UniVec

UniVec publishes vector conversion models that map embeddings from one model's vector space into another without re-embedding the source text.

What is vector conversion?

A corpus embedded with a particular model is bound to that model's vector space: queries must be encoded by the same model for nearest-neighbour search to remain meaningful. Migrating to a different embedder (whether driven by deprecation, an upgrade or a provider change) normally requires re-embedding every document. The cost scales with corpus size and recurs each time the underlying model changes.

A conversion model takes pre-computed source-space vectors and outputs target-space vectors. The training objective is retrieval-order preservation: top-K nearest neighbours in the converted space should align with top-K in the target space despite differences in dimensionality, distance distribution and noise structure.

The category sits adjacent to embedding generation (text -> vector) and embedding distillation (large model -> small model) but addresses a separate problem: relocating existing vectors between spaces.

What's published here

Two tracks, both released under Apache 2.0.

General-purpose converters are trained on broad, heterogeneous data to generalise across domains. These are the recommended models for production self-hosting. Coverage targets the most-requested source/target pairs across OpenAI, Cohere, Google, AWS Titan, Snowflake Arctic, BGE and GTE.

Benchmarking converters are trained on MTEB-aligned distributions and serve as reference points against published translation results. Open weights, open metrics, identical evaluation script.

Both tracks ship as single ONNX files. They run on CPU or GPU via ONNX Runtime, take a (batch, source_dim) array of unit-normalized vectors and return (batch, target_dim).

Available models

The catalog below lists what is currently published on the Hub. The hosted API at https://univec.ai covers additional pairs and bridge configurations.

General-purpose

Source Target Model
Alibaba GTE Large EN v1.5 Nomic Embed Text v1.5 Conversion model
Amazon Titan Embed Text v2.0 OpenAI text-embedding-ada-002 Conversion model
BAAI BGE-M3 OpenAI text-embedding-ada-002 Conversion model
BAAI BGE-M3 Snowflake Arctic Embed L v2.0 Conversion model
Cohere Embed English v3.0 OpenAI text-embedding-ada-002 Conversion model
Google EmbeddingGemma 300M OpenAI text-embedding-ada-002 Conversion model
OpenAI text-embedding-3-large Google Gemini text-embedding-004 Conversion model
OpenAI text-embedding-3-small OpenAI text-embedding-ada-002 Conversion model
Snowflake Arctic Embed L v2.0 BAAI BGE-M3 Conversion model
Snowflake Arctic Embed L v2.0 OpenAI text-embedding-ada-002 Conversion model

Benchmarking (MTEB-aligned)

Source Target Model
BAAI BGE-M3 Snowflake Arctic Embed L v2.0 Conversion model
Cohere Embed English v3.0 OpenAI text-embedding-ada-002 Conversion model
Google EmbeddingGemma 300M OpenAI text-embedding-ada-002 Conversion model
OpenAI text-embedding-3-small OpenAI text-embedding-ada-002 Conversion model
Snowflake Arctic Embed L v2.0 BAAI BGE-M3 Conversion model

Quick start

pip install onnxruntime numpy
import numpy as np
import onnxruntime as ort

session = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
input_name = session.get_inputs()[0].name

# Source-model embeddings, shape (N, source_dim), float32
embeddings = ...

converted = session.run(None, {input_name: embeddings.astype("float32")})[0]
# converted has shape (N, target_dim) in the target model's space.

Each model repository also ships univec_inference.py, a CLI covering batching, GPU execution and .npy / .jsonl input.

What's not published here

The set above is a curated subset of the UniVec catalog. The full catalog covers around 100 conversion pairs and includes bridge configurations (two-hop conversions through an intermediary model when no direct pair is trained). The hosted API at https://univec.ai exposes the unpublished pairs along with managed inference.

License

Apache 2.0. Free for commercial use, redistribution and fine-tuning.

Links