pgvector)

#Overview

A Vector Database is a specialized data store designed to index, store, and query high-dimensional vectors (embeddings). It powers similarity search for recommendation engines, semantic search, image retrieval, and RAG (Retrieval-Augmented Generation) for LLMs.

#Key Concepts

Embeddings: Dense numerical vectors (128–2048 dimensions) produced by ML models that capture semantic meaning of items, users, text, or images.
ANN (Approximate Nearest Neighbor): Algorithms that find the K most similar vectors without brute-force comparison. Trades exactness for massive speed gains.
HNSW (Hierarchical Navigable Small World): Graph-based ANN index with best recall/speed trade-off. Builds a multi-layer navigable graph.
IVF (Inverted File Index): Partitions vector space into clusters; searches only relevant clusters. Uses less memory than HNSW.
Product Quantization (PQ): Compresses vectors to reduce memory usage at the cost of some accuracy.
Distance Metrics: Cosine similarity (most common for embeddings), Euclidean (L2), Inner Product (dot product).

#Popular Implementations

Database	Type	Strengths
Milvus	Standalone, open-source	Best for large-scale (billions of vectors), GPU support
Pinecone	Managed SaaS	Easiest to operate, good for startups
pgvector	PostgreSQL extension	Good for < 10M vectors, no extra infra
Weaviate	Standalone, open-source	Built-in ML model integration
Qdrant	Standalone, open-source	Rust-based, excellent filtering support

#When to Use

Recommendation Engines: User/item embeddings → find similar items via ANN search.
Semantic Search: Convert text queries to embeddings → find similar documents.
Image/Video Retrieval: Find visually similar content.
RAG for LLMs: Retrieve relevant context from a knowledge base for grounded generation.

#When NOT to Use

Exact keyword matching (use Elasticsearch).
Structured data queries with joins (use PostgreSQL).
Small datasets < 100K vectors (brute-force cosine sim is fast enough).

#Key Performance Numbers

HNSW on 100M vectors (256-dim):
  Query latency:  < 5ms (top-200 neighbors)
  Memory:         ~40GB (vectors + graph overhead)
  Recall@200:     95%+
  Build time:     ~2 hours

#Trade-offs

Pros: Sub-10ms similarity search over billions of vectors, essential for modern ML-powered applications.
Cons: Memory-intensive (HNSW), requires embedding model management, index rebuild on model version changes.

#Used In This Repo

Recommendation Engine (#23) — User and item embedding storage + ANN candidate generation