Vector Database

Vector Database

Source: What is a Vector Database? Powering Semantic Search & AI Applications - IBM Technology on YouTube, What is a Vector database

Use case: We want to store content like image into a database. Example: store a sunset in hills.

Options are below

Use a relational database

• Store text, binary data into a database
• Store metadata of content into database
  • Tags: sunset, hills

There is a challenges with relating unstructured data with structured data

Use a vector database

Vector embeddings are an array of numbers. In the database, similar items are stored close together, semantically similar content is stored close together.

Can store unstructured data (examples: image, test, audio). Data is transformed into vector embeddings which are stored in database.

• The sunset and hills images is stored as [0.91, 0.15, 0.83] where dimensions could be:
  • [Elevation changes, buildings, colour temperature]
  • Another example of beach: [0.17, 0.08, 0.95] which is similar to the sunset image in colour temperature and lack of buildings but does not have elevation changes.

Actual vectors could have 100s to 1000s dimensions

How are vector embeddings created?

Using models that were trained on datasets. There are different embedding models trained for certain content, for example models for images, for text, and for audio.

Data is passed through multiple layers and each layer extracts features (example for images, a layer extracts shapes, another extracts words, another extracts meaning). Actual models could have 1000s of layers.

Efficient queries

The vector database makes it easier to find vectors with similar dimensions, but comparing all vectors requires a better method as it will be slow.

Vector indexing helps with queries to find approxiate close matches, example algorithms:

• Hierarchical navigable small world (HNSW) - multi-layered graphes
• Inverted file index (IVF) - divides vectors into clusters and compares with clusters

Indexes may be less accurate, but are more efficient for searches.

Answering a user’s questions with RAG and Vector Database

Vector databases are applied with Retrieval Augmented Generation (RAG) - Retrieval Augmented Generation (RAG) to store and get data to answer user questions using existing data that has been processed into a vector database.

When a user asks a question, vectors can be compared for similarity and given to a LLM for processing to give answers

See Also

• Retrieval Augmented Generation (RAG) - Retrieval Augmented Generation (RAG)