Hachi: An Image search engine (eagledot.xyz)

Hachi is a small, self‑hosted image search engine project designed to let you search across distributed personal data (local disks, remote servers or devices) via a single interface that fuses traditional metadata with ML‑derived semantic signals. The current implementation indexes images but is explicitly built to add video, text and audio. The author emphasizes privacy and hackability: minimal external dependencies, a lightweight bootstrap, and a roadmap that avoids duplicating original files — the system points to resources rather than moving them. The goal is responsive, iterative search where users can refine queries using multiple exposed attributes rather than relying on a single opaque ranking. Technically, Hachi combines a Nim-written column‑oriented meta‑index with a Python backend; JSON is used for cross‑language IPC and a few optimized Nim/C hotspots speed bottlenecks. Semantic search uses embeddings and top‑k comparison, while deterministic attributes are indexed separately so queries can blend both signals. The meta‑index is single‑threaded today but designed for auxiliary indices, cache‑friendly data structures, multi‑versioning ideas, and future multithreading/scaling. Notable practical choices: only three core Python deps (numpy, regex, markupsafe), optional requests, face‑clustering with immutable ML fields plus editable user labels, and ambitions to run distributed queries on low‑power devices—making Hachi a privacy‑first, extensible testbed for personal semantic search.