Vectors & Similarity Search: How AI Finds Meaning Instead of Keywords

Search for:

How do I scale PostgreSQL?

Now imagine the best document is titled:

Horizontal Scaling of Relational Databases

A traditional keyword search may completely miss it.

The document is clearly relevant, but it doesn't contain the exact words from the query.

Yet modern AI systems somehow find it.

GitHub Copilot, ChatGPT, Notion AI, enterprise search systems, recommendation engines, and RAG applications all solve this problem every day.

The secret is vector search.

Instead of searching words, modern AI systems search meaning.

From Embeddings to Search

In the previous article, we saw how embedding models convert text into vectors.

For example:

"PostgreSQL Replication" → [0.23, -0.81, 0.44, 0.67, ...]
"Database Sharding"      → [0.19, -0.77, 0.51, 0.61, ...]
"Pizza Recipes"          → [0.71, 0.33, -0.28, -0.52, ...]

The numbers themselves don't matter.

What matters is where those vectors end up in a high-dimensional space.

Concepts with similar meaning tend to be placed close together.

Unrelated concepts end up far apart.

This is what makes semantic search possible.

What Is a Vector?

A vector is simply an ordered list of numbers.

In AI systems, those numbers represent learned patterns extracted from text, images, audio, or other data.

Think of a vector as a location in a very large multidimensional map.

If two pieces of information are similar in meaning, their vectors tend to be near each other.

For example:

"PostgreSQL Replication"
"Database Sharding"

would likely be much closer together than:

"PostgreSQL Replication"
"Pizza Recipes"

The AI model learns these relationships during training.

A Useful Analogy

GPS coordinates locate places on Earth.

Mumbai     → (19.0760, 72.8777)
Bangalore  → (12.9716, 77.5946)

Vectors work similarly.

Except instead of locating cities, they locate ideas.

In this semantic space:

• Nearby vectors represent similar concepts
• Distant vectors represent unrelated concepts

The challenge is finding nearby vectors quickly.

That's where similarity search comes in.

Why Similarity Search Exists

Traditional search systems work primarily through keyword matching.

If a query contains:

scale PostgreSQL

the search engine looks for those exact words.

This approach is fast and reliable, but it has limitations.

It often struggles with:

• Synonyms
• Paraphrased content
• Related concepts
• Natural language questions
• Cross-language retrieval

Humans understand meaning.

Keyword search understands words.

Those are not always the same thing.

Similarity search closes that gap.

How Similarity Search Works

When a user submits a query:

How do I scale PostgreSQL?

the system converts that query into an embedding vector.

Instead of searching text directly, it searches for vectors that are closest to the query vector.

The nearest results might include documents about:

• Read replicas
• Replication strategies
• Database sharding
• Horizontal scaling
• Partitioning

Even if none of those documents contain the exact search phrase.

The retrieval is based on semantic proximity rather than keyword overlap.

Keyword Search vs Semantic Search

This isn't a battle between two approaches.

Modern systems usually combine both.

Hybrid Search Is Becoming the Standard

Most production search systems use a hybrid approach.

They combine:

• Keyword search
• Vector search
• Ranking and result fusion

Keyword search captures exact matches.

Vector search captures meaning.

Together they produce better results than either approach alone.

This is the strategy used by many modern enterprise search systems and AI applications.

Why Vector Databases Exist

Traditional databases are optimized for:

• Transactions
• Joins
• Aggregations
• Range queries

They were never designed for questions like:

Find the 10 most semantically similar documents.

That requires specialized indexing and retrieval techniques.

Vector databases were built specifically for this purpose.

Their responsibilities typically include:

• Storing embeddings
• Building vector indexes
• Running similarity searches
• Filtering using metadata
• Updating and re-indexing vectors

Popular Vector Databases

Some of the most common options today include:

For teams already running PostgreSQL, pgvector is often the easiest place to start.

Why Brute-Force Search Doesn't Scale

Imagine a system containing:

• 100 million vectors
• 1,536 dimensions each

The most straightforward approach would compare the query vector against every vector in the database.

That means:

100 million comparisons
for every search request

Clearly not practical.

Search latency would become unacceptable.

We need a faster approach.

Enter ANN Search

ANN stands for:

Approximate Nearest Neighbor

Instead of finding the exact nearest vectors, ANN algorithms find vectors that are extremely likely to be the nearest.

This small compromise produces enormous performance gains.

In practice:

• Search becomes dramatically faster
• Infrastructure costs decrease
• Retrieval quality remains very high

Most production AI systems use ANN search because exact nearest-neighbor search becomes too expensive at scale.

Common ANN Algorithms

HNSW

Hierarchical Navigable Small World.

Think of it as navigating a map.

Instead of checking every location, the algorithm quickly moves toward the most promising region before refining the search.

HNSW is widely used by:

• Qdrant
• Weaviate
• pgvector

IVF

Inverted File Index.

The idea is simple:

1. Group vectors into clusters
2. Search only the most relevant clusters

This reduces the search space dramatically.

Commonly used in:

• FAISS
• Milvus

LSH

Locality Sensitive Hashing.

Vectors with similar meaning are likely to be placed into the same bucket.

While less common in modern production systems, it remains an important concept in similarity search research.

Where You Already Use Vector Search

Most people interact with vector search daily without realizing it.

RAG Systems

When ChatGPT or another AI assistant retrieves documents before generating a response, vector search is often involved.

Documents are embedded, searched, and retrieved based on meaning.

GitHub Copilot

Copilot embeds your codebase and retrieves relevant files before generating suggestions.

Enterprise Search

Platforms like Notion AI, Slack AI, and Confluence AI use semantic retrieval to find relevant information across large knowledge bases.

Recommendation Engines

Netflix, Amazon, Spotify, and many other platforms use embeddings to recommend content that is similar to what users previously engaged with.

Common Misconceptions

"Vector databases replace SQL databases"

They solve different problems.

SQL databases manage structured data, transactions, joins, and business operations.

Vector databases specialize in semantic retrieval.

Most production systems use both.

"ANN search is inaccurate"

Modern ANN algorithms typically achieve extremely high recall while reducing search latency by orders of magnitude.

The tradeoff is usually well worth it.

"Similarity search guarantees correct answers"

Similarity search retrieves relevant information.

The language model still needs to interpret that information correctly.

Good retrieval improves answers, but it doesn't guarantee them.

"Vectors understand meaning"

Vectors encode statistical patterns learned from training data.

They don't possess human understanding.

They simply provide a mathematical representation that makes semantic retrieval possible.

Why This Matters

Vector search has quietly become one of the foundational building blocks of modern AI.

Whenever an AI system retrieves information based on meaning rather than keywords, embeddings and similarity search are usually involved.

RAG systems, AI assistants, recommendation engines, enterprise search platforms, and coding copilots all rely on the same core idea:

Convert information into vectors and search for nearby meaning.

That's what allows AI systems to move beyond matching words and start retrieving concepts.

Final Thoughts

Embeddings create the map.

Similarity search navigates the map.

Vector databases make searching that map practical at scale.

Together, these technologies power much of the retrieval layer behind modern AI systems.

Once you understand vectors and similarity search, many seemingly magical AI features start looking a lot more like engineering.

If you'd like to actually watch a query move through vector search, ANN indexes, and retrieval pipelines, check out the interactive version on SeeItFlow:

https://seeitflow.com/modern-ai-systems/ai-foundations/vectors-similarity-search