Vector Databases Explained

Why Vector Databases?

Traditional databases search by keyword. Vector databases search by meaning.

SELECT * FROM docs 
WHERE text LIKE '%python%'

vector_db.search(
  embedding("python programming"),
  top_k=5
)

How Vector Databases Work

Step 1: Convert Text to Embeddings

Embeddings are vectors (lists of numbers) that represent meaning:

Text: "The cat sat on the mat"
Embedding: [0.23, -0.45, 0.12, ..., 0.89]  (1536 dimensions)

Semantically similar sentences have similar embeddings:

"The cat sat on the mat" → [0.23, -0.45, 0.12, ...]
"A feline rested on the rug" → [0.24, -0.46, 0.11, ...]
                      ↑ Very close vectors!

Step 2: Store Vectors (Indexed)

Vector DBs use specialized indexes like HNSW (Hierarchical Navigable Small World) for fast nearest-neighbor search:

• Can search 1M vectors in <10ms
• Uses approximate nearest neighbor (ANN), not exact
• Trade-off: Speed vs accuracy (but difference is minimal)

Step 3: Query with New Embedding

Query embedding: "kitten napping on carpet"
→ [0.22, -0.47, 0.11, ...]

Find vectors closest to this embedding
Result: docs about cats, animals, resting, furniture

Vector Databases Comparison

Real Example: Building a RAG Chatbot

from pinecone import Pinecone
from openai import OpenAI

# Initialize
pc = Pinecone(api_key="your-key")
index = pc.Index("documents")
client = OpenAI()

# Step 1: Store documents (one-time)
documents = [
    "Paris is the capital of France",
    "The Eiffel Tower is 330 meters tall",
    "France has a population of 68 million"
]

for doc in documents:
    embedding = client.embeddings.create(
        model="text-embedding-3-small",
        input=doc
    ).data[0].embedding
    
    index.upsert([(doc[:50], embedding, {"text": doc})])

# Step 2: Answer questions by searching
question = "How tall is the Eiffel Tower?"
q_embedding = client.embeddings.create(
    model="text-embedding-3-small",
    input=question
).data[0].embedding

results = index.query(q_embedding, top_k=3)
context = [r['metadata']['text'] for r in results]

# Step 3: Generate answer with LLM
response = client.chat.completions.create(
    model="gpt-4",
    messages=[
        {"role": "user", "content": f"""
Context: {context}
Question: {question}
Answer:"""}
    ]
)

print(response.choices[0].message.content)
# Output: The Eiffel Tower is 330 meters tall.

When to Use a Vector Database

• ✅ Building RAG systems for document Q&A
• ✅ Semantic search across large text corpora
• ✅ Recommendation systems based on similarity
• ✅ Anomaly detection with embeddings
• ❌ Structured queries (use PostgreSQL instead)
• ❌ High-frequency transactions

Limitations and Trade-offs

• Approximate results: ANN doesn't always find the exact nearest neighbor, but close enough for most uses
• Embedding quality: Results only as good as your embedding model
• Curse of dimensionality: Very high-dimensional spaces (1500+ dims) can be slow
• Cost scaling: Pinecone pricing grows with storage/queries

Quick Start: ChromaDB (Local Development)

from chromadb import Client

client = Client()
collection = client.create_collection(name="documents")

# Add documents
collection.add(
    ids=["1", "2", "3"],
    documents=[
        "Apple is a fruit",
        "Apple is a tech company",
        "Banana is a fruit"
    ]
)

# Query
results = collection.query(
    query_texts=["What fruit is yellow?"],
    n_results=1
)
print(results)  # Returns "Banana is a fruit"

Learn RAG at Scale

Master vector databases and RAG architecture with hands-on projects:

• Understanding embeddings: OpenAI vs Hugging Face vs custom
• Choosing and deploying vector databases
• Building production RAG pipelines
• Optimizing retrieval quality
• Cost analysis: cloud vs self-hosted