approaches. Keyword search handles precise lookups (policy numbers, product codes). Vector search handles natural language questions. Together, they cover the full range of how people actually seek information.
Where It Powers Enterprise AI
Knowledge retrieval. The foundation of RAG. Employees ask questions and get answers grounded in company documents.
Document classification. Compare a new document's embedding against reference categories. The closest match determines the classification, no training data required.
Duplicate detection. Find semantically similar records even when the wording differs. Essential for compliance and data quality.
Recommendations. "Find claims similar to this one" or "show me contracts with comparable terms" - similarity search applied to enterprise content.
Getting Started
For enterprises building their AI stack, vector search is foundational infrastructure. The decisions that matter most:
- Embedding model - this determines retrieval quality more than any other choice
- Chunking strategy - how you split documents affects what gets retrieved
- Hybrid approach - combine vector and keyword search for coverage across query types
- Vector database - matters less than the above; choose based on existing infrastructure
The embedding model and chunking strategy are where you should invest your tuning effort. The database is a commodity.