The World of Vector DBs:

Search, AI, and Weaviate

Victoria Slocum

Machine Learning Engineer

@victorialslocum

/in/victorialslocum

What can you

?

Search systems run the world

multimodal search

Retrieval Augmented Generation (RAG)

Technical Deep Dive

Hybrid Search

Technical Deep Dive

Hybrid Search

Part 1: Keyword search

Data Storage

Part 1: Keyword search

BM25

Part 2: Vector search

ANN Algorithms

• Calculate similarity by calculating distances between vectors

• Manually calculating distance becomes too slow ~40,000 vectors

• ANN algorithms trade exact precision for huge speed improvements

Part 2: Vector search

Hierarchical Navigable Small Worlds (HNSW)

• Custom implementation

• CRUD operations

• Fast, scalable, and accurate

• Tuneable settings

Part 3: Fusion Algorithms

Reciprocal Rank Fusion

Uses the sum of the inverse rank from each algorithm to score and rank documents

Relative Score Fusion

Uses the normalized documents scores added together to rank documents

@victorialslocum

/in/victorialslocum