code3hr/cyxrag

Open RAG is a tool that helps AI coding assistants answer questions about a codebase without reading the whole thing every time. Instead of dumping entire files into an AI model's context window, it builds a local index of a project and pulls out small, relevant chunks of evidence when a question comes in. The goal is to cut down on wasted tokens, make answers more reliable, and speed up the process of reviewing code or getting familiar with a new repository. Everything runs locally and works offline, so no code or questions need to leave your machine unless you choose to connect an optional local model runtime for generating answers. The basic flow is simple: you set up a small config file pointing at your project, build an index from your source files, then run queries against that index to get back an evidence packet, a set of the most relevant snippets for your question. You can check the quality of that packet before spending any model tokens on it, and optionally send it to a local model server to get a plain answer. The project is organized in phases. Early phases handle building the search index and answering questions using strict evidence based prompts. A later phase adds prebuilt knowledge packs so repeated queries load faster. There is also a contract file meant for integration with a companion tool called Cyxcode, aimed at keeping code edits and reviews grounded in real evidence from the repo rather than guesswork. The README notes that small local AI models, in the 3 billion parameter range, can answer focused questions from the retrieved evidence but sometimes struggle to follow strict formatting or time out on larger packets. Because of this, the author recommends limiting how many evidence chunks are pulled per query and keeping each chunk short when speed matters, and suggests using a stronger local model for tasks that need carefully structured answers. A runnable example project is included so newcomers can try the retrieval and query flow before applying it to their own code.