wtf is activegraph-parallel-prefetch?

This project is a demo showing how to speed up a system called ActiveGraph, which processes events one at a time in a strict order so its history stays predictable and reproducible. That strict one at a time processing becomes slow when a single event needs to make many separate calls to an AI language model, since each call has to wait for the previous one to finish. In the example used here, one event triggers 200 separate AI calls to label parts of a graph, and running them one after another takes several minutes even though none of those calls actually depend on each other. The fix shown here does not change the strict ordering at all. Instead, before the normal step by step processing begins, the tool figures out ahead of time which AI calls will be needed, and fetches all of their answers at once, in parallel, using a pool of worker threads with limits, retries, and timeouts. Those answers are stored in a cache. Then the normal one at a time processing runs exactly as before, but instead of waiting on a live AI call each time, it instantly pulls the already fetched answer from the cache. If an answer is missing from the cache for any reason, the system just falls back to making a real live call, so the shortcut never breaks correctness, it only saves time when it works. The repository includes a runnable demo comparing the slow, one by one approach against this prefetch approach, and reports the parallel version was about 15 times faster in their test run, while producing an identical event history. It also includes a simulated mode that works without any real AI API key or cost, so the timing comparison can be tried immediately. The README goes on to describe how to adapt an existing AI powered behavior to use this same prefetch trick, including cases where you cannot know in advance exactly which AI calls will be needed.