Skip to content

🎯 Pointwise Re-Ranking (MonoBERT, MonoT5)

Pointwise rerankers score each query-document pair independently, then sort by score.

MonoT5

MonoT5 uses T5 to predict relevance ("true" or "false"):

from rankify.dataset.dataset import Document, Question, Context
from rankify.models.reranking import Reranking

# Create document
question = Question("What is the speed of light?")
contexts = [
    Context(text="The speed of light in vacuum is 299,792,458 m/s.", id="1"),
    Context(text="Light travels faster than sound.", id="2"),
    Context(text="Einstein developed the theory of relativity.", id="3"),
]
document = Document(question=question, contexts=contexts)

# MonoT5 reranking
reranker = Reranking(method="monot5", model_name="monot5-base-msmarco")
reranked = reranker.rank([document])

# Results
for ctx in reranked[0].reorder_contexts:
    print(f"[{ctx.score:.4f}] {ctx.text[:80]}...")

Available MonoT5 Models

Model Name Size HuggingFace ID
monot5-base-msmarco 220M castorini/monot5-base-msmarco
monot5-large-msmarco 770M castorini/monot5-large-msmarco
monot5-3b-msmarco-10k 3B castorini/monot5-3b-msmarco-10k 
# Using a larger model
reranker = Reranking(method="monot5", model_name="monot5-large-msmarco")

MonoBERT

MonoBERT uses BERT cross-encoders:

reranker = Reranking(method="monobert", model_name="monobert-large")
reranked = reranker.rank([document])

UPR (Unsupervised Passage Reranker)

UPR uses language models without task-specific training:

# T5-based UPR
reranker = Reranking(method="upr", model_name="t5-base")
reranked = reranker.rank([document])

# GPT-2 based UPR
reranker = Reranking(method="upr", model_name="gpt2")
reranked = reranker.rank([document])

Available UPR Models

Model Description
t5-small Google T5 Small
t5-base Google T5 Base
t5-large Google T5 Large
gpt2 OpenAI GPT-2
gpt-neo-2.7b EleutherAI GPT-Neo
flan-t5-xl Google Flan-T5 XL

FlashRank (Fast ONNX Reranking)

FlashRank uses ONNX for fast CPU inference:

# Fast, lightweight reranking
reranker = Reranking(method="flashrank", model_name="ms-marco-TinyBERT-L-2-v2")
reranked = reranker.rank([document])

FlashRank Models

Model Speed Quality
ms-marco-TinyBERT-L-2-v2 ⚡⚡⚡ Good
ms-marco-MiniLM-L-12-v2 ⚡⚡ Very Good
rank-T5-flan ⚡ Excellent

Batch Processing

Process multiple documents efficiently:

# Create multiple documents
documents = [
    Document(question=Question("Who invented the telephone?"), contexts=[...]),
    Document(question=Question("What is DNA?"), contexts=[...]),
    Document(question=Question("When was the moon landing?"), contexts=[...]),
]

# Batch reranking
reranker = Reranking(method="monot5", model_name="monot5-base-msmarco")
reranked_docs = reranker.rank(documents)

for doc in reranked_docs:
    print(f"Q: {doc.question.question}")
    print(f"Top result: {doc.reorder_contexts[0].text[:100]}...")

Next Steps