RAG

Retrieval-Augmented Generation (RAG) is a technique that combines a language model with a search or retrieval step, allowing the model to answer questions using specific external content rather than relying solely on what it learned during training. Instead of asking a model to answer purely from memory, a RAG system first searches a knowledge base — in this case, a collection of notes — for the most relevant pieces of content, then feeds those pieces to the model alongside the question, grounding its answer in actual source material.

The typical pipeline involves three stages: indexing (breaking content into chunks and converting each into a numerical representation called an embedding), retrieval (finding the chunks most relevant to a given query by comparing embeddings), and generation (passing the retrieved chunks and the original question to a language model to produce a final answer). This approach helps reduce a common weakness of language models — confidently stating incorrect information — by anchoring responses to real, retrievable source text rather than the model’s general training knowledge alone.

RAG is particularly well suited to personal knowledge bases like an Obsidian vault, where the value lies in a specific, evolving body of notes that no general-purpose model would otherwise know about.

Role in this vault

RAG is the intended bridge between this vault’s notes and Local AI — using a local model via Ollama or LM Studio to embed and index the vault’s own content, so questions can be answered by retrieving directly from these notes rather than from a model’s general knowledge. This is planned as a lighter-weight setup on the Linux machine hosting this vault, following the same principle already used in the broader Fastigiata Homelab.