To make RAG work, you need two key technologies: embedding models that convert text to numbers, and vector databases that store and search those numbers efficiently.

What Are Embeddings?

An embedding is a numerical representation of text — a list of numbers (a "vector") that captures the meaning of words, sentences, or documents.

• "The cat sat on the mat" → [0.12, -0.45, 0.78, ...] (1536 numbers)
• "A feline rested on the rug" → [0.11, -0.44, 0.79, ...] (very similar numbers!)
• "Stock prices rose sharply" → [-0.89, 0.23, -0.15, ...] (very different numbers)

Similar meanings produce similar vectors. This is how semantic search works — you find documents by meaning, not just keyword matching.

Popular Embedding Models

What Are Vector Databases?

A vector database is optimized for storing and searching vectors. Unlike traditional databases that match exact values (SQL WHERE clauses), vector databases find the most similar vectors to a query — even if there's no exact match.

How Vector Search Works

1. Store document embeddings with metadata
2. Convert user query to an embedding
3. Find the K nearest vectors (most similar documents)
4. Return those documents ranked by similarity

Popular Vector Databases

Cloud-hosted:

• Pinecone — Fully managed, easy to start, scales well. Best for teams that want zero infrastructure.
• Weaviate Cloud — Supports hybrid search (vector + keyword). Good for complex queries.
• Qdrant Cloud — High performance, rich filtering. Good for production workloads.

Self-hosted / Open source:

• Chroma — Lightweight, Python-native. Perfect for prototyping and small projects.
• pgvector (PostgreSQL) — Add vector search to your existing Postgres database. Great if you're already using PostgreSQL.
• Milvus — Highly scalable, designed for billions of vectors. Enterprise-grade.
• FAISS (Facebook) — In-memory vector search library. Fast but no persistence built-in.

Choosing an Embedding Model

Key factors:

• Dimensions — Higher dimensions = more nuance but more storage/compute
• Speed — How fast are embedding API calls?
• Cost — Per-token pricing adds up at scale
• Language support — Some models handle multilingual content better
• Domain — Some models are fine-tuned for specific domains (code, medical, legal)

Best Practice: Test Before Committing

Embed a sample of your documents with 2-3 different models. Run test queries and compare retrieval quality. The best model for your use case depends on your specific documents and query patterns.