<< back to Guides

AI Core Concepts (Part 18): Vector Search

Vector search allows you to find items that are semantically similar using vector embeddings, rather than exact keyword matches. It powers modern search engines, recommendation systems, and Retrieval-Augmented Generation (RAG) pipelines.

1. What Is Vector Search?

It’s a technique to search by meaning rather than literal text. It uses cosine similarity, dot product, or L2 distance to find vectors closest to a query vector.

"AI revolution" → retrieves articles like “machine learning progress”
"Heart attack symptoms" → finds related medical advice, even if phrased differently

2. Process Overview

Embed your content (e.g., documents) into dense vectors
Store them in a vector database
Embed the user query
Search using similarity (e.g., top-k nearest vectors)
Use results in your app (search, answer, recommend, etc.)

3. Use Cases

Task	Description
Semantic Search	"Explain gravity" ≠ exact match, still relevant
Recommendations	Find similar users/items by behavior
RAG	LLM retrieves relevant context chunks
Similarity Detection	Near-duplicate detection
Multimodal Search	Search images by text description

4. Example: Using FAISS for Local Vector Search

from sentence_transformers import SentenceTransformer
import faiss
import numpy as np

# Step 1: Create embeddings
model = SentenceTransformer('all-MiniLM-L6-v2')
docs = ["What is AI?", "Benefits of machine learning", "Pizza recipe"]
doc_embeddings = model.encode(docs)

# Step 2: Index with FAISS
index = faiss.IndexFlatL2(len(doc_embeddings[0]))
index.add(np.array(doc_embeddings))

# Step 3: Query
query = "Explain artificial intelligence"
query_embedding = model.encode([query])

# Step 4: Search
D, I = index.search(np.array(query_embedding), k=2)
print("Top results:", [docs[i] for i in I[0]])

5. Vector Similarity Metrics

Metric	Use Case
Cosine	Measures angle, not magnitude (text)
L2 Distance	Euclidean distance (images, audio)
Dot Product	Often used in neural retrieval

from sklearn.metrics.pairwise import cosine_similarity

cosine_similarity([vec1], [vec2])  # value between -1 and 1

6. Popular Vector Databases

Tool	Features
FAISS	Local search, fast, Facebook
Pinecone	Cloud-native, scalable, metadata filtering
Weaviate	Semantic + symbolic search, GraphQL API
Qdrant	Open source, hybrid filters
Chroma	Lightweight, built for LLM apps
Milvus	High-volume, production-grade

7. Example: Pinecone + OpenAI

import openai
import pinecone

# Init Pinecone
pinecone.init(api_key="your-key", environment="us-west1-gcp")
index = pinecone.Index("my-index")

# Create embedding
query = "How does a transformer work?"
query_vec = openai.Embedding.create(
  input=[query], model="text-embedding-3-small"
)["data"][0]["embedding"]

# Query vector DB
results = index.query(vector=query_vec, top_k=3, include_metadata=True)
for match in results['matches']:
    print(match['metadata']['text'])

8. Integration with LLMs (RAG)

# Retrieve relevant documents using vector search
query_vector = embed("What is supervised learning?")
docs = vector_db.similarity_search(query_vector, k=3)

# Inject into LLM prompt
context = "\n\n".join([doc.text for doc in docs])
prompt = f"Answer based on the context below:\n\n{context}\n\nQuestion: What is supervised learning?"
response = llm.generate(prompt)

9. Tips for Better Vector Search

✅ Use domain-specific embedding models
✅ Normalize vectors (unit length)
✅ Chunk text intelligently (e.g., 200–300 words)
✅ Store metadata (title, tags, source)
✅ Filter results with hybrid search (text + metadata)

📚 Further Resources

<< back to Guides