@inproceedings{jin2025sara,
  title={SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression},
  author={Jin, Yiqiao and Sharma, Kartik and Rakesh, Vineeth and Dou, Yingtong and Pan, Menghai and Das, Mahashweta and Kumar, Srijan},
  booktitle={ACL},
  year={2026}
}

SARA represents retrieved evidence as a mix of k natural-language passages and n − k compressed retrieval embeddings. The compressed pieces are injected at <COMPRESS> token positions in the prompt through a small trained projector that maps retriever embeddings into the language model's hidden space, letting the model attend to many more documents within a fixed context budget.

This repository supports exactly two evaluated methods, sharing one dataset, retriever, model loader, generation code, metrics, and CLI:

• Standard RAG — all n retrieved documents inserted into the Mistral prompt as ordinary text.
• SARA — the same n candidates, but k rendered as text and n − k as compressed embeddings.

Both start from the public mistralai/Mistral-7B-Instruct-v0.2 base and are trained with matched budgets, so the comparison is fair by construction.

• 🎯 Two-method, apples-to-apples: Standard RAG vs SARA under identical data, retriever, prompts, decoding, and seeds.
• 🧠 Projector + <COMPRESS> tokens: retriever embeddings projected into the LM hidden space at compress positions.
• 🪶 LoRA-first: adapters + projector only — no merged or full-model checkpoints are ever saved.
• ⚡ Fully batched generation and evaluation (dynamic padding, one GPU call per batch).
• 🔒 Leakage-safe: document-level (per-paper) train/dev/test splits.
• 🛠️ Single CLI: python -m src.run_sara {finetune, evaluate, report} --method {rag, sara}.

SARA targets Python 3.11 + CUDA 12.x. Install PyTorch from the cu128 wheel index first, then the rest.

# 1) Create and activate the env
conda create -n sara python=3.11 -y
conda activate sara
pip install -U pip setuptools wheel

# 2) Install cu128 PyTorch from the PyTorch wheel index
pip install --index-url https://download.pytorch.org/whl/cu128 torch torchvision torchaudio

# 3) Install SARA + its dependencies (editable)
pip install -e .
#   or:  pip install -r requirements.txt

💡 Flash-Attention is optional. The model loads with PyTorch SDPA by default, so no flash-attn build is required.

🔑 Set HF_TOKEN to access the gated Mistral base. WANDB_API_KEY is optional (only if you set use_wandb: true).

python -c "
import torch, transformers, peft
from src.model import XMistralForCausalLM
print('torch       ', torch.__version__, 'cuda_ok:', torch.cuda.is_available())
print('transformers', transformers.__version__)
print('peft        ', peft.__version__)
print('SARA model  ', XMistralForCausalLM.__name__, 'OK')
"

All commands run on GPU 0 (set CUDA_VISIBLE_DEVICES to choose devices). QASPER is the primary benchmark.

python -m src.data.make_qasper_splits --dev-frac 0.1 --seed 42

# Standard RAG  (LoRA only; all n documents as text)
CUDA_VISIBLE_DEVICES=0 python -m src.run_sara finetune \
    --config config/language_modeling/Finetune_RAG_qasper_pubbase.yaml

# SARA  (LoRA + projector; k text + (n-k) compressed)
CUDA_VISIBLE_DEVICES=0 python -m src.run_sara finetune \
    --config config/language_modeling/Finetune_SARA_qasper_pubbase.yaml

🪄 Optional — projector alignment warm-up. Pre-train the SARA projector to reconstruct documents from their compressed embeddings before fine-tuning (set init_projector_from in the SARA config to the output):

CUDA_VISIBLE_DEVICES=0 python -m src.train.align_projector \
    --train data/reformatted/QASPER_paragraphs_train.jsonl \
    --out checkpoints/paragraph/sara_qasper_align --steps 600 --batch-size 8 --lr 5e-4

# Standard RAG
CUDA_VISIBLE_DEVICES=0 python -m src.run_sara evaluate --method rag \
    --models rag_qasper --retrievers bm25 --datasets qasper --n 10 --batch-size 16

# SARA
CUDA_VISIBLE_DEVICES=0 python -m src.run_sara evaluate --method sara \
    --models sara_qasper --retrievers bm25 --datasets qasper --k 5 --n 10 --batch-size 16

python -m src.run_sara report \
    --rag  outputs/rag_qasper/answers_rag_qasper_bm25_qasper_k10_n10_rep1.5.jsonl \
    --sara outputs/sara_qasper/answers_sara_qasper_bm25_qasper_k5_n10_rep1.5.jsonl \
    --out-dir outputs/docs/sara_vs_rag_qasper

Evidence allocation is controlled by --k and --n:

Unsupported --method values are rejected with a clear error.

python -m pytest tests/test_xmistral.py \
                 tests/test_checkpoint_roundtrip.py \
                 tests/test_train_sara_sanity.py -q

These cover the <COMPRESS> embedding injection (batched vs single-example parity), checkpoint round-trip (LoRA + projector + added-token rows reload numerically), and the training gradient/optimizer sanity checks.

Released under the Apache License 2.0.

Built on excellent open-source work: 🤗 Transformers · 🤗 PEFT · 🤗 Accelerate · LlamaIndex.