A Fun & Absurd Introduction to Vector Databases • Alexander Chatzizacharias • GOTO 2024

Vector databases are purpose-built databases optimized for storing and querying high-dimensional vectors, primarily used as backing storage for AI systems

Vectors are mathematical constructs (arrays of numbers) with specific dimensionality that can represent text, images, audio and other data types through embedding models

Key vector database features:

• Semantic search capabilities
• Distance/similarity metrics (cosine, euclidean, etc.)
• Efficient indexing algorithms like hierarchical navigable small worlds
• Cross-modal search (find images using text queries and vice versa)

Embedding models convert data into vectors:

• Text embeddings (BERT, etc.)
• Image embeddings (CLIP, ResNet)
• Audio embeddings (Whisper)
• Results are consistent - same input always produces same vector output

Vector database landscape:

• Open source options: PostgreSQL+pgvector, Milvus, Weaviate
• Closed source/managed services: Pinecone, Azure Vector Search, AWS
• Most provide Python clients, some offer REST APIs
• Built-in embedding options available in some databases

Best practices:

• Consider vector dimensions and distance metrics
• Split large text into smaller segments
• Be aware of context limitations
• Writing to vector DBs is relatively slow
• Results are approximate but semantically relevant

Real-world applications:

• Semantic search
• Image similarity
• Audio matching
• Cross-modal search
• AI/LLM backing storage