Local Semantic Search Engine with Preloaded Models
The author has developed a lightweight, local semantic search engine that lives in the system tray, with preloaded models that automatically sync to changes, allowing for fast, accurate searches without load times.
Why it matters
This local semantic search engine provides a fast, accurate, and seamless way to search through personal files and data, without the need for internet connectivity or cloud-based services.
Key Points
- 1Lightweight, local semantic search engine
- 2Preloaded models that automatically sync to changes
- 3Supports multiple file types and modalities (text, image, OCR)
- 4Hybrid lexical/semantic search algorithm with MMR reranking
Details
The author has created a highly optimized local search engine that can embed a 20,000-file database in under an hour with 6x multithreading on GPU. It uses a hybrid lexical/semantic search algorithm with MMR reranking to provide highly accurate results, which are further boosted using an LLM that gives quality scores. The engine supports up to 49 file extensions, including vision-enabled LLMs, text and image embedding models, and OCR. It also has an optional 'Windows Recall'-like feature that takes screenshots and saves them to a folder, which can be searched separately. The author has not yet implemented the full RAG (Retrieval-Augmented Generation) functionality, as they find the LLM response to be too time-consuming, but the core retrieval part works well and can be accessed quickly from the system tray.
No comments yet
Be the first to comment