kurakula-Prashanth2004/neural-search-ai

Neural Search Engine for Technical Q&A 🧠🔍🌌

A production-grade Semantic & Hybrid Search Engine that bridges the gap between keyword retrieval and AI reasoning. This system uses a state-of-the-art Hybrid Retrieval Architecture (BM25 + FAISS) with an intelligent Senior Architect Fallback (LLM) for 100% technical coverage.

🔗 Live Production Links

• —Production Deployment (UI): https://neural-search-engine.vercel.app/
• —Direct API Endpoint (Backend): https://kurakula-prashanth2004-neural-search-ai.hf.space

🎬 Cinematic Project Overview

1. Neural Boot Sequence

When the application launches, it undergoes a tactical boot-up sequence, initializing AI protocols and establishing the retrieval matrix.

<p align="center"> <img src="https://github.com/user-attachments/assets/1ed52591-ea85-47ac-9dbd-c81f759d84d3" width="700"/> </p>

2. Intelligent Search Dashboard

A minimalist, high-impact interface where users can query the neural matrix across three distinct retrieval tiers.

<p align="center"> <img src="https://github.com/user-attachments/assets/cd1a6a9d-de68-4e03-87c3-0d48a4da30b7" width="700"/> </p> ---

🚀 How it Works

🎨 Frontend (React & Vite)

The frontend is the "Command Center" of the engine. It is hosted on Vercel for ultra-fast edge delivery.

• —It handles the Neural Boot Sequence (cinematic loading).
• —It manages user queries and communicates with the AI backend via Axios.
• —It visualizes search results with real-time scoring and markdown rendering for technical code snippets.

🧠 Backend (FastAPI & Hugging Face)

The backend is the "Engine Room," hosted on Hugging Face Spaces for massive 16GB RAM AI inference.

• —Data Layer: It processes a technical Q&A dataset (techqa.csv) upon startup.
• —Indexing Layer: It builds two simultaneous indexes: BM25 (for keyword frequency) and FAISS (for vector-based semantic meaning).
• —Inference Layer: It uses the paraphrase-MiniLM-L3-v2 transformer to convert questions into 384-dimensional mathematical vectors.
• —Decision Layer: If a search score is below 90%, it automatically triggers the Qwen-2.5-7B AI Fallback to generate a comprehensive solution from scratch.

📦 Getting Started

1. Cloning the Repository

To get started with local development, clone the repository using Git:

git clone https://github.com/kurakula-prashanth/neural-search-engine.git
cd neural-search-engine

2. How to Use the App

1. 1.Enter a Query: Type any technical question (e.g., "Explain Random Forest" or "Fix Network Error").
2. 2.Execute: Press Enter or click Execute to trigger the neural matrix.
3. 3.Explore Tabs:
4. 4.BM25: Best for finding exact error codes or keyword matches.
5. 5.Semantic: Best for finding conceptually similar questions.
6. 6.Hybrid: The "Smartest" mode—uses both methods and triggers the AI LLM if confidence is low.
7. 7.Compare All: A side-by-side benchmark of all three AI methods.

🛠️ Local Installation

Backend Setup

cd backend
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
# Set your HF_API_TOKEN in .env
uvicorn app.main:app --host 0.0.0.0 --port 8000

Frontend Setup

cd frontend
npm install
npm run dev

Created as part of an Advanced AI & Cloud Computing project for Technical Information Retrieval.