Thanks Jason, you're right. We spend so much time around RAG that we forget this is niche term.
RAG -> retrieval augmented generation. Given a query, finding the parts of the codebase most relevant to the query, and supplying it to the LLM so it can answer the query. The typical way to do this is semantic similarity, or chunking + embedding the corpus being searched, embedding the query and finding the k most cosine-similar chunks.
RAG vs fine-tuning makes a difference to you, but not to your users, who don't need to understand how they're not the same thing in order to use your product.
You’ll want to find a balance between telling people what your thing does with no jargon or acronyms whatsoever, while signaling to people who know as much as you about this stuff (all four of them) that it’s got those cool ingredients.
A simpler way to put it is it's a technique to get around the context limit/attention issues of an LLM... to inject the context that is most relevant to the user's query.
RAG -> retrieval augmented generation. Given a query, finding the parts of the codebase most relevant to the query, and supplying it to the LLM so it can answer the query. The typical way to do this is semantic similarity, or chunking + embedding the corpus being searched, embedding the query and finding the k most cosine-similar chunks.