Definition
Retrieval-Augmented Generation (RAG) combines a retrieval step (searching specific source documents such as OEM manuals or maintenance history) with generative AI's content-production capability, so the generated answer is grounded in and can cite that retrieved, verified source material, rather than relying purely on patterns the model learned during training.
Key facts
- Combines two distinct steps: retrieval (finding relevant source content) and generation (producing an answer from it).
- Aims to ground generated answers in specific, verifiable source documents, reducing the risk of plausible-but-incorrect generated content.
- Relies on a vector database and embeddings as the underlying technology for the retrieval step, matching a query to relevant source passages.
- Particularly valuable in maintenance for answering questions against OEM manuals, procedures, or maintenance history with traceable sourcing.
How does Retrieval-Augmented Generation work in maintenance?
RAG works by first retrieving the most relevant passages from a set of source documents (using embeddings and a vector database to match a question to relevant content) then feeding those retrieved passages to a generative AI model, which produces an answer grounded in and often citing that specific retrieved material.
What does Retrieval-Augmented Generation require?
Embeddings represent text as vectors capturing semantic meaning, and a vector database stores and searches those vectors efficiently. RAG relies on both as its retrieval mechanism: converting a query and source documents into embeddings, searching the vector database for the closest matches, and passing those matched passages to the generative step.
FAQ
Frequently asked questions
How does RAG connect retrieved information with a generated answer?
RAG connects them by first retrieving relevant passages from specific source documents using embeddings and a vector database, then feeding that retrieved content to a generative AI model to produce a grounded, source-traceable answer.
Last reviewed: