what is fidx fr?

fidx is a local search engine for your own files, built so that both people and AI agents can quickly find relevant documents without sending anything to the cloud. It indexes markdown files, plain text, chat exports, and code, then lets you search that content using a mix of exact keyword matching and meaning based matching, so a search can find a document even if it does not use the exact same words you typed. Everything runs on your own computer's processor, with no GPU and no API keys required. What makes fidx different from some other local search tools is speed. Many similar tools use an AI language model to rephrase your query or double check the results, which can take around ten seconds per search. fidx skips that step entirely, relying instead on a fast combination of keyword search and a single lightweight embedding calculation per query, which keeps most searches under a tenth of a second once the tool has warmed up. All of the indexed data, including the documents themselves, the keyword index, and the vector data used for meaning based search, lives inside one single SQLite database file that can be copied, backed up, or deleted as a unit. You use fidx from the command line. You register folders as named collections, run an indexing command that scans and processes the files, and then search either across everything or scoped to one collection. There is also a background mode that keeps the search model loaded in memory so that repeated searches run in milliseconds instead of needing to reload each time, which is especially useful when an AI agent is calling it repeatedly. Search results can be returned as JSON, making fidx easy to plug into an agent's toolset as a private retrieval layer. Installing it requires Python 3.11 or 3.12 with a specific SQLite feature enabled, and the project explains exactly how to check and fix this on Linux, Windows, and macOS, since macOS needs a Homebrew-installed Python rather than the default one. The first time you index files, it downloads a small embedding model, after which the tool works fully offline. It is released under the MIT license with an additional AI attribution clause.