Cross-encoder reranking via SentenceTransformers.
Overview
| Field |
Value |
| Type |
CrossEncoder |
| Library |
sentence-transformers |
| Default Model |
cross-encoder/ms-marco-MiniLM-L-2-v2 |
Installation
pip install "autorag-research[gpu]"
# or
uv add "autorag-research[gpu]"
Configuration
_target_: autorag_research.rerankers.sentence_transformer.SentenceTransformerReranker
model_name: cross-encoder/ms-marco-MiniLM-L-2-v2
Options
| Option |
Type |
Default |
Description |
| model_name |
str |
cross-encoder/ms-marco-MiniLM-L-2-v2 |
HuggingFace model name |
| max_length |
int |
512 |
Maximum input sequence length |
| device |
str |
None |
Device (auto-detected) |
| batch_size |
int |
64 |
Batch size for multiple queries |
Models
| Model |
Description |
| cross-encoder/ms-marco-MiniLM-L-2-v2 |
Fast, lightweight |
| cross-encoder/ms-marco-MiniLM-L-6-v2 |
Balanced |
| cross-encoder/ms-marco-MiniLM-L-12-v2 |
Best quality |
Usage
from autorag_research.rerankers import SentenceTransformerReranker
reranker = SentenceTransformerReranker()
results = reranker.rerank("What is RAG?", ["doc1", "doc2", "doc3"], top_k=2)
for r in results:
print(f"[{r.index}] {r.score:.3f}: {r.text[:50]}...")