What is vector search, and does your AI application need it?

Vector search retrieves items by comparing numerical representations called embeddings. It can help when a query and a relevant document use different words. Whether it helps your application is an evaluation question: some workflows need exact record lookups or keyword search instead.

A separate database is one option

The pgvector project documents exact and approximate nearest-neighbor search within PostgreSQL, alongside relational data. Its documentation explains that approximate indexes trade some recall for speed. Vector retrieval does not inherently require a separate database service. Source: https://github.com/pgvector/pgvector

Plan a retrieval experiment

For a hypothetical internal knowledge assistant, start with representative questions and a reviewed set of relevant documents. Compare keyword retrieval, vector retrieval, and a combination using the same corpus and access rules. Record missed documents as well as irrelevant matches.

• Check retrieval quality on the language and terminology your users actually use.
• Measure response latency and operating cost under representative load.
• Apply document permissions before returning context to a user or model.
• Plan updates and deletion so stale records do not remain searchable.

Retrieval is not a correctness guarantee

For a retrieval-augmented generation workflow, evaluate both the retrieved context and the generated answer. A relevant passage does not establish that every sentence of an answer follows from it. Include questions with no supported answer and define when the application should ask for help or decline to answer.

Decide from evidence

Document the baseline, failure cases, maintenance effort, and total operating cost before selecting infrastructure. Estimate business value separately from technical quality; neither a vector index nor an AI feature establishes a financial return.

Editorial revision: September 28, 2026. This is architecture planning guidance, not a benchmark or a client result.