CitationFinder

CitationFinder is an application designed to streamline the process of finding academic citations by leveraging large language models, RAG, and web scraping. It helps researchers, scholars, and students alike quickly locate supporting citations from academic literature, significantly reducing the time and effort required for this often surprisingly cumbersome task.

Overview

In scholarly contexts, providing a citation for your claims is essential. However, finding a suitable citation that perfectly matches your specific needs can be challenging. This application automates the process of searching for relevant articles, extracting pertinent information, and evaluating the quality of the citations.

(Needless to say, actual research should work the other way around — that you construct your arguments from knowledge of prior literature; not that you find prior literature to support any arbitrary argument. This project is intended primarily as a para-novelty experiment, designed to showcase and practice skills with LangChain. While it may have some real-world utility, users should not treat this as a substitute for rigorous academic research.)

Example usage

python main.py --input_sentence "Covid 19 increased the likelihood of heart complications"

=========================================== [Citation 1] ===========================================

TITLE:    Long COVID: post-acute sequelae of COVID-19 with a cardiovascular focus
YEAR:     2022
AUTHORS:  Betty Raman; David A. Bluemke; Thomas F. Lüscher; Stefan Neubauer
URL:      https://www.ncbi.nlm.nih.gov/pmc/articles/pmid/35176758
DOI:      10.1093/eurheartj/ehac031

"Presence of concurrent cardiac disease is typically associated with higher in-hospital mortality,
thrombo-embolic risk, and septic shock rates.97,194 Even in the post-acute period, patients with a
history of heart failure are at a two- to four-fold risk of decompensation and mortality.195,196
Increased rates of heart failure exacerbation may present even beyond 30 days after SARS-CoV2
infection.98"

=========================================== [Citation 2] ===========================================

TITLE:    Impact of COVID-19 on Cardiovascular Disease
YEAR:     2023
AUTHORS:  Ivan Vosko; Andreas Zirlik; Heiko Bugger
URL:      https://www.ncbi.nlm.nih.gov/pmc/articles/pmid/36851722
DOI:      10.3390/v15020508

"Twelve months following COVID-19, the HR was increased for literally all CVDs that were
investigated: stroke 1.52; transient ischemic attack 1.49; AF 1.71; sinus tachycardia 1.84; sinus
bradycardia 1.53; ventricular arrhythmias 1.84; cardiac arrest 2.45; pericarditis 1.85; myocarditis
5.38; myocardial infarction 1.63; ischemic cardiomyopathy 1.75; non-ischemic cardiomyopathy 1.62;
pulmonary embolism 2.93; deep vein thrombosis 2.09."

=========================================== [Citation 3] ===========================================

TITLE:    COVID-19 and cardiovascular diseases
YEAR:     2020
AUTHORS:  Fan Liu; Feng Liu; Lu Wang
URL:      https://www.ncbi.nlm.nih.gov/pmc/articles/pmid/33226078
DOI:      10.1093/jmcb/mjaa064

"Heart failure is also a common complication of COVID-19. A cohort study shows that the incidence of
heart failure is 23% in all 191 patients and 49% in non-survived patients (Zhou et al., 2020a).
Another study also supports the prevalence of heart failure as COVID-19-related complications and
elevated levels of amino-terminal pro-B-type natriuretic peptide in almost half of the patients
(Chen et al., 2020c)."

=========================================== [Citation 4] ===========================================

TITLE:    Cardiovascular System during SARS-CoV-2 Infection
YEAR:     2022
AUTHORS:  Maciej Koźlik; Adrianna Błahuszewska; Maciej Kaźmierski
URL:      https://www.ncbi.nlm.nih.gov/pmc/articles/pmid/35162207
DOI:      10.3390/ijerph19031184

"The results show that cardiac injury is present in about 1 in 4 patients with COVID-19 disease, and
it is an independent risk factor, which multiplies the death rate several times in comparison to
infected patients without myocardial injury."

