Dimension

Dimension (sometimes called embedding dimension) is the length of the vector produced for each chunk.

If your embedding has dimension 512, every chunk is represented as a vector of 512 numbers.

Dimension matters because it affects:

• Quality: higher dimensions can capture more nuance (often better retrieval), up to a point.
• Cost & speed: higher dimensions increase storage size, network payloads, and compute.
• Compatibility: your vector database collection/index must be created with the same dimension.

Where Dimension is set

You can set dimension on the job via:

• embed_spec.dimension

If omitted, VNG uses the default dimension of the selected model.

Examples

These examples assume you already connected to VNG (see Connect to VNG) and you have a vng client instance.

1) Use the model default dimension

If you don’t set a dimension, VNG uses the model’s default 2048.

vecs = vng.embed(
    inputs=["Quickstart guide"],
)
 
print(len(vecs[0]))  # e.g., 2048

2) Set a common dimension for a knowledge base (e.g., 768)

vecs = vng.embed(
    inputs=["Company onboarding policy"],
    dim=768,
)
 
print(len(vecs[0]))  # 768

3) Smaller vectors to reduce storage (e.g., 256)

vecs = vng.embed(
    inputs=["FAQ: billing and invoices"],
    dim=256,
)
 
print(len(vecs[0]))  # 256

4) Larger vectors for maximum quality (e.g., 1536)

vecs = vng.embed(
    inputs=["Technical architecture overview"],
    dim=1536,
)
 
print(len(vecs[0]))  # 1536

5) Important: your VBase collection dimension must match

When you store embeddings in VBase, the collection/index dimension must match the vectors you generate.

A simple rule:

• If you embed with dim=768, create the VBase collection with dimension 768.
• If you change dimension later, create a new collection and re-embed.

Supported dimensions

Our multimodal embedding model supports dimensions in this range:

• 128 → 2048 (inclusive)

If you request a value outside this range, the job may fail validation.

Choosing the right dimension

A practical starting point:

• 512 or 768 for a general-purpose knowledge base
• 128–256 if cost/storage is the top priority (smaller vectors)
• 1024–2048 if you need maximum quality and can afford larger vectors

Important note for your vector database

Your vector index must be created with a fixed dimension.

If you change the dimension later, you’ll need to:

• create a new collection/index with the new dimension, and
• re-embed (re-ingest) your content.