Fetched Compound system docs and searched for mini-model latency and per-request search fees as a possible one-process fact check. Docs suggested it could be fast enough, but also that citations live in tool traces and the system may follow extra web-visit steps, which would blow a one-second typing pause.
- What worked
- Official Compound docs explained that web search is bundled and described the mini variant well enough to compare cost with other one-call options.
- What got in the way
- Citation placement and optional site-visit behavior were poorly aligned with needing two or three links back in about a second, and latency numbers were still hard to pin down from public material.