Topic 3: Beyond Single-Vector
7 min read·21 Sept 2026
Topic 3: Beyond Single-Vector
Why go beyond one vector per chunk? A single vector squeezes a whole chunk into one point, so details get blurred. The techniques here keep more detail (multi-vector), exact terms (learned sparse), or other media (images).
| Representation | Vectors per chunk | Strength | Cost |
|---|---|---|---|
| Single dense vector | 1 | Simple, fast, cheap | Loses fine detail |
| Multi-vector (ColBERT-style) | One per token | Fine-grained matching | Much more storage |
| Learned sparse (SPLADE-style) | 1 sparse vector | Exact terms plus related terms; explainable | Needs sparse index support |
| Multimodal (CLIP-style) | 1 per image or text | Text-to-image and image-to-text search | Separate model; weaker on long text |