ask-dual-mode-makes-three-llm-calls

Summary

When the system runs in dual mode, it can make up to three separate calls to a language model: one to synthesize from the truth maintenance network, one to do keyword-based retrieval, and one to merge those results together. If either the network synthesis or the keyword search comes back empty, it skips the merge step and only makes two calls. This matters for cost and latency — every LLM call adds time and token spend, so the short-circuit behavior keeps simpler queries cheaper.