VerifAI

VerifAI is a Generative Search/Productivity engine with Verifiable answers. Please check our website https://verifai-project.com/, or deployed app at https://app.verifai-project.com/

No more searches, just verifiably accurate answers.

Table of Contents

• Project Description
• Main Features
• Support This Project
• Installation and Start-up
  • VerifAI Core
  • VerifAI BioMed
• Developed Models and Datasets
• Using our APP
• Collaborators and Contributions
• Papers and Citations
• Funding

Project Description

VerifAI is a document-based question-answering systems that aims to address problem of hallucinations in generative large language models and generative search engines. Initially, we started with biomedical domain, however, now we have expanded VerifAI to support indexing any documents in txt,md, docx, pptx, or pdf formats.

VerifAI is an AI system designed to answer users' questions by retrieving the most relevant documents, generate answer with references to the relevant documents and verify that the generated answer does not contain any hallucinations. In the core of the engine is generative search engine, powered by open technologies. However, generative models may hallucinate, and therefore VerifAI is developed a second model that would check the sources of generative model and flag any misinformation or misinterpretations of source documents. Therefore, make the answer created by generative search engine completly verifiable.

The best part is, that we are making it open source, so anyone can use it!

Check the article about VerifAI project published on TowardsDataScience

How does it work:

Main features

• Easy installation by running a single script
• Easy indexing of local files in PDF,EPUB, PPTX, DOCX, MD and TXT formats
• Combination of lexical and semantic search to find the most relevant documents
• Usage of any HuggingFace listed model for document embeddings
• Usage of any LLM that follows OpenAI API standard (deployed using vLLM, Nvidia NIM, Ollama, or via commercial APIs, such as OpenAI, Azure)
• Supports large amounts of indexed documents (tested with over 200GB of data and 30 million documents)
• Shows the closest sentence in the document to the generated claim
• User registration and log-in
• Pleasent user interface developed in React.js
• Verification that generated text does not contain hallucinations by a specially fine-tuned model
• Possible single-sign-on with AzureAD (future plans to add other services, e.g. Google, GitHub, etc.)

⭐️ Support This Project

If you find this project helpful or interesting, please consider giving it a star on GitHub! Your support helps make this project more visible to others who might benefit from it.

By starring this repository, you'll also stay updated on new features and improvements. Thank you for your support! 🙏

Installation and start-up

VerifAI Core

1. Clone the repository or download latest release
2. Create virtual python environment by running:

python -m venv verifai
source verifai/bin/activate

3. In case you get errors with installing psycopg2, you may need to install postgres by running sudo apt install postgresql-server-dev-all
4. On a clean instance you may need to run:

sudo sh -c 'echo "deb http://apt.postgresql.org/pub/repos/apt $(lsb_release -cs)-pgdg main" > /etc/apt/sources.list.d/pgdg.list'
wget --quiet -O - https://www.postgresql.org/media/keys/ACCC4CF8.asc | sudo apt-key add -
sudo apt update

5. Run requirements.txt by running pip install -r backend/requirements.txt
6. Configure system, by replacing and modifying .env.local.example in backend folder and rename it into just .env: The configuration should look in the following manner:

SECRET_KEY=6183db7b3c4f67439ad61d1b798224a035fe35c4113bf870
ALGORITHM=HS256

DBNAME=verifai_database
USER_DB=myuser
PASSWORD_DB=mypassword
HOST_DB=localhost

OPENSEARCH_IP=localhost
OPENSEARCH_USER=admin
OPENSEARCH_PASSWORD=admin
OPENSEARCH_PORT=9200
OPENSEARCH_USE_SSL=False
QDRANT_IP=localhost
QDRANT_PORT=6333
QDRANT_API=8da7725d78141e19a9bf3d878f4cb333fedb56eed9727904b46ce4b32e1ce085
QDRANT_USE_SSL=False

OPENAI_PATH=<path-to-openai/azure/vllm/nvidia_nim/ollama-interface>
OPENAI_KEY=<key-in-interface>
DEPLOYMENT_MODEL=GPT4o
MAX_CONTEXT_LENGTH=128000
SIMILARITY_METRIC=DOT
VECTOR_SIZE=768

EMBEDDING_MODEL="sentence-transformers/msmarco-bert-base-dot-v5"

INDEX_NAME_LEXICAL = 'myindex-lexical'
INDEX_NAME_SEMANTIC = "myindex-semantic"

USE_VERIFICATION=True

Please note that value of SIMILARITY_METRIC can be either DOT (dot product) or COSINE (cosine similarity). If not stated, it will resolve to Cosine similarity.

7. Run install_datastores.py file. To run this file, it is necessary to install Docker (and run the daemon). This file is designed to install necessary components, such as OpenSearch, Qdrant and PostgreSQL, as well as to create database in PostgreSQL.

python install_datastore.py

8. Index your files, by running index_files.py and pointing it to the directory with files you would like to index. It will recuresevly index all files in the directory.

python index_files.py <path-to-directory-with-files>

As an example, we have created a folder with some example files in the folder test_data. You can index them by running:

python index_files.py test_data

9. Run the backend of VerifAI by running main.py in the backend folder.

python main.py

10. Install React by following this guide, or by running following commands:

sudo apt update
sudo apt install nodejs npm
sudo npm install -g create-react-app

11. Install React requirements for the front-end in client-gui/verifai-ui folder and run front end:

cd ..
cd client-gui/verifai-ui
npm install

Create .env file in the client-gui/verifai-ui folder with the following content (or based on .env.example file):

REACT_APP_BACKEND = http://127.0.0.1:5001/ # or your API url
REACT_APP_AZURE_CLIENT_ID=<your_azure_client_id>
REACT_APP_AZURE_TENANT_ID=<your_azure_tenant_id>
REACT_APP_AZURE_REDIRECT_URL=http://localhost:3000

If you do not configure REACT_APP_AZURE_CLIENT_ID and REACT_APP_AZURE_TENANT_ID, the app will not have the option to log in with AzureAD. Your AzureAD application needs to be registered as Single-Page Application in Azure. Change REACT_APP_AZURE_REDIRECT_URL to the redirect URL matching one in Azure. Start the app by running:

npm start

12. Go to http://localhost:3000 to see the VerifAI in action.

You can check a tutorial on deploying VerifAI published on Towards Data Science

VerifAI BioMed

This is biomedical version of VerifAI. It is designed to answer questions from the biomedical domain.

