AgenticRAG - AI-Powered Agent with Smart Conversation and Retrieval-Augmented Generation

🚀 Overview

AgenticRAG is an advanced AI-powered retrieval-augmented generation (RAG) Agent designed to provide users with an interactive and intelligent conversational experience. Built using LangChain, it leverages an intelligent agent capable of retrieving relevant chunks from a custom AI index Report 2025 based on the user's query. The agent is equipped with memory to handle ongoing conversations and can determine whether to perform a RAG process based on the query’s nature.

The application allows users to interact with the AI agent, either by asking questions or engaging in casual conversation. The agent responds promptly and smartly, while using RAG for information retrieval only when needed, ensuring efficiency.

📒 DeepWiki Explanation

<a href="https://deepwiki.com/MohammedAly22/AgenticRAG" target="_blank"> <img src="https://img.shields.io/badge/Open%20in-DeepWiki-blue?logo=readthedocs&style=for-the-badge" alt="Open in DeepWiki"/> </a>

📜 Table of Contents

• Architecture
• Features
• Installation & Setup
• Usage
• Technologies Used
• Future Enhancements

---

🏗️ Architecture

The system follows a Retrieval-Augmented Generation (RAG) architecture that combines both conversational AI and information retrieval, powered by LangChain. The process involves:

1. Agent Creation: The LangChain agent is set up with the ability to perform multiple tasks: casual conversation or RAG, depending on the query type.
2. Memory & Context: The agent is designed to remember prior interactions, allowing it to engage in context-aware conversations.
3. Query Analysis: When a user submits a query, the agent first analyzes whether it’s a general conversational query or one that requires retrieving detailed data (e.g., "Provide the table of contents of this report"). In addition, there is a query re-formulation part for better retrieval.
4. RAG Execution: If the query demands more specific information, the agent performs RAG to retrieve relevant document chunks from the AI Index Report 2025.
5. Reasoning Steps: The agent can provide detailed reasoning steps for RAG queries, depending on the user's preference. The agent decides whether to show intermediate results or skip to the final answer.

---

Basic RAG Architecture

The RAG system is composed of:

• Memory: Stores prior interactions and updates context.
• Retrieval Tool: Retrieves relevant document chunks from the AI Index Report 2025.
• Generation Tool: Uses LLMs for generating responses, either as final answers or with reasoning steps.

---

✨ Features

✅ Agentic RAG System: The agent intelligently decides whether to perform a RAG process based on the query.
✅ Smart Memory: The agent remembers previous interactions, allowing for context-aware conversations.
✅ Conditional RAG Execution: If the query requires it, the agent performs RAG by retrieving relevant chunks from the AI Index Report 2025.
✅ Reasoning Steps: Users can opt to see the intermediate reasoning steps used by the agent when processing the query.
✅ Natural Conversations: The agent can handle casual conversational queries (e.g., "Hello, how are you?") without performing RAG.
✅ User-Controlled Reasoning: The user can control whether to view the reasoning steps or just the final answer, providing flexibility in how the agent responds.
✅ Streamlit Interface: A user-friendly interface that shows the agent’s responses and reasoning steps interactively.

---

🔧 Installation & Setup

1️⃣ Clone the Repository

git clone https://github.com/MohammedAly22/AgenticRAG.git
cd AgenticRAG

2️⃣ Create and Activate Virtual Environment

python -m venv agentic-rag-env
source agentic-rag-env/bin/activate  # On macOS/Linux
agentic-rag-env\Scripts\activate  # On Windows

3️⃣ Install Dependencies

pip install -r requirements.txt

4️⃣ Run the Application

streamlit run src/app.py

5️⃣ View the Interface

After following the above instructions, you may expect to see this interface:

📖 Usage

1. Open the app in your browser (default: http://localhost:8501).

2. Enter your COHERE_API_KEY in its proper place; both trial and production keys work properly.

3. Select an Embedding Model - Note: The cohere/embed-v4.0 model, when used with a trial_key, is limited to processing 100,000 tokens per minute. This rate limit may cause slower processing for large documents like the AI Index Report 2025 due to enforced waiting between batches. However, despite the slower throughput, it is much more efficient and accurate compared to sentence-transformers/all-mpnet-base-v2, especially for high-quality semantic embeddings.

4. Upload the 2025 AI Index Report in the file uploader area. Once you upload it, it starts processing the PDF, splitting it, creating chunks, and indexing it into the Chroma vector store.

5. Select how many pages you want to render in the UI. Limits the number of previewed pages from the uploaded PDF to improve performance, as rendering more pages takes longer. A maximum of 100 pages can be previewed.

6. Engage in a conversation with the AI agent or ask it to retrieve information from the AI Index Report 2025.

Examples:

• Casual Conversation: If you ask, “Hello, how are you?”, the agent will greet you without performing any RAG.

• Specific Query:
  • If you ask, “Provide me with the complete welcome message from the co-directors of the report”, the agent will perform RAG, retrieve relevant chunks, and generate an appropriate response.

• Here is the same example but with Show Reasoning Steps enabled:

🔧 Technologies Used

• LangChain - For building the intelligent agent with memory and retrieval-augmented generation capabilities.

• Cohere - LLM used for generation and embedding tasks (providing responses).

• Chroma - Vector databases for storing and retrieving document chunks.

• Streamlit - Interactive UI for easy user interaction.

🔮 Future Enhancements

• ✅ Multi-model support for more flexible generation (e.g., OpenAI GPT models).

• ✅ Multi-modal support for chatting with images and tables.

• ✅ Enhanced memory management for long-term, context-aware conversations.

• ✅ Fine-tuned retrieval with advanced filtering and re-ranking techniques.

• ✅ Multi-turn conversations with long-term memory and reasoning enhancements.

💬 Have Questions?

Reach out on GitHub or open an issue!

---

🎯 AgenticRAG - Your Intelligent AI Agent for Smart Conversations and Data Retrieval! 🚀