# Supabase — Matryoshka embeddings: faster OpenAI vector search using Adaptive Retrieval

- Company: Supabase (supabase.com)
- Announced: 2024-02-13
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://supabase.com/blog/matryoshka-embeddings
- Record: https://forck.live/items/21447-matryoshka-embeddings-faster-openai-vector-search-using-adaptive-retrieval
- Subject: Supabase
- Models affected: text-embedding-3-small, text-embedding-3-large

OpenAI released text-embedding-3-small and text-embedding-3-large, which outperform the previous text-embedding-ada-002 model on MTEB and MIRACL benchmarks. The models include a dimensions API parameter that allows shortening the embedding vector to any size, enabled by Matryoshka Representation Learning (MRL), a training technique that preserves semantic meaning in truncated sub-vectors.

## Evidence

Verbatim from https://supabase.com/blog/matryoshka-embeddings:

> OpenAI's text-embedding-3-large produces 3072 dimensions by default. But with their new dimensions API parameter, you can shorten the number of dimensions to any size

---

Record: https://forck.live/items/21447-matryoshka-embeddings-faster-openai-vector-search-using-adaptive-retrieval
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md