Retrieval-Augmented Fine-Tuning (RAFT)
A training technique that combines retrieval-augmented generation with fine-tuning, teaching models to better leverage retrieved context. Produces models that are both knowledgeable and grounded in source documents.
Why it matters
RAG lets a model pull in outside documents at answer time, but models aren't always good at actually using what's retrieved, and sometimes ignore it. RAFT trains the model specifically to read retrieved passages well and lean on them instead of guessing. For teams building AI over their own knowledge base, this means more answers grounded in the real documents and fewer confident-sounding mistakes.
A concrete example
A company builds an internal assistant over its policy manuals. With plain RAG, the model sometimes retrieves the right page but still answers from memory. After RAFT, it's been trained on examples where the correct answer comes straight from the retrieved text, and also to ignore irrelevant passages. Now when an employee asks about a leave policy, the reply reliably reflects what the actual manual says.
How to use it
Worth considering when you already run retrieval and the model handles the retrieved passages poorly — ignoring them, mixing them with its own priors, or failing to notice when they do not answer the question. Training it specifically to work with retrieved context, including distracting passages, addresses that. It is a refinement of a working retrieval system, not a starting point.
The common mistake
Reaching for it before the retrieval itself is good. If the wrong passages are being retrieved, no amount of training on how to use passages helps — fix retrieval first and re-measure.
Related terms
RAG (Retrieval-Augmented Generation)
A technique where the AI retrieves relevant information from a knowledge base before generating a response, reducing hallucinations and grounding outputs in real data.
Fine-Tuning
The process of further training a pre-trained AI model on a specific dataset to improve performance for a particular task or domain.
Grounding
The process of connecting AI model outputs to verified, real-world information sources. Grounded AI responses cite specific documents, databases, or web sources — reducing hallucinations and increasing factual reliability.
AI Dubbing
Automated translation and re-voicing of audio/video content into other languages while preserving the original speaker's voice characteristics, timing, and emotional delivery.
AI Orchestration
Coordinating multiple AI models, tools, and data sources in a unified pipeline. An orchestration layer manages prompt routing, context passing, error handling, and output aggregation across different AI services.
API (Application Programming Interface)
A way for software applications to communicate with each other. AI APIs let developers integrate AI capabilities into their own applications programmatically.