Vector Knowledge Base: A Semantic Search Engine for Personal Document Management

This paper presents Vector Knowledge Base (VKB), a full-stack semantic search engine designed for personal document management. Traditional keyword search fails to capture conceptual similarity, leading to vocabulary mismatch and retrieval inefficiencies. VKB addresses this by embedding documents into a high-dimensional semantic space and performing cosine similarity-based retrieval.

The system uses the all-mpnet-base-v2 SentenceTransformers model to generate 768-dimensional embeddings, stored in a Qdrant vector database with HNSW indexing for efficient approximate nearest neighbor search. The document pipeline supports 23 file formats via 9 specialized extractors, including OCR for images and AST-aware chunking for source code.

Core features include:

• Sub-50ms semantic search latency for collections under 10,000 vectors

• HDBSCAN clustering with automatic TF-IDF-based semantic cluster naming

• Interactive 3D embedding visualization using PCA and Three.js

• Hierarchical folder management via SQLite with drag-and-drop UI

• Batch upload with GPU-accelerated embedding generation (CUDA / MPS)

• Model Context Protocol (MCP) integration for AI agent connectivity

• Dockerized deployment with FastAPI backend and vanilla JavaScript frontend

The system scales efficiently beyond 100,000 vectors and includes a fault-tolerant document registry for O(1) document listing. VKB demonstrates a practical, production-ready architecture for self-hosted semantic knowledge management.