Small knowledge bases are injected directly into the LLM context to avoid added latency; RAG is only enabled once the KB is large enough that retrieval is more efficient than direct inclusion. (elevenlabs.io)
When RAG is enabled, the system retrieves only relevant chunks and is engineered to keep per‑query overhead low and stable even on large enterprise KBs (median RAG latency ~155 ms after their redesign). In practice, overall KB size is not the primary driver of latency. (elevenlabs.io)
What does influence latency are RAG configuration choices (e.g., embedding model, maximum retrieved chunks, vector distance), which the docs note can impact latency and cost. (elevenlabs.io)
earnest mirage
wraith gale