Fidx – local semantic search in one SQLite file, no LLM at query (github.com)

A new local AI search engine, Fidx, has been launched, aimed at providing fast semantic and keyword search capabilities for various file types, including markdown, text, and code, all operated without cloud reliance or GPU support. The entire index is compactly stored in a single SQLite file, enabling users to conduct searches using a command-line interface (CLI) with responses generated in milliseconds. Unlike many existing solutions that rely on large language models (LLMs) for processing queries, Fidx employs a hybrid search approach that combines BM25 for exact match retrieval and 768-dimensional vector searches, resulting in impressive performance with query latencies ranging from 18 to 49 milliseconds in optimal conditions. This development is significant for the AI/ML community as it challenges the performance and operational models of traditional LLM-based search systems, which can take significantly longer to process queries due to their complexity. By demonstrating lower latency and a strong recall performance across various benchmarks, Fidx presents an attractive option for users and applications that require quick, locally-managed searches without the complications of remote data handling. With its easy-to-backup SQLite file structure, scoped search capabilities for organized collections, and provisions for agent-friendly responses, Fidx is positioned as a practical tool for developers, data analysts, and knowledge workers seeking efficient local search solutions.