1. What is vector database?

A vector database is a type of database optimized to store, index, and search high-dimensional vectors, which are typically used to represent unstructured data like text, images, audio, and video.

A vector database stores data entries as records with the following main components:

  • ID: primary key. E.g. “doc_1234”
  • Vector: the main feature, a high-dimensional float array representing the data. E.g. [0.12, -0.45, 0.33, …, 0.98]
  • Metadata: optional, it is used for tags or attributes. E.g.

    {
  "type": "support_doc",
  "language": "en",
  "created_at": "2025-01-01"
}

Here is the sample record in a vector database:

    {
  "id": "doc_1234",
  "vector": [0.12, -0.45, 0.33, ..., 0.98],
  "metadata": {
    "category": "account",
    "language": "en",
    "priority": "high"
  }
}
photo

2. Vector database applications

Here are some popular applications of Vector DB:

2.1. Semantic search

Keyword search is too literal. Vector DB supports to find the results that are semantically similar.

E.g. A user goes to a company’s help center and types: “I can’t get into my account”. Traditional keyword search might fail because it just looks for the documents with matching keywords (“can’t”, “account”). It will miss the helpful articles like:

– “How to reset your password”

– “Fix login issues”

…

 

2.2. Embeddings from ML models

Models like OpenAI, Cohere, and CLIP convert text/images to vectors. We need a place to store and search those vectors.

 

2.3. Fast similarity search

Comparing 1 vector against millions is very slow with brute force. Vector DB use optimized indexes (HNSW, IVF) to make it fast.

E.g. Recommend similar songs from a database of 1M audio clips.

 

2.4. Building LLM based apps

Large Language Models (like GPT-4) don’t know your data. Vector DB is used to enable semantic retrieval of relevant context that the language model (LLM) can use to generate accurate, grounded, and useful response.

E.g. A legal chatbot retrieves case law snippets to help answer questions.

 

2.5. Personalization or recommendations

Match users with similar behaviors, preferences, or content.

 

3. Retrieval Augmented Generation

A typical traditional system may include:

  • A relational DB (e.g., MySQL, PostgreSQL)
  • Admin dashboard or website
  • REST APIs
  • Keyword search, form input,…

 

3.1. What is the enhancement of the Vector Database?

Traditional Component Vector DB Enhancement Example
🔍 Keyword Search Add semantic search for queries
📦 Product Catalog Recommend similar products using product descriptions
📚 FAQ or Document Center Retrieve relevant documents by meaning, not keywords
👤 CRM or Helpdesk Auto-suggest replies based on semantic case history
🛒 E-commerce Suggestions based on vector similarity
🧾 SQL Database Store metadata or primary data in Postgres + use vector DB for similarity search

 

3.2. How to integrate Vector Database into a traditional system?

photo

  • Embedding model:
    • Hosted Embedding Services (API-based): OpenAI, Cohere Embed API, Google Vertex AI Embedding, AWS Bedrock / SageMaker, Azure OpenAI,… You send input text or code to their API and get back a vector embedding.
    • Custom (Self-Hosted) Embedding Models: SentenceTransformers, BGE, CodeBERT, CodeT5,…In this approach, you run an embedding model on your own infrastructure.
  • Generate vector embeddings: Use a model like OpenAI’s text-embedding-ada-002 to convert text into vectors. Some common embedding model input types: text (text-embedding-ada-002, sentence-transformers), images (CLIP, OpenCLIP, ResNet, DINO, BLIP), audio (Whisper, Wav2Vec2, AudioCLIP), video (VideoCLIP, XCLIP, TimeSformer), code (CodeBERT, GraphCodeBERT).
  • Vector Database: The traditional database holds your structured business data (rows, columns), while the vector database holds the semantic representation (vectors) of selected fields (e.g., descriptions, documents, images).

photo

  • User enters a natural language query, this is just a text input and we want to search by meaning, not exact words.
  • Use an embedding model to convert the query into a vector.
  • Query the vector DB to find the top K nearest vectors (ANN search).
  • Use those IDs to get the actual data from traditional DB.
  • Return the most relevant items to the user.

 

3.3. Some main features

 

# Feature Description Example / Use Case
1 Semantic Similarity Search Retrieve results based on meaning, not exact keywords. Input: “healthy fruit” => Output: “red apple”, “cherry juice”
2 K-Nearest Neighbors (k-NN) / ANN Search Find the top K most similar vectors using cosine similarity or dot product. Top product matches, similar documents, related user profiles
3 Personalized recommendations Recommend items close in vector space to user preferences. Based on: user behavior, product likes, past interactions.
4 Contextual Retrieval (RAG for LLMs) Retrieve relevant vectors to provide context for a language model. Used in: AI chatbots, document Q&A, internal knowledge assistants