Lesson 16 / 28
Multi-Vector, Multilingual and Multimodal Embeddings
Know the extensions: many vectors per item, many languages, many media.
By Rahul Agarwal · 15 years in web development
One space for more than English text
Multilingual embedding models place sentences from many languages in one shared space, so an English query can retrieve a Hindi document and the reverse; test Hindi, English and mixed Hinglish, as quality varies by language and script. Multimodal models (such as CLIP-style models) embed images and text together, enabling "search images by text" and similar-image search. Multi-vector (late-interaction) methods such as ColBERT keep one vector per token and compare sets of vectors, often more accurate but needing more storage and specialised indexes. A parent-child design embeds small chunks for precise matching but returns the larger parent section. Pick the extension that matches your problem; do not add complexity before the single-vector baseline is measured.
Test Hinglish queries
Users type Hindi in Latin letters and mix languages. Include such queries in your golden set.
Quick check: What does a multilingual embedding model enable?
- Removing the need for chunking
- Translating text word for word
- Matching a query in one language to documents in another
- Encrypting text
Answer
Matching a query in one language to documents in another — A shared space lets meaning match across languages.