Forward-Looking Active REtrieval augmented generation (FLARE)

This repository contains the code and data for the paper Active Retrieval Augmented Generation.

Overview

FLARE is a generic retrieval-augmented generation method that actively decides when and what to retrieve using a prediction of the upcoming sentence to anticipate future content and utilize it as the query to retrieve relevant documents if it contains low-confidence tokens.

Install environment with Conda

Create a conda env and follow setup.sh to install dependencies.

Quick start

Download Wikipedia dump

Download the Wikipedia dump from the DPR repository using the following command:

mkdir data/dpr
wget -O data/dpr/psgs_w100.tsv.gz https://dl.fbaipublicfiles.com/dpr/wikipedia_split/psgs_w100.tsv.gz
pushd data/dpr
gzip -d psgs_w100.tsv.gz
popd

Build Wikipedia index

We use Elasticsearch to index the Wikipedia dump.

wget -O elasticsearch-7.17.9.tar.gz https://artifacts.elastic.co/downloads/elasticsearch/elasticsearch-7.17.9-linux-x86_64.tar.gz  # download Elasticsearch
tar zxvf elasticsearch-7.17.9.tar.gz
pushd elasticsearch-7.17.9
nohup bin/elasticsearch &  # run Elasticsearch in background
popd
python prep.py --task build_elasticsearch --inp data/dpr/psgs_w100.tsv wikipedia_dpr  # build index

Setup Bing search

This is only required for experiments on the WikiASP dataset.

Create a bing search API key following instructions on https://www.microsoft.com/en-us/bing/apis/bing-web-search-api.
Run a local bing search server with caching functionality to save credits: export BING_SEARCH_KEY=$YOUR_KEY; python bing_search_cache_server.py &> bing_log.out &.

Setup OpenAI keys

Put OpenAI keys in the keys.sh file. Multiple keys can be used to accelerate experiments. Please avoid uploading your keys to Github by accident!

Run FLARE

Use the following command to run FLARE with text-davinci-003.

./openai.sh 2wikihop configs/2wikihop_flare_config.json  # 2WikiMultihopQA dataset
./openai.sh wikiasp configs/wikiasp_flare_config.json  # WikiAsp dataset

Be careful, experiments are relatively expensive because FLARE calls OpenAI API multiple times for a single example. You can decrease max_num_examples to run small-scale experiments to save credits. Set debug=true to active the debugging mode which walks you through the iterative retrieval and generation process one example at a time.

Citation

@article{jiang2023flare,
      title={Active Retrieval Augmented Generation}, 
      author={Zhengbao Jiang and Frank F. Xu and Luyu Gao and Zhiqing Sun and Qian Liu and Jane Dwivedi-Yu and Yiming Yang and Jamie Callan and Graham Neubig},
      year={2023},
      eprint={2305.06983},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

jzbjyb / FLARE