Vector Databases

Vector databases are specialized data stores designed to handle high-dimensional vector representations—typically generated from embeddings—and enable efficient similarity search at scale.

Purpose

Traditional databases are inefficient for similarity-based search (e.g., “find documents similar to this one”). Vector DBs solve this by indexing dense vectors and optimizing for Approximate Nearest Neighbor (ANN) retrieval.

Core Features

• Indexing Methods:
  • HNSW (Hierarchical Navigable Small World)
  • IVF (Inverted File Index)
  • PQ (Product Quantization)
  • Flat (brute force, for small datasets)
• Similarity Metrics:
  • Cosine Similarity
  • Euclidean Distance
  • Dot Product
• Operations:
  • Insert / upsert vectors
  • Search (KNN queries)
  • Delete / update vectors
  • Metadata filtering

Use Cases

• RAG (Retrieval-Augmented Generation)
• Semantic Search / Q&A
• Recommendation Systems
• Anomaly Detection
• Clustering and Classification

Popular Tools

• Pinecone — Scalable managed vector DB with filtering and namespaces
• Weaviate — Open source, schema-based with hybrid search
• Qdrant — Rust-based with strong filtering support
• Chroma — Local-first, easy to use for prototyping
• FAISS — Facebook’s ANN library (not a DB but often used as backend)

Considerations

• Embedding model and dimensionality must be consistent across insert/search.
• Choose index type based on dataset size and performance needs.
• Normalize vectors if using cosine similarity.

Example Flow

1. Generate embeddings for a corpus
2. Store vectors in DB alongside metadata
3. At query time, embed the query
4. Search the DB for top K nearest vectors
5. Use results in your application (e.g., LLM context)

Related Notes

• Embeddings
• Semantic Search
• RAG (Retrieval-Augmented Generation)
• Cosine Similarity
• LLM Infrastructure