QuerynAi/queryn-adapter-fastembed-bge-small_to_pplx-embed-1

Queryn adapter — fastembed-bge-small → pplx-embed-1

Translates an embedding produced by fastembed-bge-small into the embedding space of pplx-embed-1, so a corpus already embedded with fastembed-bge-small can be served against a pplx-embed-1 index without re-embedding it. Part of the Queryn embedding-translation engine.

Specs

Source model	fastembed-bge-small (384-d)
Target model	pplx-embed-1 (1024-d)
Architecture	deep (1 hidden layer, GELU, compressed latent)
Parameters	~271.6K
Best test cosine similarity	0.6420 (epoch 15)
ONNX opset	17

Architecture ablation (best test cosine): linear 0.6351, deep 0.6420 ← saved.

Input / output contract

• Input source_embedding — float32, shape [batch, 384]. Raw fastembed-bge-small embeddings; the graph L2-normalizes them itself, so pre-normalization is neither required nor harmful.
• Output target_embedding — float32, shape [batch, 1024], unit-normalized, in pplx-embed-1 space.
• Batch axis is dynamic.

Usage

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

path = hf_hub_download("QuerynAi/queryn-adapter-fastembed-bge-small_to_pplx-embed-1", "model.onnx")
sess = ort.InferenceSession(path, providers=["CPUExecutionProvider"])

src = np.random.rand(4, 384).astype(np.float32)   # your fastembed-bge-small embeddings
tgt = sess.run(["target_embedding"], {"source_embedding": src})[0]
assert tgt.shape == (4, 1024)                     # unit vectors in pplx-embed-1 space

Training

Trained on paired embeddings over a unified multi-domain corpus — arXiv abstracts, Australian case law, SQuAD passages, PubMed abstracts, and crypto/markets news (~350k rows spanning science, legal, QA, medical, and finance). Loss: 1 - mean cosine similarity, Adam, ReduceLROnPlateau, best-epoch checkpoint. Both a linear baseline and the MLP are trained for every pair; the higher-scoring one is published (ties go to linear).

Plots

fastembed-bge-small → all targets: learning curves and best scores (this pair included).

Linear vs. deep for every pair (black ring = saved architecture).

Full adapter set: Queryn Embedding Adapters

Provenance

• Source checkpoint: models/v1/fastembed-bge-small_to_pplx-embed-1.pt (sha256 5e0eb66340f68769…)
• Converted: 2026-08-30T20:45:22+00:00 · torch 2.13.0 · ptConverter.py

License

Released under the MIT license.