Vector Databases (Pinecone, Milvus)

What Are Vector Databases?

Vector DB Architecture

Vector databases store and index high-dimensional vectors (embeddings) for fast similarity search. They are the backbone of every RAG system.

How They Work

1. Embed: Convert your text into vectors using an embedding model (e.g., OpenAI text-embedding-3-small)
2. Index: Store vectors in an optimized index structure (HNSW, IVF)
3. Query: Convert the user's question into a vector, find the nearest neighbors

Popular Options

• Pinecone: Managed, serverless. Great for getting started quickly.
• Milvus / Zilliz: Open-source, highly scalable. Good for large datasets.
• Weaviate: Open-source with built-in vectorization modules.
• ChromaDB: Lightweight, great for prototyping and local development.
• pgvector: PostgreSQL extension. Use your existing Postgres infrastructure.

Similarity Metrics

• Cosine Similarity: Most common. Measures angle between vectors. Best for text.
• Euclidean Distance: Measures straight-line distance. Good for spatial data.
• Dot Product: Fast computation. Works when vectors are normalized.

1from openai import OpenAI
2import chromadb
3
4client = OpenAI()
5chroma = chromadb.Client()
6
7# Create a collection
8collection = chroma.create_collection("docs")
9
10# Add documents
11docs = [
12    "RAG improves LLM accuracy by providing relevant context.",
13    "Vector databases store embeddings for fast similarity search.",
14    "Chunking strategy significantly affects retrieval quality.",
15]
16
17# Embed and store
18for i, doc in enumerate(docs):
19    embedding = client.embeddings.create(
20        input=doc, model="text-embedding-3-small"
21    ).data[0].embedding
22    
23    collection.add(
24        embeddings=[embedding],
25        documents=[doc],
26        ids=[f"doc_{i}"]
27    )
28
29# Query
30query = "How does RAG work?"
31query_embedding = client.embeddings.create(
32    input=query, model="text-embedding-3-small"
33).data[0].embedding
34
35results = collection.query(query_embeddings=[query_embedding], n_results=2)
36print(results["documents"])