Vector Search Explained - The Technology Behind SearchAF

Vector search is revolutionizing how we build answer engines and intelligent search systems. In this post, we'll explore the technology that powers SearchAF's answer generation capabilities.

What Are Vector Embeddings?#

Vector embeddings are numerical representations of data that capture semantic meaning. Instead of storing text as words, we convert it into high-dimensional vectors (arrays of numbers).

"running shoes" → [0.24, -0.18, 0.91, ... 0.33]  // 768 dimensions
"athletic footwear" → [0.26, -0.17, 0.89, ... 0.31]  // similar vector!

Notice how similar concepts have similar vectors, even though the words are different.

How Vector Search Works#

Step 1: Generate Embeddings#

When you add a product to SearchAF, we:

Extract text from title, description, and attributes
Pass it through an embedding model (like OpenAI's text-embedding-3)
Store the resulting vector alongside your product data

Step 2: Query Embedding#

When a user searches:

Convert the query to a vector using the same embedding model
This query vector captures the semantic intent

Step 3: Similarity Search#

Find products with vectors closest to the query vector using:

Cosine Similarity

similarity = (A · B) / (||A|| × ||B||)

This measures the angle between vectors - closer angles mean more similar meaning.

Why Vector Databases?#

Traditional databases struggle with vector operations:

High Dimensionality: Vectors can have 768+ dimensions
Similarity Computation: Requires specialized indexes
Scale: Millions of products need efficient search

Antfly's Advantage#

SearchAF is built on Antfly, which provides:

HNSW Indexes: Hierarchical navigable small world graphs for fast similarity search
Distributed Architecture: Horizontal scaling for massive catalogs
Hybrid Queries: Combine vector and traditional filters

SELECT * FROM products
WHERE vector_similarity(embedding, query_vector) > 0.7
  AND price < 100
  AND in_stock = true
ORDER BY vector_similarity(embedding, query_vector) DESC
LIMIT 10

Embedding Models#

Different models excel at different tasks:

OpenAI text-embedding-3-small#

Dimensions: 1536
Best for: General-purpose search
Pros: High quality, widely supported
Cons: API-dependent, cost per query

Sentence Transformers#

Dimensions: 384-768
Best for: Self-hosted solutions
Pros: Free, customizable
Cons: Need GPU for inference

Custom Fine-tuned Models#

Dimensions: Variable
Best for: Domain-specific search
Pros: Optimized for your data
Cons: Requires ML expertise

Hybrid Search: Best of Both Worlds#

Pure vector search isn't always optimal. SearchAF uses hybrid search:

final_score = α × vector_score + (1-α) × keyword_score

Where α controls the balance:

α = 0: Pure keyword (traditional BM25)
α = 0.5: Equal weight
α = 1: Pure semantic

This catches:

Exact brand names and SKUs (keyword)
Conceptual matches (vector)
Typos and variations (both)

Real-World Performance#

Our benchmarks show:

Metric	Traditional Search	Vector Search	Hybrid
Relevance (NDCG)	0.72	0.84	0.91
Latency (p95)	12ms	45ms	28ms
Zero Results	18%	4%	2%

Challenges and Solutions#

Challenge 1: Cold Start#

Problem: New products lack interaction data Solution: Vector embeddings work immediately from product descriptions

Challenge 2: Query Latency#

Problem: Vector operations are compute-intensive Solution: Antfly's distributed architecture and HNSW indexes

Challenge 3: Embedding Drift#

Problem: Models can become outdated Solution: Automated reindexing and A/B testing of new models

Getting Started#

Want to leverage vector search in your application?

Use SearchAF: Get started in minutes with our managed platform
Explore Antfly: Self-host for complete control
Read the Docs: API Reference

The Future#

We're excited about upcoming developments:

Multimodal Search: Combine text, images, and metadata
Retrieval Augmented Generation: Use vectors for AI chatbots
Real-time Personalization: User-specific vector adjustments

Want to learn more? Check out our Semantic Search Guide