What replaced text-embedding-ada-002 in OpenAI's lineup?
OpenAI replaced text-embedding-ada-002 with text-embedding-3-small (1536d, cheap) and text-embedding-3-large (3072d, top quality), both Matryoshka-truncatable via the dimensions API parameter.
Think of OpenAI's lineup of text to fingerprint services like a coffee shop menu. The old standard was one fixed-size cup called ada-002. The new menu offers two cups: a smaller one called 3-small that costs less and tastes nearly as good for most customers, and a larger one called 3-large for those who want the premium version. Both new cups also offer something the old cup never did: you can ask the barista for a half-pour at the counter, paying only for what fits in your cup, instead of always getting the full pour. The old cup is still on the menu for loyal customers, but the staff quietly steer newcomers toward the new pair.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
text-embedding-ada-002 was the embedding default for an entire generation of RAG systems. From its release in December 2022 through early 2024, it was the model that 'just worked' for a large fraction of production retrieval pipelines. In January 2024 OpenAI released the text-embedding-3 family, and ada-002 quietly moved from 'recommended' to 'still supported.'
This deep dive covers what changed, why the new lineup is cleanly better on cost and quality, the one new capability (Matryoshka truncation) that did not exist in ada-002, and how to migrate a production index without serving mixed-vintage vectors.
The shape of the new lineup
OpenAI replaced one model with two. text-embedding-3-small is the new cost-optimized default at 1536 dimensions and $0.02 per million tokens. text-embedding-3-large is the quality-optimized option at 3072 dimensions and $0.13 per million tokens.
Both are trained with a Matryoshka representation learning objective, which means any prefix of the output vector is itself a valid lower-dimensional embedding. The API exposes this through a 'dimensions' parameter. Call the model with dimensions=512 and you get a 512-d vector that is meaningful (though weaker than the full vector). ada-002 had no such control; you took 1536 dimensions or you did not call the API.
What the two SKUs are for
text-embedding-3-small is the new starting point for any general-purpose English RAG system. It is cheaper than ada-002 and scores higher on MTEB. text-embedding-3-large is for systems where retrieval quality is the load-bearing constraint and the extra 4x storage cost (3072 vs 768 if you truncate aggressively) is acceptable. In practice many teams pick 3-large and truncate to 1024 or 1536 dimensions for a quality-cost compromise that beats both ada-002 and 3-small at default sizes.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Model | Dimensions | Price per 1M tokens | MTEB lift over ada-002 | Matryoshka |
|---|---|---|---|---|
| text-embedding-ada-002 | 1536 (fixed) | $0.10 (deprecated default) | baseline | no |
| text-embedding-3-small | 1536 (truncatable) | $0.02 | +5 points typical | yes |
| text-embedding-3-large | 3072 (truncatable) | $0.13 | +8 to +10 points typical | yes |
Real products, models, and research that use this idea.
- OpenAI announcement of the text-embedding-3 family in January 2024, including the dimensions parameter that exposes Matryoshka truncation.
- Pinecone migration guides describe parallel-index patterns specifically for ada-002 to text-embedding-3-large cutovers.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you decide between text-embedding-3-small at 1536 dim and text-embedding-3-large truncated to 1536 dim?
Run both on a labeled retrieval eval set for your domain; 3-large truncated usually wins but the gap shrinks on simple corpora, and the cost difference matters at scale.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Still defaulting to ada-002 in new builds because the docs you read in 2023 said so, missing roughly 50% cost savings and a 5 to 10 point MTEB lift.
60 second bullets to scan on the way to the call.
Names of the two replacement OpenAI embedding models
Dimension sizes of small and large
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.