What is vector_database?

This project is a vector database and semantic search engine built from scratch in C++, paired with a Python backend and a React frontend. A vector database stores pieces of text as long lists of numbers, called embeddings, that capture their meaning, so you can search by similarity rather than exact keyword matches. The core search engine here uses a graph based technique called HNSW, which organizes vectors into layers so it can find close matches quickly without comparing a query against every single item. The author focused heavily on low level performance. The engine stores its data in flat, contiguous arrays instead of the usual nested data structures, aligned to 64 bytes in memory so the CPU can read them efficiently. Distance calculations, the math used to measure how similar two vectors are, are hand written using AVX2 instructions that let the processor handle eight numbers at once instead of one at a time. On top of the C++ engine, text is turned into vectors using Facebook's pretrained FastText model, which can handle misspelled or unfamiliar words because it works with word pieces rather than whole words only. The C++ engine is exposed to Python through pybind11, a tool for connecting C++ and Python code directly, and a FastAPI server handles incoming search requests, also filtering out near duplicate results using a text similarity measure called Levenshtein distance. The project reports benchmark numbers on the standard SIFT1M dataset of one million vectors: about 99.3 percent accuracy at returning the correct top 10 matches, over 2,200 queries handled per second, and typical response times under one millisecond. The backend is packaged with Docker and deployed on Hugging Face Spaces, while the frontend is hosted on Vercel. To run it locally you compile the C++ engine with CMake, start the FastAPI backend after downloading the FastText model file, and run the React frontend separately with npm.