A traditional database finds records that match exactly: the invoice with this number, customers in this postcode. A vector database finds records that are similar in meaning: documents about late delivery penalties, even if none of them uses those words. It does this by storing embeddings, lists of numbers produced by an AI model that capture what a piece of text, an image or an audio clip means, and by searching for the nearest ones. That is the retrieval engine behind most company AI assistants. The two types are complementary, and many businesses can add vector search to the database they already run.
What is an embedding?
An embedding model reads a chunk of content and outputs a vector, typically a few hundred to a few thousand numbers. Content with similar meaning produces vectors that sit close together. 'Refund policy for damaged goods' and 'how do I return a broken item' land near each other even though they share almost no words. Searching by meaning then becomes a geometry problem: find the stored vectors closest to the vector of the question.
Vector database vs traditional database: side by side
| Traditional (relational) database | Vector database | |
|---|---|---|
| Stores | Rows and columns of structured values | Embeddings plus the original text and metadata |
| Query | Exact match, ranges, joins (SQL) | Nearest neighbours by similarity |
| Answers | Precise and complete | Approximate and ranked by relevance |
| Index type | B-tree, hash | Approximate nearest neighbour, such as HNSW or IVF |
| Best for | Transactions, finance, inventory | Semantic search, RAG, recommendations, deduplication |
| Hardware pressure | CPU, storage IOPS | RAM for the index; GPU optional |
Vector search is approximate by design. It trades a small chance of missing the very best match for a large gain in speed, and that trade-off is tunable.
How does a vector database fit into an AI assistant?
- Documents are split into chunks and each chunk is embedded.
- The vectors are stored with the text, the source and access permissions.
- A user's question is embedded with the same model.
- The database returns the most similar chunks the user is allowed to see.
- The language model writes an answer from those chunks and cites them.
This is retrieval-augmented generation (RAG). It is the indexing stage of the AI data pipeline, and we walk through a full build in our private knowledge assistant guide and analysing company documents locally.
Do you need a dedicated vector database?
| Option | Examples | Suits |
|---|---|---|
| Extension to your existing database | pgvector for PostgreSQL; vector features in other mainstream databases | Up to a few million vectors; teams that want one system to run and back up |
| Embedded or lightweight store | Chroma, SQLite-based extensions, FAISS | Prototypes and single-application tools |
| Dedicated vector database | Milvus, Qdrant, Weaviate | Tens of millions of vectors and up, heavy filtering, high query rates |
| Search engine with vectors | Elasticsearch, OpenSearch | Teams that already run search and want hybrid results |
For most SMEs, the honest answer is that an extension such as pgvector is enough to begin with. A few thousand documents produce perhaps tens or hundreds of thousands of chunks, which any of these handle comfortably.
A vector database doesn't replace your SQL database. It answers a different question. SQL tells you exactly which records match; vectors tell you which ones are about the same thing.
Why is hybrid search usually better?
Pure semantic search is weak on exact identifiers. Part numbers, model codes, invoice numbers and names look like noise to an embedding model. Hybrid search combines keyword matching with vector similarity and then re-ranks the results. For a business that sells products with SKUs, hybrid search is close to essential. Metadata filters, such as department, date and permission, matter just as much as the vectors.
How much memory does a vector index need?
Raw vectors take number of vectors, times dimensions, times 4 bytes at standard 32-bit precision.
| Vectors | 768 dimensions | 1,024 dimensions | 1,536 dimensions |
|---|---|---|---|
| 100,000 | 0.3 GB | 0.4 GB | 0.6 GB |
| 1 million | 3.1 GB | 4.1 GB | 6.1 GB |
| 10 million | 31 GB | 41 GB | 61 GB |
| 100 million | 307 GB | 410 GB | 614 GB |
CALC: vectors x dimensions x 4 bytes, decimal GB, raw vectors only. In-memory indexes such as HNSW add overhead on top, and stored text and metadata are extra. Vector quantization can cut memory several times at some cost in accuracy.
So the index for a typical company knowledge base fits in the RAM of an ordinary server or NAS. Very large indexes are a server-memory question, where technologies such as CXL memory may help in future.
What hardware does vector search need?
- Embedding the documents. A one-off batch job, plus updates. A modest GPU does it quickly; a CPU can do small collections.
- Serving queries. Mostly RAM and CPU. GPU-accelerated indexing and search exist for very large collections.
- Storage. SSD for the database, plus the original documents on a NAS with backup.
- The language model. Usually the largest hardware cost, and separate from the database.
If you change the embedding model later, every document has to be re-embedded, so keep the originals and the pipeline that produced the index.
Buying in Singapore
For an all-in-one private assistant, the NVIDIA DGX Spark 4TB (S$9,450) or ASUS Ascent GX10 128GB/1TB (S$10,350) can run the embedding model, the vector store and the language model together for a small team. For embedding at volume, the NVIDIA RTX PRO 4000 Blackwell (S$5,640). For the document repository, the Synology DS925+ (S$1,260) or DS1825+ (S$2,148), and for fast storage with room for virtual machines, the QNAP TVS-h674T (S$3,715) (listed 11 Nov 2026; prices change; ask for a current quotation). Server memory such as the Micron 64GB DDR5 RDIMM is price on request. To start, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
What is a vector database? A vector database stores embeddings, numeric representations of the meaning of text, images or audio, and retrieves the items most similar to a query. It powers semantic search and retrieval-augmented generation.
How is a vector database different from a SQL database? A SQL database returns exact matches on structured values. A vector database returns approximate, ranked matches by similarity of meaning. Most AI applications use both.
Do I need a dedicated vector database for RAG? Not always. Extensions such as pgvector handle up to a few million vectors well. Dedicated vector databases suit very large collections, heavy filtering and high query rates.
How much RAM do one million vectors need? About 4.1 GB for raw 1,024-dimension vectors at 32-bit precision, plus index overhead and the stored text and metadata.
