emilresearch/ragless

RAGless is a question-answering system for your own documents that works without calling an AI model at query time. Most documentation chatbots send each user question to a language model at runtime, which costs money, introduces lag, and sometimes produces incorrect answers. RAGless takes a different approach: an AI model is used only once during setup to convert your documents into pre-written question-and-answer pairs, and all subsequent user queries are handled by fast local vector search with no API calls. The workflow has three steps. First, you run a preparation script that reads your PDF, text, or Markdown files and uses the Gemini API to extract structured question-and-answer blocks from them. Each block contains the answer text, several ways a user might phrase the question, and a source quote. Second, an ingestion script turns those blocks into vector embeddings and loads them into a local Qdrant database stored on your disk. Third, users interact with a command-line chatbot that compares their question to the stored question vectors, sums up scores across multiple phrasing variants of the same answer, and returns the best matching pre-written answer verbatim. Because answers are fixed at ingestion time, the system cannot hallucinate at runtime. It also cannot generate novel answers or combine information from multiple sources into a single synthesized response. It works best for support documentation, FAQ content, or policy documents where there is a known, finite set of correct answers. The README includes a detailed comparison table showing where this approach beats classic retrieval-augmented generation and where it falls short. Setup requires Python 3.10 or newer and a free Gemini API key from Google. There is no Docker requirement: Qdrant runs embedded directly on your machine. Installing the Python dependencies, configuring the API key in a .env file, placing your documents in the source folder, and running the three scripts in sequence is the full setup process. The tool is aimed at developers who need a reliable, low-cost Q&A layer over a stable knowledge base and want deterministic behavior rather than generated responses.