Vector Databases & Similarity Search (Pinecone, Qdrant, pgvector)

Traditional keyword search engines (like Elasticsearch) rely on exact token matches or stemming. However, they struggle with semantic synonymy—searching for "canine doctor" will fail to match a document containing "veterinarian".

Modern AI models (OpenAI, Cohere, HuggingFace) transform unstructured data (text, images, audio) into dense mathematical vectors of floating-point numbers (Embeddings, typically 768, 1536, or 3072 dimensions).

In this high-dimensional vector space, geometric proximity corresponds directly to semantic similarity:

• Concepts with similar meanings map to nearby coordinates: vec("puppy") ≈ vec("dog") ≈ vec("canine").
• Vector algebra captures relationships: vec("king") - vec("man") + vec("woman") ≈ vec("queen").

Distance Metrics:

1. Cosine Similarity: Measures the angle \theta between two vectors, ignoring magnitude:

\cos(\theta) = \frac{A \cdot B}{\|A\| \|B\|}

Ideal for normalized text embeddings (range: -1 to +1). 2. Euclidean Distance (L2): Measures straight-line geometric distance. 3. Dot Product (Inner Product): Fastest to compute if vectors are pre-normalized to unit length (length = 1.0).