🔬 SecuRAG: Academic Research Agent with Hybrid RAG Pipeline

A production-style RAG Agent for academic literature research, featuring multi-stage retrieval, agentic routing, and automated evaluation.

📖 Introduction

SecuRAG is an intelligent research assistant built on a full agentic architecture. Unlike simple RAG systems, SecuRAG implements a multi-stage retrieval pipeline and an autonomous intent router that decides whether to retrieve from the knowledge base or respond directly.

Key Features

• Agentic Intent Router: Autonomously classifies user queries into SEARCH or CHAT, achieving 85% accuracy on a 20-case test set.
• Hybrid Retrieval Pipeline: Combines BM25 keyword retrieval and dense vector search for higher recall.
• CrossEncoder Reranking: Two-stage retrieval with CrossEncoder re-scoring for improved precision.
• Query Rewriting: LLM-based coreference resolution to handle multi-turn ambiguous queries.
• Multi-turn Memory: Session-based conversation history management.
• LLM-as-a-Judge Eval: Automated evaluation of answer relevance and completeness.
• Local Privacy Mode: Fully offline using Ollama + ChromaDB. Zero data egress.
• Microservices Architecture: Decoupled FastAPI backend + Streamlit frontend, Docker-ready.

---

📊 Performance

Metric	Result
Intent Router Accuracy	85% (17/20 test cases)
Retrieval	BM25 + Vector hybrid, top-5 recall
Reranking	CrossEncoder ms-marco-MiniLM-L-6-v2
Eval	LLM-as-a-Judge (relevance + completeness)

---

🚀 Installation & Setup

Prerequisites

• Python 3.10+
• Ollama installed and running
• Docker (optional)

1. Clone & Install

git clone https://github.com/HITGoose/SecuRAG-Agent.git cd SecuRAG-Agent pip install -r requirements.txt

2. Configure Environment

APP_MODE=local OLLAMA_HOST=http://localhost:11434 DEEPSEEK_API_KEY=your_key

3. Pull Local Model

ollama pull deepseek-r1:1.5b

4. Run

Option A: Docker docker-compose up

Option B: Manual Terminal 1: python api_server.py Terminal 2: streamlit run frontend.py

---

🛠️ Tech Stack

Layer	Technology
LLM	DeepSeek / Ollama
Vector DB	ChromaDB
Retrieval	BM25 + Dense Vector
Reranking	CrossEncoder
Backend	FastAPI
Frontend	Streamlit
Deployment	Docker

---

📁 Project Structure

securag_engine.py - Core RAG Agent engine hybrid_retriever.py - BM25 + Vector hybrid retrieval reranker.py - CrossEncoder reranking security_guard.py - Input sanitization pdf_loader.py - PDF ingestion pipeline api_server.py - FastAPI backend frontend.py - Streamlit UI test_intent.py - Intent router evaluation test_eval.py - LLM-as-a-Judge evaluation Dockerfile docker-compose.yml