NeuraX — Vector Intelligence Platform

NeuraX is a custom-engineered, in-memory Vector Database and Retrieval-Augmented Generation (RAG) pipeline designed for lightning-fast semantic retrieval and local intelligence. Optimized for macOS and powered by a Python FastAPI backend and a premium, responsive glassmorphic dark UI, NeuraX implements high-performance search algorithms and PCA-based multidimensional vector visualization side-by-side.

---

⚡ Core Architecture

NeuraX orchestrates text processing, high-dimensional indexing, and AI generation into a seamless, unified local pipeline:

flowchart TD
    A[User Query] --> B[Ollama nomic-embed-text]
    B -->|768D Vector| C{Nearest Neighbors Search}
    C -->|HNSW Graph ANN| D[Top K Document Chunks]
    C -->|KD-Tree Exact| D
    C -->|Brute Force Exact| D
    D -->|Context Retrieval| E[LLM Prompt Builder]
    A --> E
    E --> F[Ollama llama3.2 LLM]
    F --> G[Streamed Response Output]
    
    style A fill:#050508,stroke:#00e676,stroke-width:2px,color:#fff
    style B fill:#0e0e13,stroke:#1d1d24,stroke-width:1px,color:#fff
    style C fill:#00e676,stroke:#00e676,stroke-width:2px,color:#000
    style D fill:#0e0e13,stroke:#1d1d24,stroke-width:1px,color:#fff
    style E fill:#0e0e13,stroke:#1d1d24,stroke-width:1px,color:#fff
    style F fill:#0e0e13,stroke:#1d1d24,stroke-width:1px,color:#fff
    style G fill:#050508,stroke:#ffc400,stroke-width:2px,color:#fff

Loading

---

🚀 Key Features

1. Multi-Algorithmic Vector Search

NeuraX features three core search algorithms running side-by-side so you can evaluate performance and accuracy tradeoffs in real-time:

• HNSW (Hierarchical Navigable Small World): A production-grade approximate nearest neighbor (ANN) search graph algorithm. By constructing a multi-layer graph index, it achieves logarithmic $O(\log N)$ search complexity, making it scale effortlessly to high dimensions.
• KD-Tree (K-Dimensional Tree): A spatial partitioning tree structured to divide high-dimensional coordinates. It provides exact nearest neighbor searches but degrades in speed as dimensions grow due to the curse of dimensionality.
• Brute Force (Flat Index): The exact search baseline that calculates distance metrics across every node in the database. Operates at linear $O(N)$ complexity.

flowchart TD
    subgraph HNSW Graph Structure
    L2[Layer 2: Sparse Expressway] -->|Zoom in neighborhood| L1[Layer 1: Transitway]
    L1 -->|Greedy local search| L0[Layer 0: Dense All Nodes]
    end
    
    style L2 fill:#ff3d00,stroke:#ff3d00,color:#fff
    style L1 fill:#ff9100,stroke:#ff9100,color:#fff
    style L0 fill:#00e676,stroke:#00e676,color:#000

Loading

2. Multi-Metric Distance Support

Supports three standard mathematical formulas to calculate similarity or distance vectors:

• Cosine Similarity: Measures the angular cosine between vectors (perfect for text and semantic comparisons).
• Euclidean Distance: Calculates the straight-line distance in Euclidean space.
• Manhattan Distance: Calculates distance along axis-aligned grid paths.

3. High-Tech 2D PCA Visualizer

• Displays 16-dimensional semantic demo vectors projected down to a 2D coordinate space using Principal Component Analysis (PCA).
• Renders real-time connections, crawling neon pulses from the target to its K-nearest neighbors, and concentric sonar radar waves to illustrate the search sweep.
• Displays a dedicated Category Legend in the side panel showing color-coded vectors mapped across CS/Algorithms, Mathematics, Food & Cooking, Sports & Games, and Ingested Documents.

4. Semantic Document Ingestion

• Automatically splits long documents into overlapping 250-word chunks.
• Connects to local Ollama endpoints to generate 768-dimensional embeddings using nomic-embed-text.
• Stores document chunks in an independent, high-performance HNSW index.

5. Local AI Core (RAG Pipeline)

• Interrogates your document database with user-defined Context Depth ($K$) parameters.
• Retrieves the top $K$ context chunks, packages them into a clean system prompt context, and streams the answer using local LLMs (e.g. llama3.2 or llama3.2:1b).
• Formats streamed outputs with clean code blocks, custom scrollbars, and user-friendly chat bubbles.

6. Neural Analytics & Benchmarking

• Displays search latency metrics down to the microsecond level.
• Features a real-time Benchmarks visualizer showing execution times for each algorithm, highlights the FASTEST algorithm with a neon badge, and calculates the exact speedup ratio.
• Visualizes active node levels and layer topologies for HNSW indexing graphs.

---

🛠 Tech Stack

• Frontend: Vanilla CSS (Premium Glassmorphic Dark UI), HTML5 Canvas (Visualizer), JavaScript (DOM, Fetch API).
• Backend: Python 3.14+ (FastAPI, Uvicorn, NumPy).
• AI Engine: Ollama Local API (nomic-embed-text and llama3.2).