Vector Search (AI Search)

Concept

If a user searches for "cozy sweater", traditional keyword engines (like Elasticsearch’s Inverted Index) look for documents containing the exact words “cozy” and “sweater”. If your database contains a product named "warm knitted pullover", standard search will return 0 results, because none of the letters match.
Vector Search (Semantic Search) uses Artificial Intelligence to understand the meaning and context of words. It realizes that a “cozy sweater” and a “warm pullover” are conceptually the exact same thing, and returns the result perfectly.

Mental Model

Imagine a 3-Dimensional graph.

• The X-axis is “Warmth”.
• The Y-axis is “Clothing Type”.
• The Z-axis is “Comfort”.

When you index the product "warm knitted pullover", an AI model calculates its position on this graph. It gets coordinates [0.9, 0.8, 0.9]. This array of numbers is called an Embedding Vector.
When the user searches for "cozy sweater", the AI model calculates its coordinates: [0.85, 0.8, 0.88].
The Database simply measures the physical distance between the two points on the graph. Because they are grouped closely together in “concept space”, the database returns it as a match.

How It Works

1. The Embedding Model: You take all the text/images in your database and pass them through a Machine Learning model (like OpenAI’s text-embedding-ada-002). The model returns a massive array of floats (e.g., a 1536-dimensional vector) representing the “meaning” of the text.
2. The Vector Database: You save these massive arrays into a specialized database (like Pinecone, Milvus, or pgvector). Standard B-Trees cannot search 1536-dimensional space. These databases use advanced algorithms (like HNSW - Hierarchical Navigable Small World graphs) to cluster similar vectors together.
3. The Query: The user types a query. Your backend passes the query to the Embedding Model to turn it into a 1536-dimensional vector.
4. KNN / ANN Search: Your backend asks the Vector Database to perform a K-Nearest Neighbors (KNN) or Approximate Nearest Neighbors (ANN) search. The database uses math (Cosine Similarity or Euclidean Distance) to find the vectors stored in the database that are physically closest to the user’s query vector, returning the matches.

Trade-Offs

• Pros: Unbelievably powerful. It powers Spotify’s song recommendations, Netflix’s movie matching, and ChatGPT’s “Retrieval-Augmented Generation (RAG)” memory. It can search images, audio, and text interchangeably (e.g., uploading an image of a sweater to find text descriptions of similar sweaters).
• Cons:
  • Latency & Cost: Generating embeddings via AI APIs takes hundreds of milliseconds and costs money.
  • The Exact Match Problem: Vector search is terrible at finding exact serial numbers or precise IDs. Searching for iPhone 14 might return iPhone 13 as the top result because conceptually they are 99% identical.

Real-World Usage

Because Vector Search struggles with exact keyword matching, the modern industry standard is Hybrid Search.
You run a traditional Elasticsearch keyword query AND a Vector Search query simultaneously. You take both sets of results, feed them into a secondary “Re-ranking” algorithm (like Reciprocal Rank Fusion - RRF), and combine them to get the absolute best of both worlds.

Interview Questions

Q: In a Vector Database, why do we use ANN (Approximate Nearest Neighbors) instead of exact KNN (K-Nearest Neighbors)?
A: Exact KNN mathematically calculates the distance between the user’s query vector and every single vector stored in the database. If you have 1 billion documents, calculating 1 billion 1536-dimensional mathematical distances will take hours and consume massive CPU.
ANN (Approximate Nearest Neighbors) sacrifices perfect accuracy for speed. It uses graphing algorithms (like HNSW) to intelligently skip 99% of the database and only measure distances in the general “neighborhood” of the query. It returns results in 10 milliseconds, and while it might accidentally miss the absolute #1 best match, finding the #2 or #3 best match is perfectly acceptable for semantic search.