=========================================== [Citation 5] ===========================================

TITLE:    COVID-19 and Cardiovascular Disease
YEAR:     2021
AUTHORS:  Mina K. Chung; David A. Zidar; Michael R. Bristow; Scott J. Cameron; Timothy Chan; Clifford V. Harding, III; Deborah H. Kwon; Tamanna Singh; John C. Tilton; Emily J. Tsai; Nathan R. Tucker; John Barnard; Joseph Loscalzo
URL:      https://www.ncbi.nlm.nih.gov/pmc/articles/pmid/33856918
DOI:      10.1161/CIRCRESAHA.121.317997

"Acute cardiac injury, inferred from elevations in cTn (cardiac troponin) levels, is reported in 8%
to 62% of patients hospitalized with COVID-19 and is associated with greater disease severity,
including need for mechanical ventilation, and death."

Supported Databases

As of now, CitationFinder includes support only for the PubMed database.

• PubMed is a free resource that primarily comprises MEDLINE, a database of references and abstracts on life sciences and biomedical topics. It is maintained by the National Center for Biotechnology Information (NCBI) at the U.S. National Library of Medicine (NLM). PubMed provides access to a vast repository of medical literature, including research articles, reviews, and clinical studies.

(Including additional databases is primarily a matter of developing article scraping logic for these databases.)

How It Works

The workflow of CitationFinder is as follows:

1. Input Sentence: The user inputs a sentence into the application, for which they seek supporting citations from academic literature.
2. Search Query Translation: A language model transforms the sentence into N (default: 5) search queries / phrases optimized for database searches.
3. Database Search: The application searches an article database for the K (default: 10) most relevant articles for each search query, retrieving their texts and metadata.
4. Document Processing: The application breaks down the texts into smaller paragraphs, treating each paragraph as a "document".
5. Embedding Documents: Each document is embedded, i.e., converted into a numerical vector that encodes its semantic content.
6. Vector Storage: The vectors are stored in a vector database, allowing for efficient retrieval based on similarity search.
7. User Input Sentence Embedding: The user's original input sentence is embedded into a vector.
8. Document Retrieval: The application retrieves J (default: 10) documents from the vector database whose vectors most closely match the user's input sentence vector.
9. Citation Evaluation: A language model evaluates each of the retrieved documents to determine whether it can be used as a direct source for the user's input sentence. In addition, the LLM extracts the most relevant sentence or sentences from the paragraph.

Usage Instructions

Environment Setup

Before running the application, you need to set up the required environment variables:

1. OPENAI_API_KEY: This is necessary to use the language models. You can set it up in your terminal as follows:

   export OPENAI_API_KEY='<your-openai-api-key>'

2. LANGCHAIN_API_KEY: This is required if you wish to use LangSmith for enhanced tracing and analysis. Set it up with:

   export LANGCHAIN_API_KEY='<your-langchain-api-key>'

   Make sure to enable LangSmith usage in the configuration file by setting use_langsmith to True.

Dependency management

You can manage dependencies using either Poetry or pip:

• Using poetry

  poetry install

• Using pip

  pip install requirements.txt

Running the Application

To run the application, navigate to the src directory and execute the following command:

python main.py --sentence "<Your input sentence here>"

If you want to see how the application works with the default configuration, you can simply run:

python main.py

This will automatically search for citations for the sentence "Covid 19 increased the likelihood of heart complications."

Configuration

Please refer to the following flags for how to conifgure the application:

• --input_sentence: The sentence for which you want to find citations. This argument is required.
• --n_articles: The number of articles to retrieve per search query (default: 10).
• --n_docs: The number of documents to retrieve from the vector database (default: 10).
• --model_name: The name of the OpenAI language model to use (default: "gpt-4o-2024-08-06").
• --temperature: The temperature setting for the language model, which controls the randomness of the output (default: 0.0).
• --reset_vectorstore: Whether to reset the vector store after retrieval (default: True).
• --use_langsmith: Whether to enable LangSmith integration for enhanced tracing and analysis (default: False; only necessary if reset_vectorstore == True).
• --langchain_project: The name of the LangChain project to use (default: "citation-finder"; only necessary if reset_vectorstore == True).
• --langchain_tracing_v2: The setting for LangChain tracing version 2 (default: "true"; only necessary if reset_vectorstore == True).
• --langchain_endpoint: The API endpoint for LangChain (default: "https://api.smith.langchain.com"; only necessary if reset_vectorstore == True).
• --langchain_user_agent: The user agent string for LangChain API requests (default: "myagent"; only necessary if reset_vectorstore == True).

License

This project is licensed under the CC BY-NC 4.0 License. See the LICENSE file for details.