# RAGless

A retrieval-only question-answering system for documentation that makes zero LLM calls at query time, matching questions against pre-extracted answer blocks using local vector search instead of generation.

RAGless answers questions about documentation with no LLM call at runtime. Source documents are converted ahead of time into self-contained question-and-answer blocks, using Gemini in JSON mode only during preparation, indexed into a local Qdrant vector database, and queried at runtime using asymmetric Gemini embeddings, so answering a question is purely a vector lookup with no generation cost or hallucination risk.

It aggregates scores across multiple question variants that map to the same answer, which the project reports makes retrieval more stable than classic top-1 matching, chunks only documents that exceed a token threshold measured with the model's real tokenizer, and logs below-threshold queries for later analysis. An optional second LLM pass, enabled with a `--judge` flag, can discard extracted blocks that are not supported by the source text.

It consists of three independent scripts: one to extract Q&A blocks from source documents, one to generate embeddings and populate the local Qdrant database, and a CLI chatbot that retrieves the most relevant stored answer for a user's query.

Website: https://github.com/EmilResearch/RAGless