Choosing Dimensions
The number of dimensions in your vectors is determined entirely by your embedding model. You cannot choose it independently - you must match the collection dimensions to your model’s output size exactly.
Match your embedding model
Every embedding model outputs a fixed-size vector. You must set your collection’s dimensions to exactly this number.
| Model | Provider | Dimensions |
|---|
| text-embedding-ada-002 | OpenAI | 1536 |
| text-embedding-3-small | OpenAI | 1536 (default) |
| text-embedding-3-large | OpenAI | 3072 (default) |
| embed-english-v3.0 | Cohere | 1024 |
| embed-multilingual-v3.0 | Cohere | 1024 |
| all-MiniLM-L6-v2 | Hugging Face | 384 |
| all-mpnet-base-v2 | Hugging Face | 768 |
| e5-large-v2 | Hugging Face | 1024 |
| bge-large-en-v1.5 | Hugging Face | 1024 |
If you create a collection with 768 dimensions and try to upsert vectors with
1536 dimensions, VecLabs will return a DimensionMismatchError. There is no
automatic conversion.
Higher dimensions = higher quality (up to a point)
More dimensions allow the model to encode more nuance. Generally:
- 384 dims - good for lightweight, fast applications. Open-source models. Good enough for most use cases.
- 768-1024 dims - solid quality. Good balance of speed and accuracy for most production workloads.
- 1536 dims - best-in-class quality for most text tasks. OpenAI’s most widely used size. VecLabs benchmarks at this size.
- 3072 dims - marginal improvement over 1536 for most tasks. 2x memory and compute cost.
The quality gap between 1536 and 3072 dimensions is small for most applications. Unless you’re working on a task that specifically requires maximum precision, 1536 is the practical ceiling.
Matryoshka models: flexible dimensions
Some newer models (like OpenAI’s text-embedding-3 series) support Matryoshka Representation Learning (MRL) - they can output vectors at multiple dimension sizes with graceful quality degradation.
For example, text-embedding-3-small can output:
- 1536 dims (full quality)
- 512 dims (~90% of full quality, 3x smaller)
- 256 dims (~85% of full quality, 6x smaller)
If storage or compute is constrained, MRL models let you trade a small amount of quality for significant efficiency gains. Specify the dimensions when generating embeddings:
// OpenAI text-embedding-3-small with reduced dimensions
const response = await openai.embeddings.create({
model: "text-embedding-3-small",
input: "Your text here",
dimensions: 512, // reduce from default 1536
});
// Collection must match
const collection = sv.collection("my-collection", { dimensions: 512 });
Memory implications
Each vector dimension is stored as a 32-bit float (4 bytes). At 100K vectors:
| Dimensions | Memory for vectors |
|---|
| 384 | ~147 MB |
| 768 | ~294 MB |
| 1536 | ~587 MB |
| 3072 | ~1.17 GB |
You cannot change dimensions after collection creation
Like the distance metric, dimensions are fixed at collection creation time. To change them, create a new collection and re-embed all your documents.
This is why it’s important to commit to an embedding model before building your production index. Switching from ada-002 to a 3072-dim model later requires a full re-index.