ResearchAssistant

An AI-powered assistant for querying your personal Zotero library and synthesizing literature across sources like PubMed using GPT-4.

This tool lets you interrogate your own research collection using natural language, and combines it with up-to-date PubMed searches to generate integrated answers, identify gaps, and recommend papers to add.

---

Features

• ✅ Query your Zotero library using natural language
• 🔍 Automatically search PubMed for new or related findings
• 🧠 Synthesizes both sources with GPT-4 (or GPT-4o)
• 🗃️ Local vector search with FAISS for your own library
• 🧾 Logs and stores queries and results in a local SQLite database
• 📬 Watch PubMed for new papers on past queries and send email alerts
• 📄 Prompts factored into editable Markdown files in /prompts

---

Requirements

• Python 3.10+
• A Zotero library (.bib file)
• Conda (Miniconda or Anaconda)
• An OpenAI API key with access to:
  • text-embedding-3-large
  • gpt-4 or gpt-4o

---

Installation

1. Clone the repository

git clone https://github.com/BridgesLab/ResearchAssistant.git
cd ResearchAssistant

2. Export your Zotero library

You can either:

Export manually from Zotero: File → Export Library → Format: BibTeX Save the file as library.bib in the project root.

Or create a symbolic link to your .bib file:

ln -s /path/to/your/library.bib ./library.bib

3. Set up the Environment

ln -s /path/to/your/library.bib ./library.bib

4. Add your OpenAI API key and email settings

Create a .env file in the root folder containing at minimum:

OPENAI_API_KEY=your-api-key-here

# For email alerts (adjust according to your SMTP provider)
EMAIL_USER=your.email@example.com
EMAIL_PASS=your-email-password-or-app-token
EMAIL_SMTP=smtp.example.com
EMAIL_PORT=587
EMAIL_TO=your.email@example.com

5. Build the Zotero index

Run this once after setup.

python scripts/build_index.py

This creates:

• zotero.index — FAISS vector index of embedded papers
• zotero_meta.pkl — metadata used for retrieval and synthesis

To update your index run this (or set up a cron job to do it regularly)

python scripts/update_index.py

Running a Query (Multi-Agent Mode)

Use the manager agent to query both Zotero and PubMed, and get a synthesized answer:

python scripts/manager_agent.py "What is the relationship between calcium and cholesterol?"

• Log your query to a local SQLite database
• Search your Zotero library using semantic similarity
• Use GPT-4 to create a PubMed search string and query PubMed
• Synthesize results and highlight any gaps or missing references
• Store results and synthesis in the database for future reference

⏰ Keeping Updated with PubMed Watcher

Periodically run the watcher script to search PubMed for new results related to your past queries and receive email alerts:

python scripts/watch_pubmed.py

Or do do this manually checking recent papers run this to see if there are any new publications relevant to your logged queries (in this case in the last 60 days).

python scripts/find_new_papers.py --days 60

Automate this with cron or task scheduler to run weekly or biweekly.

🌐 Optional: Streamlit UI

Launch a simple browser interface for querying your Zotero library:

streamlit run scripts/app.py

Open the provided local address in your browser.

🧾 Prompt Customization

All GPT prompts are stored in Markdown format under /prompts. These are easy to edit and version, and follow a structured format with role, task, input, output, and guardrails. Current prompts include:

• pubmed_search.md — converts questions to PubMed Boolean searches
• synthesis.md — integrates and summarizes findings
• zotero_search.md — (optional/future) for explainable Zotero querying

Coming Soon / Possible Extensions

• 🔁 Auto-add recommended PubMed papers to Zotero
• 📄 Export summaries or citations to BibTeX, Markdown, or Notion
• 📊 Interactive filters for Zotero results
• 🧪 Jupyter or VS Code extension for integrated research