Retrieval Augmented Fine-Tuning (RAFT)

What Is Retrieval Augmented Fine-Tuning?

For companies running AI systems with their own data, RAFT solves a central problem: pure RAG systems often deliver inaccurate answers when faced with irrelevant or contradictory documents. RAFT trains the model to filter out noise and use only the truly relevant passages. Especially for chatbots in specialized domains like law, medicine, or technology, RAFT significantly improves answer quality.

Retrieval Augmented Fine-Tuning (RAFT) combines two powerful approaches: RAG and fine-tuning. While RAG provides a model with relevant documents at runtime and fine-tuning fundamentally trains the model for new tasks, RAFT combines both methods. The model is specifically trained to extract the correct information from a mix of relevant and irrelevant documents.

The training follows a well-designed concept: the model is presented with question-answer pairs along with a mix of oracle documents (containing the correct answer) and distractor documents (irrelevant documents). This way the model learns not only to find information, but also to ignore distractions. The answers are trained with chain-of-thought reasoning so the model makes its thought process traceable.

For companies building domain-specific AI applications, RAFT offers a clear advantage over pure RAG or pure fine-tuning. The combination delivers more precise answers, especially in specialized knowledge domains like law, medicine, or technical documentation. In combination with agentic RAG, RAFT-trained models can also reliably perform complex multi-step research.