scientific_agent

Scientific Agent: A Retrieval-Augmented Generation (RAG) System for Domain-Aware Literature Review Automation

---

Overview

scientific_agent is an AI-powered literature review system that combines the power of Retrieval-Augmented Generation (RAG) with structured metadata from scientific databases. It enables the generation of accurate, context-rich, and citation-ready literature reviews across various research domains including—but not limited to—biomedicine, computer science, engineering, social sciences, and environmental studies.

The system leverages modern language models, real-time knowledge retrieval, and multi-source citation formatting to support scalable, credible, and multilingual scholarly writing.

---

Key Features

• 🔍 Domain-Agnostic RAG Pipeline Capable of generating literature reviews on any topic where academic publications are available—across disciplines.

• 📚 Multi-Source Retrieval Retrieves metadata and abstracts from Crossref, with optional support for Scopus and Web of Science, ensuring comprehensive coverage.

• 💡 Knowledge-Grounded Generation Uses OpenAI embeddings and FAISS vector search to ground language model outputs in retrieved, relevant documents.

• 🌐 REST API Access Offers a FastAPI-based endpoint for integrating RAG-based review generation into research tools or academic workflows.

• 🧾 Flexible Citation Formats Outputs citations in BibTeX, APA 7, or raw key/DOI formats based on user preference.

• 🌍 Multilingual Support Literature reviews can be generated in multiple languages by specifying the target language.

---

Use Cases

• 📖 Academic Literature Reviews (e.g., AI ethics, climate change, neural network architectures)
• 🧬 Biomedical Research Summaries
• ⚖️ Legal and Policy Analysis with Reference Tracking
• 📊 Technical Surveys in Engineering and Data Science
• 🧠 Scientific Knowledge Gap Identification

---

Installation & Setup

pip install -r requirements.txt

Create a .env file with your API credentials:

OPENAI_API_KEY=your_openai_key
CROSSREF_MAILTO=[email protected]
# Optional:
ELSEVIER_API_KEY=...
WOS_API_KEY=...

---

Usage

CLI Mode

python scientific_rag_agent.py

You will be prompted for a topic. The resulting literature_review.md will be created in the current directory.

API Mode

Start the FastAPI server:

uvicorn main:app --host 0.0.0.0 --port 7001

Then send a POST request to:

POST /literature-review

Example Request

{
  "topic": "The impact of generative AI on educational assessment",
  "citation_format": "apa7",
  "language": "English"
}

---

System Architecture

• Retrievers: Modular design for metadata extraction from Crossref, Scopus, and Web of Science.
• Embedding + Retrieval: Uses OpenAI embeddings and FAISS for nearest-neighbor search.
• LLM Backend: gpt-4o and gpt-4o-mini models for natural language generation.
• Citation Formatter: Styles include raw, BibTeX, and APA 7.
• Prompt Templates: Designed for domain-aware, citation-rich synthesis.

---

Limitations

• Retrieval Dependency: Quality depends on the abstracts and metadata available from external APIs.
• Citation Bias: May inherit biases from underlying datasets or retrieval algorithms.
• Factual Constraints: Only generates from retrieved content; does not fabricate knowledge.

---

Future Work

• 📑 PDF ingestion and multimodal source support
• 📌 Full-text summarization (beyond abstracts)
• 🧠 Continual learning for evolving knowledge bases
• 🧮 Improved metrics for coherence, factual grounding, and citation coverage

---

License

Licensed under the MIT License © 2025 Emirhan Soylu

---

References

Literature used for model synthesis and benchmarking is cited in BibTeX format in the output files. For a conceptual background, see:

@article{Genesis2025,
  title={Large Language Models (LLMs)...},
  author={Genesis},
  year={2025},
  doi={10.20944/preprints202504.0443.v1}
}
@article{Li2024,
  title={Biomedrag: A Retrieval Augmented Large Language Model for Biomedicine},
  author={Li},
  year={2024},
  doi={10.2139/ssrn.4910081}
}
@article{Huang2024,
  title={Adapting LLMs for Biomedicine through Retrieval-Augmented Generation},
  author={Huang},
  year={2024},
  doi={10.1109/bibm62325.2024.10822725}
}
@article{Singh2024,
  title={A Multimodal Framework for Quantifying RAG Efficacy},
  author={Singh},
  year={2024},
  doi={10.36227/techrxiv.173152556.61823435/v1}
}

---

Contributing

Pull requests and feature suggestions are welcome, particularly around new retriever integrations, multilingual capabilities, or review evaluation methods.