What I Learned Today About “Vector Embeddings”

Introduction

What I learned today about “Vector Embeddings” from the Python + AI: Level Up series of Microsoft Reactor.

I’m thinking about semantic search these days, and this session was so timely.

Here are my main takeaways:

What Are Vectors?

• Vectors are words or data converted into numbers using embedding models.
• Each embedding model understands only its own vectors.
• Computers understand numbers, not words — that’s why we convert text, images, and video into numeric form.

Where to Use in Business

• Add semantic, multilingual, and multimodal search to websites using vector similarity search.
• Use vector embeddings for recommendation systems and fraud detection.

Making Vector Searches Faster

• Use Approximate Nearest Neighbor (ANN) algorithms such as HNSW, IVFFlat, Faiss, or DiskANN instead of exhaustive search.
• HNSW (Hierarchical Navigable Small Worlds) works well for frequently updated data and scales logarithmically with large indexes.

How to Reduce Storage Needs and Make Them Faster

• Vector quantization reduces the size of vectors by lowering numeric precision.
• Scalar quantization converts 64-bit floating-point numbers into smaller integers (16-bit, 8-bit, or 4-bit).
• Binary quantization (1-bit) gives extreme compression while still retaining semantic information.
• In Azure AI Search, quantization can reduce storage by ~74% (8-bit) and ~96% (1-bit).
• Dimensionality reduction reduces the number of vector dimensions.
• Matryoshka Representation Learning (MRL) can reduce dimensions while keeping semantic meaning (for supported models).

How to Mitigate Information Loss

• Combine quantization and dimensionality reduction carefully.
• Use a two-stage retrieval process:
  1. Retrieve top N results from the compressed index (fast).
  2. Re-score those results using uncompressed vectors (accurate).
• This approach ensures both speed and quality in vector search.

It was a great session with a lot of insightful information.

The session link is here: https://developer.microsoft.com/en-us/reactor/events/25084/