RAGless: A FAQ Search System Based on Question-to-Question Matching

RAGless is a semantic search system based on Question-to-Question matching that removes the generation step of traditional RAG, providing a structure optimized for FAQ systems.

How It Works

• Ingestion: Using an LLM, 3 to 5 question variants are generated per answer, and embeddings are extracted for each.
• Query time: After embedding the user's question, the most similar question variants (Top-K) are retrieved. Scores are then aggregated by answer_id, and the answer with the highest score is returned.
• Threshold logic: A dual-gate approach combining a minimum aggregate score (0.70) and a single-hit top score (0.82) is used to prevent False Negatives.

Key Features

• RAG vs RAGless: While traditional RAG retrieves document chunks and has an LLM generate an answer, RAGless retrieves pre-generated question variants and immediately returns a prepared answer.
• Improved Precision: Because it operates at the question level rather than the document level, it shows high precision in closed-domain FAQ environments with a fixed answer scope.
• Target Audience: Suitable for engineers and researchers building FAQ systems with a fixed answer space, rather than generative Q&A.