what is auditrag fr?

AuditRAG is a tool for building question-answering systems on top of your own documents, in a way that lets you actually check whether the AI's answer is telling the truth. This kind of system is often called retrieval-augmented generation, or RAG, where an AI model answers questions by first pulling relevant text from your files and then writing a response based on it. The README points out that most RAG tools cite a whole document under a paragraph of generated text, which does not really prove anything, since the model could still be making claims the source never actually said. AuditRAG's approach is to treat exact sourcing as the whole point of the tool. Every sentence the AI writes is linked to a specific chunk of text and a specific page number, and a separate model checks each individual claim against the exact wording of the source it cites, flagging anything unsupported. The whole question-and-answer exchange, including every claim, its verdict, and the exact cited text, can be exported as a timestamped PDF report, the kind of document you could hand to an auditor or compliance reviewer as evidence. Using it involves installing the Python package, pointing it at a folder of PDF, Markdown, or text files to index them, and then either asking questions from the command line or running a small web interface in your browser. It works with a local, no-API-key embedding model by default, but can also be configured to use OpenAI or a local alternative like Ollama for both the search and the answer-generation steps. Under the hood, it combines a search technique called BM25 with vector-based semantic search, stores the retrievable text chunks in a local SQLite database, and keeps embeddings in a separate ChromaDB index, always linked back to the same source chunk identifiers so a citation can never point to text that does not exist. The README notes that this version deliberately does not support multi-turn chat or bring in heavyweight frameworks like LangChain, aiming to keep citation tracking simple and verifiable. The project is open source under the MIT license.