Glossary
What is Vector database?
Also known as: vector store, embedding database
A vector database is a specialized data store optimized for similarity search over high-dimensional vectors — typically the embedding vectors produced by language models. Unlike a traditional database that finds rows matching exact values, a vector database returns the rows whose stored vectors are mathematically closest to a query vector. This is the underlying primitive that makes retrieval-augmented generation (RAG), semantic search, and recommendation systems work at production scale.
Why vector databases are different
Traditional relational databases excel at exact matches and range queries — find the customer with id=123, find orders between two dates. Vector databases solve a different problem: given a query vector, find the N closest stored vectors by some distance metric (cosine similarity, dot product, Euclidean distance). Doing this efficiently over millions of high-dimensional vectors requires specialized indexing structures (HNSW, IVF, ScaNN) that traditional databases don’t implement well.
The "vectors" themselves are usually embeddings — fixed-length numeric arrays produced by a language model from text, images, or audio. Two pieces of text with similar meaning produce embeddings that are close in vector space. The vector database stores those embeddings and finds similar ones at query time.
Common vector databases
The vector database landscape has several major players. Pinecone is the most widely used managed offering and most common for first deployments. Weaviate is open source with strong filtering and hybrid search. Qdrant is open source with a developer-friendly API. pgvector is a PostgreSQL extension that lets you do vector search in your existing Postgres database — useful when you don’t want a separate piece of infrastructure. Chroma is a lighter option for development and small deployments. Choice depends on scale, infrastructure preferences, and whether you’re doing pure vector search or hybrid (vector + keyword + metadata filtering).
Vector databases in AI legal-intake context
For law firms running AI voice agents or chatbots, vector databases typically hold firm-specific reference content: practice-area FAQs, intake decision trees, prior matter summaries, and ethics-rule documentation. When the agent needs to ground a response in firm-specific facts ("what languages does the firm speak?"), it embeds the question, queries the vector database, retrieves the most relevant chunks, and includes them in the LLM prompt. Without a vector database, the agent has to either hard-code all firm facts in the system prompt (doesn’t scale) or hallucinate.