netease-youdao/qanything

QAnything is a question and answering system that works on top of your own documents. The name stands for Question and Answer based on Anything. You point it at files you have stored locally, ask questions in plain language, and it gives you answers drawn from the content of those files. It is built by NetEase Youdao and is open source under the Apache license. The README lists the file types it can read, including PDF, Word, PowerPoint, Excel, Markdown, email, plain text, images, CSV, and web links, with more planned. A main selling point is privacy: it can be installed and run with the network cable unplugged, so your documents never leave your machine. It also supports asking questions in Chinese or English no matter which language the document is written in, and it can search across several knowledge bases at once. A technical section explains how it finds relevant text. It uses a two step process: a first pass that gathers candidate passages and a second pass called reranking that reorders them for accuracy. The authors argue this two stage method holds up well as the amount of data grows, where a single step search would get worse. The retrieval relies on their own models, BCEmbedding for the first stage and a matching reranker for the second, and the document includes comparison tables showing how these score against other models. The answering itself is built on the open Qwen language model, fine tuned on many question answering datasets. The rest of the README covers getting started: prerequisites, installation through a pure Python setup or Docker, offline install, an FAQ, and usage with an API. The full README is longer than what was shown.