Queryn — Embedding Translation

Move a corpus between embedding models without re-embedding it. Given a text chunk's embedding in Model A's space, a Queryn adapter returns the equivalent vector in Model B's space — so a vector index built with one model can be served against another after a lightweight transform instead of a full, expensive backfill.

What's in this org

The Queryn Embedding Adapters collection — one small ONNX model per directed model pair (queryn-adapter-<source>_to_<target>). Each repo contains:

model.onnx — the adapter, opset 17, dynamic batch axis. Runs anywhere onnxruntime runs; no PyTorch needed.
model.safetensors — the same weights, for retraining / inspection.
config.json — dimensions, the I/O contract, and provenance.
a model card with per-pair metrics and training plots.

Every adapter L2-normalizes its input and output internally, so you feed raw embeddings straight in and get unit vectors back.

Using an adapter

import numpy as np, onnxruntime as ort
from huggingface_hub import hf_hub_download

repo = "QuerynAi/queryn-adapter-ada-002_to_bge-m3"
sess = ort.InferenceSession(hf_hub_download(repo, "model.onnx"),
                            providers=["CPUExecutionProvider"])

src = np.random.rand(8, 1536).astype(np.float32)          # your ada-002 embeddings
tgt = sess.run(["target_embedding"], {"source_embedding": src})[0]
#  tgt: (8, 1024) unit vectors in bge-m3 space

How the adapters are built

One linear projection and one 1-hidden-layer GELU MLP (with a compressed latent below both dimensions) are trained per pair; whichever scores higher on held-out cosine similarity is published, ties to linear. Each pair's card reports both, so the linear-vs-nonlinear trade-off is visible.
Trained on a unified ~349,674-row, five-domain corpus (arXiv abstracts, Australian case law, SQuAD passages, PubMed RCT abstracts, financial/markets news), optimizing 1 − mean cosine similarity with Adam and LR scheduling.
Some embedding spaces are near-isomorphic and align well with a single matrix; others need the nonlinear map and still lose accuracy. Check the card metrics for your pair before relying on it — translation is an approximation, not a substitute for re-embedding when fidelity is critical.

Model coverage

Model	Dim	Role in v1
`ada-002`	1536	source only (deprecated as a target)
`te3-small`	1536	source + target
`qwen3-emb-8b`	4096	source + target
`bge-m3`	1024	source + target
`me5-large`	1024	source + target
`pplx-embed-1`	1024	source + target
`nemotron-1b-free`	2048	source + target
`fastembed-bge-small`	384	source + target

49 directed pairs in the current (v1) generation.

License

Adapter models: MIT. The Queryn codebase: Apache-2.0. Training data keeps each source corpus's own license — see the dataset's LICENSE manifest.

Status

An independent research project on practical embedding-space alignment. Issues and findings welcome via the repo; the adapters are provided as-is.