Vector database là gì và khi nào bạn thật sự cần nó?

Vector DB giải quyết vấn đề gì?

Traditional search: keyword matching (LIKE '%query%')

Vector search: semantic matching (ý nghĩa tương tự)

Ví dụ: search "cách nấu phở" → vector search cũng trả về "recipe for Vietnamese beef noodle soup"

Cách hoạt động

1. Text → Embedding model → Vector (array of 768-3072 numbers)
2. Vectors lưu vào specialized database
3. Query → embed → find nearest vectors (cosine similarity)

Khi nào CẦN vector DB?

• RAG system (search internal docs)
• Semantic search (product search, FAQ matching)
• Recommendation (similar items/content)
• Anomaly detection (find outliers in data)

Khi nào KHÔNG cần?

• Data < 10K documents → pgvector đủ
• Only keyword search needed → PostgreSQL full-text
• Real-time updates critical → vector DBs have indexing lag
• Budget tight → managed vector DB = expensive

Options

| DB | Type | Best for |

|----|------|----------|

| pgvector | Extension | Already use PostgreSQL, small-medium data |

| Pinecone | Managed | Production, don't want to manage infra |

| Qdrant | Self-hosted | Full control, good performance |

| Weaviate | Self-hosted | GraphQL API, hybrid search |

| ChromaDB | Embedded | Prototyping, small projects |