Research AgentX 🤖

A conversational, multi-turn, and stateful research agent powered by LangGraph. It follows a deterministic workflow to search multiple data sources, intelligently refines its queries based on conversational context, and synthesizes its findings into a coherent answer. All conversation history is persistent, powered by a SQLite database backend.

---

🖼️ Application UI

The user interface supports multiple, persistent conversations. Users can switch between chats, create new ones, and see the agent work through its research process in real-time.

---

✨ Overview

Research AgentX has evolved into a sophisticated conversational research assistant. It maintains the context of a conversation across multiple turns, allowing for follow-up questions and iterative exploration of topics.

Given a user's query, it executes a structured, deterministic plan to gather information from:

• General Web Search: For up-to-date information using Tavily Search.
• Academic Papers: For scientific knowledge from ArXiv.
• Internal Knowledge Base: For specialized information using a Retrieval-Augmented Generation (RAG) pipeline with a local ChromaDB instance.

The agent uses a stateful graph built with LangGraph and a SQLite checkpointer to manage its workflow and persist all conversation history, ensuring no data is lost between sessions.

---

⚙️ How It Works

The agent's logic has been refactored from an LLM-based planner to a more robust and predictable deterministic workflow.

1. Query Refinement: Every user query first enters the refine_query node. This node uses an LLM with a highly detailed prompt to analyze the full conversation history. It rewrites the user's latest message into a complete, standalone search query, resolving context and handling follow-up questions.
2. Sequential Search & Grading: The agent proceeds through its tools in a fixed order: web_search, then arxiv_search, then rag_search.
   • After each search, the results are passed to the grade_and_filter node, which uses an LLM to discard irrelevant documents.
3. Conditional Routing: The route_after_grading function checks if the last search found any relevant documents.
   • If YES, it proceeds to the next tool in the sequence (e.g., from web_search to arxiv_search).
   • If NO, it loops back to the refine_query node to try a different query for the same tool. It will retry up to a defined limit before moving on.
4. Synthesis: Once all search tools have been tried, the graph proceeds to the synthesize node. This node takes all the relevant documents collected from all sources and generates a final, comprehensive answer based on the user's last question.
5. Persistence: The entire state of the conversation, including every message and intermediate step, is automatically saved to a db.sqlite file using LangGraph's SqliteSaver.

---

🧠 Graph Architecture

The new architecture is a deterministic, sequential graph that prioritizes reliability and predictability. It uses conditional edges to loop for query refinement when a search tool fails to find useful information. This design removes the "LLM-in-the-loop" for planning, resulting in a more stable and debuggable agent.

---

🚀 Key Features

• Conversational Memory: Maintains the full context of a conversation, allowing for natural follow-up questions.
• Persistent Chat History: Uses a SQLite database to save all conversations, so you never lose your work.
• Multi-Conversation Management: A sidebar UI allows you to create, delete, and switch between different chat sessions.
• Automatic Conversation Titling: Automatically generates a concise title for new chats based on your first query.
• Intelligent Query Refinement: A dedicated node analyzes user intent and conversational history to create highly effective search queries.
• Multi-Tool Orchestration: Sequentially queries the web, ArXiv, and an internal RAG pipeline.
• Relevance-Based Grading: Employs an LLM-powered grader to filter out irrelevant noise from search results.

---

🛠️ Tech Stack

• Orchestration: LangChain & LangGraph
• LLM: Google Gemini (gemini-2.0-flash)
• Database: SQLite (for conversation state) & ChromaDB (for RAG)
• Web Search: Tavily Search API
• Academic Search: ArXiv Python Library
• Embeddings: Hugging Face Sentence Transformers (all-MiniLM-L6-v2)
• UI Framework: Streamlit

---

📦 Setup & Installation

1. Clone the Repository

   git clone https://github.com/eslammohamedtolba/Research-AgentX.git
   cd Research-AgentX

2. Create and Activate a Virtual Environment

   # Create the environment
   python -m venv venv
   
   # Activate on Windows PowerShell
   .\venv\Scripts\Activate.ps1
   
   # On macOS/Linux, use: source venv/bin/activate

3. Install Dependencies

   pip install -r requirements.txt

4. Set Up Environment Variables

   • Create a file named .env in the root directory.
   • Add your API keys to the .env file:

   GOOGLE_API_KEY="YOUR_GOOGLE_API_KEY"
   TAVILY_API_KEY="YOUR_TAVILY_API_KEY"

---

▶️ How to Run

1. Launch the Application

   streamlit run app.py

2. First-Time Setup: The first time you run the application, it will download the ML paper dataset from Hugging Face and build the ChromaDB vector store. This may take a few minutes. Subsequent runs will be much faster as it will load the existing database.

3. Start Chatting: Open your browser to the Streamlit URL. A new chat will be created automatically. Type your research question and press Enter.

---

🤝 Contributing

Contributions are welcome and greatly appreciated! Whether it's fixing a bug, proposing a new feature, or improving documentation, your help makes this project better.