Practical overview
When it fits
Use Sara for local GPU experiments comparing ordinary text-in-context RAG with a method that compresses some retrieved evidence into embeddings, particularly within its documented QASPER workflow.
Who it is for
- Researchers studying retrieval-augmented generation and context compression
- Developers building local RAG experiments who want a documented baseline and SARA comparison
- Teams evaluating retrieval and compression approaches on the QASPER benchmark
Documented uses
- Prepare document-level QASPER train, development and test splits
- Fine-tune standard RAG and SARA adapters from the Mistral base
- Run batched evaluation using the documented methods, retriever and dataset settings
- Generate a comparison report from RAG and SARA evaluation outputs
- Optionally warm up the SARA projector by training it to reconstruct documents from compressed embeddings
Requirements & limits
The repository is public and licensed under Apache-2.0. The documented setup targets Python 3.11 and CUDA 12.x. Access to the gated Mistral base requires an HF_TOKEN.
- The implementation is a research repository, not a hosted service or API.
- The documented baseline uses Mistral-7B-Instruct-v0.2, QASPER and BM25.
- Training and evaluation commands are GPU-oriented; the README directs commands to GPU 0 by default.
- The SARA variant adds a trained projector to LoRA adapters, while standard RAG uses LoRA only.
- Flash-Attention is optional; the documented default is PyTorch SDPA.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up Sara’s research environment
Prepare the documented Python environment and install Sara and its dependencies.
Before you start
- Linux environment with Python 3.11 and CUDA 12.x
- Conda
- Access to the gated Mistral base via an HF_TOKEN for later model use
Steps
Create and activate a Python 3.11 Conda environment named sara.
Install PyTorch, torchvision and torchaudio from the cu128 PyTorch wheel index.
Install Sara in editable mode with pip install -e ., or install its requirements file.
Expected outcome
Sara’s documented dependencies are installed in the environment for the subsequent benchmark workflow.
Guide limitations
- The guide covers installation only; it does not verify successful execution.
- Model access requires an HF_TOKEN for the gated Mistral base.
- Flash-Attention is optional; the README says PyTorch SDPA is used by default.
Prepare QASPER splits and train the two benchmark methods
Create the documented QASPER splits and launch fine-tuning for standard RAG and SARA.
Before you start
- Sara environment installed
- Python 3.11 and CUDA 12.x with GPU access
- HF_TOKEN for access to the gated Mistral base
Steps
Create the document-level splits with python -m src.data.make_qasper_splits --dev-frac 0.1 --seed 42.
Launch standard RAG fine-tuning with the Finetune_RAG_qasper_pubbase.yaml configuration.
Launch SARA fine-tuning with the Finetune_SARA_qasper_pubbase.yaml configuration.
Expected outcome
The documented commands prepare QASPER data and start training a standard RAG adapter and a SARA adapter; the latter includes a projector.
Guide limitations
- The README’s training commands target GPU 0 and do not provide hardware sizing guidance.
- This recipe stops at fine-tuning; evaluation and reporting are separate CLI stages.
- The SARA method represents k candidates as text and n − k as compressed embeddings.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 12 October 2026
- Original website title
- Ahren09/SARA · Official homepage title · retained in its source language