solid resource, the hierarchical indexing (small chunks for precision, parent chunks for context) is underrated — most RAG tutorials skip that entirely and wonder why retrieval sucks.

one thing i'd add from building production agent memory systems: conversation memory alone isn't enough once you go multi-session. you need persistent memory that survives process restarts, not just in-memory state within a LangGraph run. we hit this wall hard. agent was brilliant within a session and completely amnesic the next day lol.

the self-correction loop is interesting tho, we do something similar but at the knowledge level instead of the query level — if an agent retrieves a memory that turns out to be wrong, the confidence score on that memory gets penalized automatically so it surfaces less next time. basically reinforcement learning on your RAG results instead of just re-querying.

curious if the multi-agent map-reduce handles contradictory sub-query results? thats where things get spicy in production — two sub-agents return conflicting answers and you need a merge strategy that isn't just "pick the longer one" lol