One requirement to run locally is to have installed Postgres SQL. You can install it for example on mac by running brew install postgresql.

1. Clone the repository
2. Run requirements.txt by running pip install -r backend/requirements.txt
3. Download Medline. You can do it by executing download_medline_data.sh for core files for the current year and download_medline_data_update.sh for Medline current update files.
4. Install Qdrant following the guide here
5. Run the script: python medline2json.py to transform MEDLINE XML files into JSON
6. Run python json2selected.py to selects the fields that should be inported into the index
7. Run python abstarct_parser.py to concatinate abstract titles and abstracts and splits texts to 512 parts that can be indexed using a transformer model
8. Run python embeddings_creation.py to create embeddings.
9. Run python scripts/indexing_qdrant.py to create qdrant index. Make sure to point to the right folder created in the previous step and to the qdrant instance.
10. Install OpenSearch following the guide here
11. Create OpenSearch index by running python scripts/indexing_lexical_pmid.py. Make sure to configure access to the OpenSearch and point the path variable to the folder created by json2selected script.
12. Set up system variables that are needed for the project. You can do it by creating .env file with the following content:

OPENSEARCH_IP=open_search_ip
OPENSEARCH_USER=open_search_user
OPENSEARCH_PASSWORD=open_search_pass
OPENSEARCH_PORT=9200
QDRANT_IP=qdrant_ip
QDRANT_PORT=qdrant_port
QDRANT_API=qdrant_api_key
QDRANT_USE_SSL=False
OPENSEARCH_USE_SSL=False
MAX_CONTEXT_LENGTH=32000


EMBEDDING_MODEL="sentence-transformers/msmarco-bert-base-dot-v5"

INDEX_NAME_LEXICAL = 'medline-faiss-hnsw-lexical-pmid'
INDEX_NAME_SEMANTIC = "medline-faiss-hnsw"

USE_VERIFICATION=True

13. Run backend by running python backend/main.py
14. Install React by following this guide
15. Run npm run-script build
16. Run frontend by running npm start in client-gui/verifai-ui

Developed models and datasets

1. Fine tuned QLoRA addapted for Mistral 7B-instruct v01
2. Fine tuned QLoRA addapted for Mistral 7B-instruct v02
3. PQAref dataset
4. Verification model based on DeBERTa, fine-tuned on SciFact dataset

Using our APP

You can use our app here. You need to create a free account by clicking on Join now.

Collaborators and contributions

Currently, two institutions are the main drivers of this project, namely Bayer A.G and Institute for Artificial Intelligence Research and Development of Serbia. Current contrbiutors are by institutions

• Bayer A.G.
  • Nikola Milosevic
  • Lorenzo Cassano
• Institute for Artificial Intelligence Research and Development of Serbia:
  • Adela Ljajic
  • Milos Kosprdic
  • Bojana Basaragin
  • Darija Medvecki
  • Angela Pupovac
  • Nataša Radmilović
  • Petar Stevanović

We welcome contribution to this project by anyone interested in participating. This is an open source project under AGPL license. In order to prevent any legal issues, before sending the first pull request, we ask potential contributors to sign Individual Contributor Agreement and send to us via email ([email protected]).

Papers and Citations

• Adela Ljajić, Miloš Košprdić, Bojana Bašaragin, Darija Medvecki, Lorenzo Cassano, Nikola Milošević, “Scientific QA System with Verifiable Answers”, The 6th International Open Search Symposium 2024
• Košprdić, M., Ljajić, A., Bašaragin, B., Medvecki, D., & Milošević, N. "Verif. ai: Towards an Open-Source Scientific Generative Question-Answering System with Referenced and Verifiable Answers." The Sixteenth International Conference on Evolving Internet INTERNET 2024 (2024).
• Bojana Bašaragin, Adela Ljajić, Darija Medvecki, Lorenzo Cassano, Miloš Košprdić, Nikola Milošević "How do you know that? Teaching Generative Language Models to Reference Answers to Biomedical Questions", Accepted at BioNLP 2024, Colocated with ACL 2024
• Adela Ljajić, Lorenzo Cassano, Miloš Košprdić, Bašaragin Bojana, Darija Medvecki, Nikola Milošević, "Enhancing Biomedical Information Retrieval with Semantic Search: A Comparative Analysis Using PubMed Data", Belgrade Bioinformatics Conference BelBi2024, 2024
• Košprdić, M.; Ljajić, A.; Medvecki, D.; Bašaragin, B. and Milošević, N. (2024). Scientific Claim Verification with Fine-Tuned NLI Models. In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management - KMIS; ISBN 978-989-758-716-0; ISSN 2184-3228, SciTePress, pages 15-25. DOI: 10.5220/0012900000003838

Funding

This project was in September 2023 funded by NGI Search project of the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or European Commission. Neither the European Union nor the granting authority can be held responsible for them. Funded within the framework of the NGI Search project under grant agreement No 101069364