What is vic_aisaq_demo?

This project is a demonstration of a smart search system built to work well even on devices with very little memory. It is designed for situations where you want to search through a large pile of files, like documents on a hard drive, and find the ones that actually answer a question, without needing to load everything into RAM first. The system works in three stages, called tiers. First, it filters files using simple details like file type, size, or date, to quickly shrink a huge folder down to a much smaller set of candidates. Second, it uses a technique called vector search, which compares the meaning of text rather than just matching keywords, to find files that are semantically related to the question. This step relies on a storage-friendly search method from KIOXIA called AiSAQ, built to work directly off disk instead of needing everything loaded into memory. Third, only the small number of files that survive the first two rounds get read in full detail, so the expensive part of the process only happens on a short list. The project combines two existing open-source pieces: a retrieval controller called lc0_vic that plans and coordinates the tiered search, and aisaq-diskann, which handles the actual disk-based similarity search. It also uses Ollama, a tool for running AI language models locally, both to turn a plain-English question into a search plan and to convert text into the vectors used for semantic matching. Setting the project up involves compiling the AiSAQ backend from source inside a WSL Ubuntu environment, setting up a Python virtual environment, and running specific local models through Ollama for planning and embeddings. The README frames the project as a systems demonstration rather than a finished product, showing the tradeoff of using less memory in exchange for somewhat higher search latency. The full README is longer than what was shown.