rdxtremity/jev-reranking

Search Stack — query towers

The query-side encoders for a two-stage product search over a multilingual (EN + AR) fashion catalogue. These are the exact weights that produced the vectors in the Qdrant collection they are used against — a cosine ANN is only meaningful if query and documents share an encoder, so these are part of the retrieval contract, not an interchangeable dependency.

path	size	what it is
adapter/	4.7 MB	LoRA (r=8, q/k/v) on BAAI/bge-m3 — the dense query tower
tokenizer/	17 MB	matching bge-m3 tokenizer
splade_v6_v2/	545 MB	fine-tuned SPLADE sparse encoder (distilbert-base-multilingual-cased)

The BAAI/bge-m3 backbone is not included — it is unmodified upstream and transformers fetches it on first run.

Use

from huggingface_hub import snapshot_download
snapshot_download("rdxtremity/jev-reranking",
                  allow_patterns=["adapter/*", "tokenizer/*", "splade_v6_v2/*"],
                  local_dir="models")

Dense query encoding is CLS-pool + L2-normalise, max 64 tokens, producing a 1024-d unit vector:

from transformers import AutoModel, AutoTokenizer
from peft import PeftModel
import torch

tok = AutoTokenizer.from_pretrained("models/tokenizer")
model = PeftModel.from_pretrained(AutoModel.from_pretrained("BAAI/bge-m3"),
                                  "models/adapter").eval()
enc = tok(["red dress"], return_tensors="pt", truncation=True, max_length=64)
with torch.no_grad():
    out = model(**enc)
vec = torch.nn.functional.normalize(out.last_hidden_state[:, 0], p=2, dim=-1)

Sparse encoding uses sentence_transformers.SparseEncoder on splade_v6_v2/, with max_seq_length = 128.

Only one short forward pass runs at query time, so CPU is sufficient.