Vector Database

What is a Vector Database?

A vector database converts text and images into “sequences of numbers” for storage and enables fast retrieval of semantically similar information. Where traditional databases seek “exact matches,” vector databases find “similar items.” It’s the technological foundation of RAG and AI search.

In a nutshell: Like a librarian who, when asked “Give me books on this theme,” recommends “How about this book?” Finding related books—searching by meaning similarity rather than keyword exact match.

Key points:

• What it does: Converts text and images into “vectors” (numerical lists) and searches by semantic similarity
• Why it’s needed: Enables instant search for “semantically similar” information that keyword searches cannot find
• Who uses it: RAG systems, AI chatbots, recommendation systems, and general AI applications

Why it matters

Traditional databases (like SQL) excel at “exact matches” but struggle with “semantic similarity.” Searching for “customer dissatisfaction” won’t retrieve the similar information “customer anger.” However, with the advent of LLMs that process text as meaning, databases optimized for that processing became necessary.

RAG (the technique of retrieving relevant information from external databases to give to AI) cannot exist without vector databases. To make AI “understand” company internal documents, you need vector databases to quickly extract semantically related information and provide it to the AI.

How it works

Vector databases are most easily understood through “coordinate axes.” Imagine a plane with two axes: “taste” and “price.” Sweet items at the top, spicy at the bottom, expensive on the right, inexpensive on the left. Each data point is expressed as “coordinates,” and nearby coordinates are found. This is the essence of vector databases.

In practice, this occurs in high-dimensional spaces like 256 or 512 dimensions. Relationships like “apple sweetness and pear sweetness are semantically similar” are embedded in numerical space (a process called “embedding”). This is automatically generated by specialized models (called “embedding models”) and stored in vector databases. At search time, queries are also embedded by the same model, and the “closest” data in space is retrieved.

A representative algorithm is HNSW (Hierarchical Navigable Small World), which uses graph structures to find needed information from billions of vectors in just milliseconds.

Real-world use cases

Customer support AI

When a user asks “Wi-Fi won’t connect,” the vector database instantly retrieves related articles like “wireless connection issues” and “network configuration” from the internal knowledge base. Providing those articles to the LLM generates user responses. Articles missed by keyword search are found through semantic similarity.

E-commerce product recommendations

When recommending “running shoes” to a user who previously bought “sneakers,” the vector database embeds product descriptions and searches for products similar to the “sneaker” vector. Recommendations consider design, functionality, and use case similarity beyond just category tags.

Medical literature search

When medical students search for “diabetes and heart disease relationships,” the vector database retrieves semantically related papers like “metabolic abnormalities” and “vascular disease,” supporting physician decision-making.

Benefits and considerations

Vector databases’ merit is enabling “meaning-aware search.” Combined with RAG systems, they can significantly reduce LLM hallucinations (fabrication). Providing fact-based information to AI improves reliability.

However, there are also considerations. Everything depends on embedding model quality. Data embedded with low-quality models has poor search accuracy. When re-embedding data with new models, all data must be reprocessed, incurring computational costs. Additionally, the “curse of dimensionality” (increasing complexity with higher dimensions) is a challenge.

Related terms

• RAG — Retrieval Augmented Generation. A technique that retrieves related information from vector databases and provides it to LLMs for more accurate answer generation.
• LLM — Large Language Model. AI that understands text retrieved from vector databases and generates responses.
• Semantic Search — Search based on semantic relevance rather than keyword exact match. The primary feature of vector databases.
• Machine Learning — Foundation technology for training models (embedding models) that generate vector embeddings.
• Natural Language Processing — Technology that processes text as meaning. Vector embeddings are based on this.

Frequently asked questions

Q: What’s the difference between vector databases and regular databases (like SQL)?

A: Regular databases search with “exact matches” or ranges like “name is ‘Tanaka’” or “age is 30.” Vector databases specialize in ambiguous searches like “meaning is similar.” Though “customer dissatisfaction,” “complaints,” and “anger” are different words, they’re semantically related—vector databases can retrieve them together.

Q: Can you build an AI chatbot with just a vector database?

A: No. Vector databases only handle “information retrieval.” To actually generate answers, an LLM is needed. A typical approach combines “retrieving information with vector DB → giving to LLM to generate answers” (RAG).

Q: How are embeddings created before storing data in a vector database?

A: Passing text to an embedding model (like OpenAI’s embedding model or Sentence Transformers) returns a numerical list (vector). That’s then stored in the vector database.