Vector Databases

Vector databases store and search high-dimensional vectors — mathematical representations of data like text, images, and audio. Instead of keyword matching, vector search finds semantically similar items: “running shoes” matches “athletic sneakers” even though they share no keywords. This is the foundation of semantic search, RAG systems, and recommendation engines.

---

How Vector Search Works

Traditional search:
  Query: "running shoes"
  Match: Documents containing the exact words "running" AND "shoes"
  Miss: "athletic sneakers", "jogging footwear", "nike air max"

Vector search:
  Query: "running shoes" → embed → [0.23, -0.45, 0.67, ...]
  Search: Find nearest vectors in database
  Match: "athletic sneakers" (similarity: 0.95)
         "jogging footwear" (similarity: 0.91)
         "nike air max running" (similarity: 0.89)
         "hiking boots" (similarity: 0.42) ← correctly low

---

Embedding Generation

# Text embeddings with OpenAI
from openai import OpenAI
client = OpenAI()

response = client.embeddings.create(
    model="text-embedding-3-small",
    input="Running shoes for marathon training",
)
embedding = response.data[0].embedding  # 1536-dimensional vector

# Sentence Transformers (open source)
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("all-MiniLM-L6-v2")
embeddings = model.encode([
    "Running shoes for marathon training",
    "Athletic sneakers for jogging",
    "Formal leather dress shoes",
])
# Shape: (3, 384)

# Similarity
from sklearn.metrics.pairwise import cosine_similarity
sims = cosine_similarity(embeddings)
# [running, sneakers] = 0.87 (similar)
# [running, formal]   = 0.23 (dissimilar)

---

Vector Database Comparison

Database	Type	Best For	Hosting
Pinecone	Purpose-built	Production, managed, fast	Fully managed
pgvector	PostgreSQL extension	Postgres shops, small scale	Self-hosted or cloud Postgres
Weaviate	Purpose-built	Multi-modal, built-in ML	Managed or self-hosted
Qdrant	Purpose-built	Performance, Rust-based	Managed or self-hosted
ChromaDB	Embedded	Prototyping, small datasets	In-process
Milvus	Purpose-built	Large scale, enterprise	Self-hosted

---

Pinecone Usage

from pinecone import Pinecone

pc = Pinecone(api_key="your-api-key")
index = pc.Index("products")

# Upsert vectors
index.upsert(vectors=[
    {
        "id": "shoe-001",
        "values": embed("Running shoes for marathon"),
        "metadata": {
            "category": "footwear",
            "price": 129.99,
            "brand": "Nike",
        }
    },
    # ... more vectors
])

# Query with metadata filtering
results = index.query(
    vector=embed("shoes for running"),
    top_k=10,
    filter={
        "category": {"$eq": "footwear"},
        "price": {"$lte": 150.00},
    },
    include_metadata=True,
)

for match in results.matches:
    print(f"{match.id}: {match.score:.3f} - {match.metadata}")

---

Hybrid Search

Vector-only search:
  Pro: Semantic understanding
  Con: Misses exact keyword matches ("SKU-12345")

Keyword-only search:
  Pro: Exact matching, well-understood
  Con: Misses semantic similarity

Hybrid search:
  Combine both: vector similarity + keyword matching
  Score = α * vector_score + (1-α) * keyword_score

  Implementation:
  1. Sparse embedding (BM25/TF-IDF) for keywords
  2. Dense embedding for semantics
  3. Combine scores with weighted fusion
  4. Or: separate queries, reciprocal rank fusion

---

Anti-Patterns

Anti-Pattern	Consequence	Fix
Wrong embedding model	Poor similarity results	Match model to your domain and data type
No metadata filtering	Slow search over entire index	Pre-filter with metadata before vector search
Embedding everything unstructured	Noise in results	Chunk text properly, clean data
No evaluation metrics	Cannot measure search quality	Precision@k, recall@k, NDCG
Single vector per document	Miss multi-faceted content	Multiple embeddings or chunk-level vectors

Vector databases are the infrastructure layer for AI-native applications. They enable the shift from “search by keywords” to “search by meaning.”