Vector Databases (LLM Search & Retrieval)

Vector databases store, index, and query high-dimensional vector embeddings generated by machine learning models.

1. Key Concepts

Traditional databases index text fields literally. Vector databases represent text semantically as arrays of floats (embeddings) and query using vector distance metrics:

Distance calculations: Cosine Similarity, Euclidean Distance (L2), Dot Product

Search Algorithms (ANN)

To find matches within milliseconds across billions of high-dimensional vectors, databases use Approximate Nearest Neighbor (ANN) indices:

• HNSW (Hierarchical Navigable Small World): A multi-layered graph representation offering fast logarithmic searches.
• IVF (Inverted File): Clusters vectors into subsets using k-means, querying only the closest centroids.
• PQ (Product Quantization): Compresses vector footprints to fit inside RAM.

2. Dynamic Retrieval-Augmented Generation (RAG)

Clients generate vector query embeddings using OpenAI/Cohere APIs, query the vector database for matching context indices, and append the results to LLM system prompts.

3. References & Tech Blogs

• Pinecone: What is a Vector Database and How Does it Work?
• Meta Research: Billion-Scale Similarity Search with Faiss