Vector Databases: How Embedding Search Powers Modern AI Applications
Vector databases are specialized storage systems optimized for similarity search on high-dimensional embedding vectors. They enable semantic search — finding items by meaning rather than keywords — by storing numerical representations (embeddings) of text, images, or other data and quickly finding the nearest neighbors to a query vector. Key implementations include pgvector (PostgreSQL extension), Pinecone, Weaviate, Chroma, Qdrant, and Milvus. They are the retrieval backbone of RAG (Retrieval-Augmented Generation) systems.
A vector database is a storage system optimized for storing, indexing, and searching high-dimensional numerical vectors — the embedding representations produced by machine learning models. Their primary function is similarity search: given a query vector, find the stored vectors most similar to it. ## How They Work **Embedding generation:** Text, images, audio, or other data are converted into fixed-length numerical vectors (typically 256-4096 dimensions) by an embedding model. These vectors encode semantic meaning — similar concepts produce similar vectors regardless of the specific words used. **Indexing:** Vector databases build specialized index structures (HNSW — Hierarchical Navigable Small World graphs, IVF — Inverted File Index, or others) that enable approximate nearest-neighbor search without comparing the query against every stored vector. This makes search sublinear in the number of stored vectors. **Similarity metrics:** Cosine similarity (angle between vectors, most common for text), Euclidean distance (absolute distance), and dot product (magnitude-weighted similarity) are the standard metrics. ## Key Implementations | Database | Type | Notes | |----------|------|-------| | **pgvector** | PostgreSQL extension | Adds vector columns to existing Postgres tables. Good for moderate scale, integrates with existing data. | | **Pinecone** | Managed cloud service | Fully hosted, auto-scaling. Popular for production RAG. | | **Weaviate** | Open source, self-hosted | GraphQL API, supports hybrid search (vector + keyword). | | **Chroma** | Open source, lightweight | Popular for prototyping and local development. Python-native. | | **Qdrant** | Open source, Rust-based | High performance, rich filtering capabilities. | | **Milvus** | Open source, distributed | Designed for billion-scale vector collections. | ## Use Cases The dominant use case is RAG (Retrieval-Augmented Generation): storing document chunk embeddings and retrieving relevant context for LLM prompts. Other applications include image search (find visually similar images), recommendation systems (find similar items/users), anomaly detection (identify outliers in embedding space), and semantic deduplication (find near-duplicate content even when worded differently). RAG (Retrieval-Augmented Generation): How LLMs Access External Knowledge The LLM Wiki Pattern: Personal Knowledge Bases Without RAG