TART

Task-Aware Retrieval with Instructions.

Overview

Field	Value
Type	Instruction-T5
Library	torch, transformers
Default Model	facebook/tart-full-flan-t5-xl
Paper	Asai et al., 2022

Installation

pip install "autorag-research[gpu]"
# or
uv add "autorag-research[gpu]"

Configuration

_target_: autorag_research.rerankers.tart.TARTReranker
model_name: facebook/tart-full-flan-t5-xl

Options

Option	Type	Default	Description
model_name	str	facebook/tart-full-flan-t5-xl	TART model name
instruction	str	Find passage to answer given question	Task instruction
max_length	int	512	Maximum input sequence length
device	str	None	Device (auto-detected)
batch_size	int	64	Batch size for multiple queries

How It Works

1. Prepend task instruction to query: "{instruction} [SEP] {query}"
2. Encode instruction-query with document as input pair
3. Apply softmax to classification logits
4. Use positive class probability as relevance score

Usage

from autorag_research.rerankers import TARTReranker

reranker = TARTReranker(instruction="Find passage to answer given question")
results = reranker.rerank("What is RAG?", ["doc1", "doc2", "doc3"], top_k=2)