In the basic RAG module, retrieval was unconditional. Every query triggered the same pipeline: embed the query, search the vector store, stuff the top chunks into the prompt, generate. The model had n
Markdown supported · ⌘↵ to post
No responses yet.